Web Neural Network API

W3C Candidate Recommendation Draft,

More details about this document
This version:
https://www.w3.org/TR/2024/CRD-webnn-20240424/
Latest published version:
https://www.w3.org/TR/webnn/
Editor's Draft:
https://webmachinelearning.github.io/webnn/
Previous Versions:
History:
https://www.w3.org/standards/history/webnn/
Implementation Report:
https://wpt.fyi/results/webnn?label=master&label=experimental&aligned&q=webnn
Test Suite:
https://github.com/web-platform-tests/wpt/tree/master/webnn
Feedback:
GitHub
Inline In Spec
Editors:
Ningxin Hu (Intel Corporation)
Dwayne Robinson (Microsoft Corporation)
Former Editor:
Chai Chaoweeraprasit (Microsoft Corporation)
Explainer:
explainer.md
Polyfill:
webnn-polyfill / webnn-samples

Abstract

This document describes a dedicated low-level API for neural network inference hardware acceleration.

Status of this document

This section describes the status of this document at the time of its publication. A list of current W3C publications and the latest revision of this technical report can be found in the W3C technical reports index at https://www.w3.org/TR/.

This document was published by the Web Machine Learning Working Group as a Candidate Recommendation Draft using the Recommendation track.

Publication as a Candidate Recommendation does not imply endorsement by W3C and its Members. A Candidate Recommendation Draft integrates changes from the previous Candidate Recommendation that the Working Group intends to include in a subsequent Candidate Recommendation Snapshot.

This is a draft document and may be updated, replaced or obsoleted by other documents at any time. It is inappropriate to cite this document as other than work in progress.

The Web Machine Learning Working Group maintains a list of all bug reports that the group has not yet addressed. Pull requests with proposed specification text for outstanding issues are strongly encouraged.

This document was produced by a group operating under the W3C Patent Policy. W3C maintains a public list of any patent disclosures made in connection with the deliverables of the group; that page also includes instructions for disclosing a patent. An individual who has actual knowledge of a patent which the individual believes contains Essential Claim(s) must disclose the information in accordance with section 6 of the W3C Patent Policy.

This document is governed by the 03 November 2023 W3C Process Document.

Since the initial Candidate Recommendation Snapshot the Working Group has gathered further implementation experience and added new operations and data types needed for well-known transformers to support generative AI use cases. In addition, informed by this implementation experience, the group removed MLCommandEncoder, support for synchronous execution, and higher-level operations that can be expressed in terms of lower-level primitives in a performant manner. The group has also updated the specification to use modern authoring conventions to improve interoperability and precision of normative definitions. The group is developing a new feature, a backend-agnostic storage type, to improve performance and interoperability between the WebNN, WebGPU APIs and purpose-built hardware for ML and expects to republish this document as a Candidate Recommendation Snapshot when ready for implementation. This document is maintained and updated at any time. Some parts of this document are work in progress and further improvements are expected to be reflected in revised Candidate Recommendation Drafts and Snaphots.

Before requesting transition to Proposed Recommendation, the Working Group will seek to demonstrate that:

1. Introduction

The Web Neural Network API defines a web-friendly hardware-agnostic abstraction layer that makes use of Machine Learning capabilities of operating systems and underlying hardware platforms without being tied to platform-specific capabilities. The abstraction layer addresses the requirements of key Machine Learning JavaScript frameworks and also allows web developers familiar with the ML domain to write custom code without the help of libraries.

For an illustrated introduction, please see the explainer.

2. Use cases

2.1. Application Use Cases

This section illustrates application-level use cases for neural network inference hardware acceleration. All applications in those use cases can be built on top of pre-trained deep neural network (DNN) [models].

Note: Please be aware that some of the use cases described here, are by their very nature, privacy-invasive. Developers who are planning to use the API for such use cases should ensure that the API is being used to benefit users, for purposes that users understand, and approve. They should apply the Ethical Principles for Web Machine Learning [webmachinelearning-ethics] and implement appropriate privacy risk mitigations such as transparency, data minimisation, and users controls.

2.1.1. Person Detection

A user opens a web-based video conferencing application, but she temporarily leaves from her room. The application is watching whether she is in front of her PC by using object detection (for example, using object detection approaches such as [SSD] or [YOLO] that use a single DNN) to detect regions in a camera input frame that include persons.

When she comes back, the application automatically detects her and notifies other online users that she is active now.

2.1.2. Semantic Segmentation

A user joins a teleconference via a web-based video conferencing application at her desk since no meeting room in her office is available. During the teleconference, she does not wish that her room and people in the background are visible. To protect the privacy of the other people and the surroundings, the application runs a machine learning model such as [DeepLabv3+], [MaskR-CNN] or [SegAny] to semantically split an image into segments and replaces segments that represent other people and background with another picture.

2.1.3. Skeleton Detection

A web-based video conferencing application tracks a pose of user’s skeleton by running a machine learning model, which allows for real-time human pose estimation, such as [PoseNet] to recognize her gesture and body language. When she raises her hand, her microphone is automatically unmuted and she can start speaking on the teleconference.

2.1.4. Face Recognition

There are multiple people in the conference room and they join an online meeting using a web-based video conferencing application. The application detects faces of participants by using object detection (for example, using object detection approaches such as [SSD]) and checks whether each face was present at the previous meeting or not by running a machine learning model such as [FaceNet], which verifies whether two faces would be identical or not.

2.1.5. Facial Landmark Detection

A user wants to find new glasses that beautifully fits her on an online glasses store. The online store offers web-based try-on simulator that runs a machine learning model such as Face Alignment Network [FAN] to detect facial landmarks like eyes, nose, mouth, etc. When she chooses a pair of glasses, the simulator properly renders the selected glasses on the detected position of eyes on her facial image.

2.1.6. Style Transfer

A user is looking for cosmetics on an online store and wondering which color may fit her face. The online store shows sample facial makeup images of cosmetics, and offers makeup simulator that runs a machine learning model like [ContextualLoss] or [PairedCycleGAN] to transfer the makeup style of the sample makeup image to her facial image. She can check how the selected makeup looks like on her face by the simulator.

2.1.7. Super Resolution

A web-based video conferencing is receiving a video stream from its peer, but the resolution of the video becomes lower due to network congestion. To prevent degradation of the perceived video quality, the application runs a machine learning model for super-resolution such as [SRGAN] to generate higher-resolution video frames.

2.1.8. Image Captioning

For better accessibility, a web-based presentation application provides automatic image captioning by running a machine learning model such as [im2txt] which predicts explanatory words of the presentation slides.

2.1.9. Text-to-image

Images are a core part of modern web experiences. An ability to generate images based on text input in a privacy-preserving manner enables visual personalization and adaptation of web applications and content. For example, a web application can use as an input a natural language description on the web page or a description provided by the user within a text prompt to produce an image matching the text description. This text-to-image use case enabled by latent diffusion model architecture [LDM] forms the basis for additional text-to-image use cases. For example, inpainting where a portion of an existing image on the web page is selectively modified using the newly generated content, or the converse, outpainting, where an original image is extended beyond its original dimensions filling the empty space with generated content.

2.1.10. Machine Translation

Multiple people from various countries are talking via a web-based real-time text chat application. The application translates their conversation by using a machine learning model such as [GNMT] or [OpenNMT], which translates every text into different language.

2.1.11. Emotion Analysis

A user is talking to her friend via a web-based real-time text chat application, and she is wondering how the friend feels because she cannot see the friend’s face. The application analyses the friend’s emotion by using a machine learning model such as [DeepMoji], which infers emotion from input texts, and displays an emoji that represents the estimated emotion.

2.1.12. Video Summarization

A web-based video conferencing application records received video streams, and it needs to reduce recorded video data to be stored. The application generates the short version of the recorded video by using a machine learning model for video summarization such as [Video-Summarization-with-LSTM].

2.1.13. Noise Suppression

A web-based video conferencing application records received audio streams, but usually the background noise is everywhere. The application leverages real-time noise suppression using Recurrent Neural Network such as [RNNoise] for suppressing background dynamic noise like baby cry or dog barking to improve audio experiences in video conferences.

2.1.14. Speech Recognition

Speech recognition, also known as speech to text, enables recognition and translation of spoken language into text. Example applications of speech recognition include transcription, automatic translation, multimodal interaction, real-time captioning and virtual assistants. Speech recognition improves accessibility of auditory content and makes it possible to interact with such content in a privacy-preserving manner in a textual form. Examples of common use cases include watching videos or participating in online meetings using real-time captioning. Models such as [Whisper] approach humans in their accuracy and robustness and are well positioned to improve accessibility of such use cases.

2.1.15. Text Generation

Various text generation use cases are enabled by large language models (LLM) that are able to perform tasks where a general ability to predict the next item in a text sequence is required. This class of models can translate texts, answer questions based on a text input, summarize a larger body of text, or generate text output based on a textual input. LLMs enable better performance compared to older models based on RNN, CNN, or LSTM architectures and further improve the performance of many other use cases discussed in this section. Examples of LLMs include [t5-small], [m2m100_418M], [gpt2], and [llama-2-7b].

2.1.16. Detecting fake video

A user is exposed to realistic fake videos generated by ‘deepfake’ on the web. The fake video can swap the speaker’s face into the president’s face to incite a user politically or to manipulate user’s opinion. The deepfake detection applications such as [FaceForensics++] analyze the videos and protect a user against the fake videos or images. When she watches a fake video on the web, the detection application alerts her of the fraud video in real-time.

2.2. Framework Use Cases

This section collects framework-level use cases for a dedicated low-level API for neural network inference hardware acceleration. It is expected that Machine Learning frameworks will be key consumers of the Web Neural Network API (WebNN API) and the low-level details exposed through the WebNN API are abstracted out from typical web developers. However, it is also expected that web developers with specific interest and competence in Machine Learning will want to interface with the WebNN API directly instead of a higher-level ML framework.

2.2.1. Custom Layer

A web application developer wants to run a DNN model on the WebNN API. However, she has found that some of activation functions like [LeakyReLU], [ELU], etc. are not included in the WebNN API. To address this issue, she constructs custom layers of the additional activation functions on top of the WebNN API. Note that the scope of custom layers may include convolution, normalization, etc. as well as activation.

2.2.2. Network Concatenation

A web application uses a DNN model, and its model data of upper convolutional layers and lower fully-connected layers are stored in separate files, since model data of the fully-connected layers are periodically updated due to fine tuning at the server side.

Therefore, the application downloads both partial model files at first and concatenates them into a single model. When the model is updated, the application downloads fine-tuned part of the model and replace only the fully-connected layers with it.

2.2.3. Performance Adaptation

A web application developer has a concern about performance of her DNN model on mobile devices. She has confirmed that it may run too slow on mobile devices which do not have GPU acceleration. To address this issue, her web application refers to the WebNN API to confirm whether acceleration is available or not, so that the application can display the warning for devices without acceleration.

After several weeks, she has developed a tiny DNN model that can even run on CPU. In order to accommodate CPU execution, she modifies the application so that the application loads the tiny model in the case of CPU-only devices.

2.2.4. Operation Level Execution

A JavaScript ML framework is responsible for loading, interpreting and executing a ML model. During the model execution phase, the framework iterates through the operations of the model and executes each operation on the hardware device, like CPU, GPU or ML accelerator. To avoid the unnecessary data copying across devices, the framework selects the same device to execute the operations. For a compute intensive operation, such as convolution 2D or matrix multiplication, the framework uses WebNN API to execute it with the ML-specific acceleration available on that selected device.

2.2.5. Integration with real-time video processing

The user experience of WebRTC-based video conferencing is enhanced using real-time video processing. For example, background blur implemented using a § 2.1.2 Semantic Segmentation model blurs the background in the user’s live camera feed. To satisfy the performance requirements of this use case, the WebNN API integrates with primitives from other Web APIs that make up the media pipeline to allow WebNN API-based transformation of real-time video streams.

3. Security Considerations

This specification defines a low-level API for neural network inference hardware acceleration. This API is considered a powerful feature [POWERFUL-FEATURES] because it grants low-level access to a user’s computer. To meet the authentication and confidentiality expectations of a powerful feature and to prevent man-in-the-middle attacks, all interfaces defined by this specification are only available in a secure context.

This API is disabled by default in all cross-origin frames using the § 6.4 Permissions Policy Integration. This prevents third-party content from using this API unless the embedding page explicitly sets a policy that grants permission.

This API allows creation of an MLContext from a GPUDevice defined by WebGPU specification. See WebGPU Security Considerations for more information regarding security characteristics of this context.

Once the graph is fully constructed and compiled, the input shapes into each of the operations in the graph are inferred and finalized. The bounds checking occurs when the compute method is invoked that executes the graph against the actual data. No actual data is bound to the compiled graph before this stage. It is the implementation’s responsibility to make sure proper bounds checking occurs against the shapes of the data already inferred by that time.

Document operations susceptible to out-of-bounds access as a guidance to implementers.

As a future-proofing measure, the API design allows certain operations that can be generically emulated to be deprecated for security, performance, or other reasons without breaking compatibility. This is made possible by high-level functions that are defined in terms of smaller primitive operations defined in this specifications. This enables a native implementation of a high-level function to be replaced with a polyfill implementation.

Investigate side channel attack feasibility considering the current state where CPU is shared between processes running renderers.

In order to not allow an attacker to target a specific implementation that may contain a flaw, the § 6.2 Device Selection mechanism is a hint only, and the concrete device selection is left to the implementation - a user agent could for instance choose never to run a model on a device with known vulnerabilities. As a further mitigation, no device enumeration mechanism is defined.

Hinting partially mitigates the concern. Investigate additional mitigations.

The API design minimizes the attack surface for the compiled computational graph. The MLGraphBuilder interface that hosts the various operations is a data definition API and as such doesn’t execute anything, only constructs data. What follows, is that the potential for an attack is limited to when binding the data to the graph before executing it by invoking the MLContext.compute() method. This enables implementers to focus on hardening the MLContext.compute() method. For example, by making sure it honors the boundary of data and fails appropriately when the bounds are not respected.

Purpose-built Web APIs for measuring high-resolution time mitigate against timing attacks using techniques such as resolution reduction, adding jitter, detection of abuse and API call throttling [hr-time-3]. The practical deployment of WebNN implementations are likely to bring enough jitter to make timing attacks impractical (e.g. because they would use IPC) but implementers are advised to consider and test their implementations against timing attacks.

3.1. Guidelines for new operations

To ensure operations defined in this specification are shaped in a way they can be implemented securely, this section includes guidelines on how operations are expected to be defined to reduce potential for implementation problems. These guidelines are expected to evolve over time to align with industry best practices:

In general, always consider the security and privacy implications as documented in [security-privacy-questionnaire] by the Technical Architecture Group and the Privacy Interest Group when adding new features.

4. Privacy Considerations

This API enhances privacy compared to cloud-based inference, since input data such as locally sourced images or video streams stay within the browser’s sandbox.

This API exposes the minimum amount of information necessary to address the identified § 2 Use cases for the best performance and reliability of results.

No information from the underlying platform is exposed directly. An execution time analysis may reveal indirectly the performance of the underlying platform’s neural network hardware acceleration capabilities relative to another underlying platform.

Note: The group is soliciting further input on the proposed execution time analysis fingerprinting vector and will augment this section with more information and mitigations to inform the implementers of this API.

Unlike WebGPU, this API does not intrinsically support custom shader authoring; and as a result is not prone to timing attacks that rely on shader caches, or other persistent data. The API builds upon pre-existing shaders and lower level primitives of the browser or the underlying OS. Web developers who interface with GPUDevice are expected to be aware of WebGPU compilation cache considerations.

The WebGPU API identifies machine-specific artifacts as a privacy consideration. Similarly, the WebNN API’s compute unit scheduling may under certain circumstances introduce a fingerprint. However, similarly to WebGPU, such fingerprints are identical across most or all of the devices of each vendor, mitigating the concern. Furthermore, software implementations can be used to further eliminate such artifacts.

The WebNN API defines two developer-settable preferences to help inform § 6.2 Device Selection and allow the implementation to better select the most appropriate underlying execution device for the workload. An MLDeviceType normatively indicates the kind of device and is either "cpu" or "gpu". If this type cannot be satisfied, an "OperationError" DOMException is thrown, thus this type can in some cases add two bits of entropy to the fingerprint. An MLPowerPreference indicates preference as related to the power consumption and is considered a hint only and as such does not increase entropy of the fingerprint.

If a future version of this specification introduces support for a new MLDeviceType that can only support a subset of MLOperandDataTypes, that may introduce a new fingerprint.

In general, implementers of this API are expected to apply WebGPU Privacy Considerations to their implementations where applicable.

5. Ethical Considerations

The Working Group has started documenting ethical issues associated with using Machine Learning on the Web, to help identify what mitigations its normative specifications should take into account. The Working Group publishes and maintains an Ethical Principles for Web Machine Learning document [webmachinelearning-ethics] open to contributions from the wider community via a dedicated GitHub repository.

6. Programming Model

6.1. Overview

At the heart of neural networks is a computational graph of mathematical operations. These operations are the building blocks of modern machine learning technologies in computer vision, natural language processing, and robotics. The WebNN API is a specification for constructing, compiling, and executing computational graphs of neural networks.

The MLGraph interface represents a compiled computational graph that is immutable (that is, a model).

The MLGraphBuilder interface serves as a builder (factory) to construct a computational graph (its graph) that is then compiled to create an MLGraph.

In WebNN, a computational graph is composed of operators which act on data, and are the nodes of the graph. MLOperands are a representation of data that flows within the computational graph, and are the edges of the graph. MLOperands include a computational graph's input values for inference, constants (including trained weights) used for inference, intermediate values (often referred to as activations) computed during inference, as well as the output values of inference. An operator's input is one or more MLOperands. An operator's output is one or more MLOperands. Operators have operator-specific parameters that control their behavior, which can include zero or more activation functions, which are MLActivations.

A key part of the MLGraphBuilder interface are methods such as gemm() and softmax() which create an operator which represents the actual operation to perform on the input data when the computation is run, and return a new MLOperand or MLActivation holding the operator. Methods that create an MLOperand connect any inputs and activations to the operator. Each method invocation returns a distinct new value, without changing the value of any other MLOperand.

At inference time, every MLOperand will be bound to a tensor (the actual data), which are essentially multidimensional arrays. The representation of the tensors is implementation dependent, but it typically includes the array data stored in some buffer (memory) and some metadata describing the array data (such as its shape).

Operations within the computational graph have functional semantics. This allows the implementation to potentially share the array data between multiple tensors. For example, the implementation of operations such as reshape, or slice may return a view of its input tensor that shares the same buffer as the input tensor. (In the case of reshape, the entire data is shared, while in the case of slice, a part of the input data is shared.) The implementation may use views, as above, for intermediate values.

Before the execution, the computation graph that is used to compute one or more specified outputs needs to be compiled and optimized. The key purpose of the compilation step is to enable optimizations that span two or more operations, such as operation or loop fusion.

The MLGraphBuilder.build() method compiles the graph in the background without blocking the calling thread, and returns a Promise that resolves to an MLGraph. The compilation step produces an MLGraph that represents a compiled graph for optimal execution.

The MLGraph underlying implementation will be composed of platform-specific representations of operators and operands which correspond to the MLGraphBuilder's operators and MLOperands, but which are not script-visible and may be compositions or decompositions of the graph as constructed by script.

Once the MLGraph is constructed, the MLContext.compute() method performs the execution of the graph asynchronously either on a parallel timeline in a separate worker thread for the CPU execution or on a GPU timeline in a GPU command queue. This method returns immediately without blocking the calling thread while the actual execution is offloaded to a different timeline. The caller supplies the input values using MLNamedArrayBufferViews, binding the input MLOperands to their values. The caller then supplies pre-allocated buffers for output MLOperands using MLNamedArrayBufferViews. The execution produces the results of the computation from all the inputs bound to the graph. The computation results will be placed at the bound outputs at the time the operation is successfully completed on the offloaded timeline at which time the calling thread is signaled. This type of execution supports both the CPU and GPU device.

6.2. Device Selection

An MLContext interface represents a global state of neural network execution. One of the important context states is the underlying execution device that manages the resources and facilitates the compilation and the eventual execution of the neural network graph. In addition to the default method of creation with MLContextOptions, an MLContext could also be created from a specific GPUDevice that is already in use by the application.

In a situation when a GPU context executes a graph with a constant or an input in the system memory as an ArrayBufferView, the input content is automatically uploaded from the system memory to the GPU memory, and downloaded back to the system memory of an ArrayBufferView output buffer at the end of the graph execution. This data upload and download cycles will only occur whenever the execution device requires the data to be copied out of and back into the system memory, such as in the case of the GPU. It doesn’t occur when the device is a CPU device. Additionally, the result of the graph execution is in a known layout format. While the execution may be optimized for a native memory access pattern in an intermediate result within the graph, the output of the last operation of the graph must convert the content back to a known layout format at the end of the graph in order to maintain the expected behavior from the caller’s perspective.

When an MLContext is created with MLContextOptions, the user agent selects and creates the underlying execution device by taking into account the application’s MLPowerPreference and MLDeviceType options.

6.3. Task Source

The ML task source is a task source to be used for all tasks related to asynchronous compilation and execution of MLGraphs and creation of MLContexts.

To queue an ML task given a global object global and a series of steps steps, queue a global task on the ML task source with global and steps.

6.4. Permissions Policy Integration

This specification defines a policy-controlled feature identified by the string "webnn". Its default allowlist is 'self'.

7. API

7.1. The navigator.ml interface

An ML object is available in the Window and DedicatedWorkerGlobalScope contexts through the Navigator and WorkerNavigator interfaces respectively and is exposed via navigator.ml.

interface mixin NavigatorML {
  [SecureContext, SameObject] readonly attribute ML ml;
};
Navigator includes NavigatorML;
WorkerNavigator includes NavigatorML;

7.2. ML interface

enum MLDeviceType {
  "cpu",
  "gpu"
};

enum MLPowerPreference {
  "default",
  "high-performance",
  "low-power"
};

dictionary MLContextOptions {
  MLDeviceType deviceType = "cpu";
  MLPowerPreference powerPreference = "default";
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface ML {
  Promise<MLContext> createContext(optional MLContextOptions options = {});
  Promise<MLContext> createContext(GPUDevice gpuDevice);
};

7.2.1. MLContextOptions

The deviceType option is an MLDeviceType and indicates the application’s preference for the kind of device used for the context. It is one of the following:

"cpu"
Provides the broadest compatibility and usability across all client devices with varying degrees of performance.
"gpu"
Provides the broadest range of achievable performance across graphics hardware platforms from consumer devices to professional workstations.

The powerPreference option is an MLPowerPreference and indicates the application’s preference as related to power consumption. It is one of the following:

"default"
Let the user agent select the most suitable behavior.
"high-performance"
Prioritizes execution speed over power consumption.
"low-power"
Prioritizes power consumption over other considerations such as execution speed.

7.2.2. createContext()

To create a context given realm realm and options (a GPUDevice or MLContextOptions), run these steps:
  1. Let context be a new MLContext object with realm.

  2. If options is a GPUDevice object,

    1. Set context.[[contextType]] to "webgpu".

    2. Set context.[[deviceType]] to "gpu".

    3. Set context.[[powerPreference]] to "default".

  3. Otherwise,

    1. Set context.[[contextType]] to "default".

    2. If options["deviceType"] exists, then set context.[[deviceType]] to options["deviceType"]. Otherwise, set context.[[deviceType]] to "cpu".

    3. If options["powerPreference"] exists, then set context.[[powerPreference]] to options["powerPreference"]. Otherwise, set context.[[powerPreference]] to "default".

  4. If the user agent cannot support context.[[contextType]], context.[[deviceType]] and context.[[powerPreference]], return failure.

  5. Return context.

The createContext(options) steps are:
  1. Let global be this's relevant global object.

  2. If global’s associated Document is not allowed to use the webnn feature, return a new promise rejected with a "SecurityError" DOMException.

  3. Let realm be this's relevant realm.

  4. Let promise be a new promise.

  5. Run the following steps in parallel.

    1. Let context be the result of creating a context given realm and options. If that returns failure, then queue an ML task with global to reject promise with a "NotSupportedError" DOMException and abort these steps.

    2. Queue an ML task with global to resolve promise with context.

  6. Return promise.

The createContext(gpuDevice) method steps are:
  1. Let global be this's relevant global object.

  2. If global’s associated Document is not allowed to use the webnn feature, return a new promise rejected with a "SecurityError" DOMException.

  3. Let realm be this's relevant realm.

  4. Let promise be a new promise.

  5. Run the following steps in parallel.

    1. Let context be the result of creating a context given realm and gpuDevice. If that returns failure, then queue an ML task with global to reject promise with a "NotSupportedError" DOMException and abort these steps.

    2. Queue an ML task with global to resolve promise with context.

  6. Return promise.

7.3. MLContext interface

The MLContext interface represents a global state of neural network compute workload and execution processes. Each MLContext object has associated context type, MLDeviceType and MLPowerPreference.
typedef record<DOMString, ArrayBufferView> MLNamedArrayBufferViews;

dictionary MLComputeResult {
  MLNamedArrayBufferViews inputs;
  MLNamedArrayBufferViews outputs;
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLContext {
  Promise<MLComputeResult> compute(
      MLGraph graph, MLNamedArrayBufferViews inputs, MLNamedArrayBufferViews outputs);
};
MLContext has the following internal slots:
[[contextType]] of type context type.

The MLContext's context type.

[[deviceType]] of type MLDeviceType.

The MLContext's MLDeviceType.

[[powerPreference]] of type MLPowerPreference.

The MLContext's MLPowerPreference.

The context type is the type of the execution context that manages the resources and facilitates the compilation and execution of the neural network graph:

"default"
Context created per user preference options.
"webgpu"
Context created from WebGPU device.
When the [[contextType]] is set to default with the MLContextOptions.deviceType set to "gpu", the user agent is responsible for creating an internal GPU device that operates within the context and is capable of ML workload submission on behalf of the calling application. In this setting however, only ArrayBufferView inputs and outputs are allowed in and out of the graph execution since the application has no way to know what type of internal GPU device is being created on their behalf. In this case, the user agent is responsible for automatic uploads and downloads of the inputs and outputs to and from the GPU memory using this said internal device.
inputs, of type MLNamedArrayBufferViews

An object where the keys are the graph input names, and the values are the transferred ArrayBufferViews for the supplied input tensor values.

outputs, of type MLNamedArrayBufferViews

An object where the keys are the graph output names, and the values are the transferred ArrayBufferViews for the computed output tensor values.

To validate buffer with descriptor given ArrayBufferView bufferView and MLOperandDescriptor descriptor, run the following steps:
  1. If bufferView’s element type does not match to descriptor.dataType according to this table, return false.

  2. If bufferView.[[ByteLength]] is not equal to descriptor’s byte length, return false.

To execute graph, given MLGraph graph, MLNamedArrayBufferViews inputs and MLNamedArrayBufferViews outputs, run the following steps. They return undefined, or an error.
  1. Let inputResources denote the input resources of graph.[[implementation]].

  2. For each nameinputValue of inputs:

    1. Let inputDescriptor be graph.[[inputDescriptors]][name].

    2. Let inputTensor be a new tensor for graph.[[implementation]] as follows:

      1. Set the data type of inputTensor to the one that matches inputValue’s element type.

      2. Set the dimensions of inputTensor to inputDescriptor.dimensions.

      3. Set the values of elements in inputTensor to the values of elements in inputValue.

    3. Request the underlying implementation of graph to bind inputResources[name] to inputTensor.

  3. For each nameoutputValue of outputs:

    1. Issue a compute request to graph.[[implementation]] given name and inputResources and wait for completion.

      1. If that returns an error, then return an "OperationError" DOMException.

      2. Otherwise, store the result in outputTensor.

    2. Let outputDesc be graph.[[outputDescriptors]][name].

    3. If the byte length of outputTensor is not equal to outputDesc’s byte length, then return a TypeError.

    4. If outputTensor’s element type doesn’t match outputValue’s element type, then return a TypeError.

    5. Request the underlying implementation of graph to set the values of elements in outputValue to the values of elements in outputTensor.

  4. Return undefined.

7.3.1. MLNamedArrayBufferViews transfer algorithm

To transfer an MLNamedArrayBufferViews views with realm realm:
  1. Let transferredViews be a new MLNamedArrayBufferViews.

  2. For each nameview of views:

    1. Let transferredBuffer be the result of transferring view’s underlying buffer.

    2. Let constructor be the appropriate view constructor for the type of ArrayBufferView view from realm.

    3. Let elementsNumber be the result of view’s byte length / view’s element size.

    4. Let transferredView be Construct(constructor, transferredBuffer, view.[[ByteOffset]], elementsNumber).

    5. Set transferredViews[name] to transferredView.

  3. Return transferredViews.

7.3.2. compute()

Asynchronously carries out the computational workload of a compiled graph MLGraph on a separate timeline, either on a worker thread for the CPU execution, or on a GPU timeline for the submission of GPU workload on the command queue. The asynchronous nature of this call avoids blocking the calling thread while the computation for result is ongoing. This method of execution requires an MLContext created with MLContextOptions. Otherwise, it throws an "OperationError" DOMException.
In accordance with the Web IDL warning, to prevent the calling thread from modifying the input and output resources while the computation is ongoing, this method transfers the input and output MLNamedArrayBufferViews to new views that share the same backing memory allocations. The transferred views are returned to the caller via the promise fulfillment with the computation result written into the backing memory of the output views.
Arguments:

Returns: Promise<MLComputeResult>.

Note: Invocations of compute() will fail if any of the graph's inputs are not provided as inputs, or if any requested outputs do not match the graph's outputs.

The compute(graph, inputs, outputs) method steps are:
  1. Let global be this's relevant global object.

  2. Let realm be this's relevant realm.

  3. If graph.[[context]] is not this, then return a new promise rejected with a TypeError.

  4. If graph.[[context]].[[contextType]] is not "default", then return a new promise rejected with an "OperationError" DOMException.

  5. For each namedescriptor of graph.[[inputDescriptors]]:

    1. If inputs[name] does not exist, then return a new promise rejected with a TypeError.

    2. If validating buffer with descriptor given inputs[name] and descriptor returns false, then return a new promise rejected with a TypeError.

  6. For each nameresource of outputs:

    1. If graph.[[outputDescriptors]][name] does not exist, then return a new promise rejected with a TypeError.

    2. If validating buffer with descriptor given resource and graph.[[outputDescriptors]][name] returns false, then return a new promise rejected with a TypeError.

  7. Let transferredInputs be the result of transferring MLNamedArrayBufferViews inputs with realm. If that threw an exception, then return a new promise rejected with that exception.

  8. Let transferredOutputs be the result of transferring MLNamedArrayBufferViews outputs with realm. If that threw an exception, then return a new promise rejected with that exception.

  9. Let promise be a new promise.

  10. Run the following steps in parallel:

    1. Invoke execute graph given graph, transferredInputs and transferredOutputs. If that returns an error, then queue an ML task with global to reject promise with an equivalent error in realm and abort these steps.

    2. Let result be a new MLComputeResult with realm.

    3. Set result.inputs to transferredInputs.

    4. Set result.outputs to transferredOutputs.

    5. Queue an ML task with global to resolve promise with result.

  11. Return promise.

7.3.2.1. Examples
The following code showcases the asynchronous computation.
const operandType = {dataType: 'float32', dimensions: [2, 2]};
const context = await navigator.ml.createContext();
const builder = new MLGraphBuilder(context);
// 1. Create a computational graph 'C = 0.2 * A + B'.
const constant = builder.constant(0.2);
const A = builder.input('A', operandType);
const B = builder.input('B', operandType);
const C = builder.add(builder.mul(A, constant), B);
// 2. Compile it into an executable.
const graph = await builder.build({'C': C});
// 3. Bind inputs to the graph and execute for the result.
const bufferA = new Float32Array(4).fill(1.0);
const bufferB = new Float32Array(4).fill(0.8);
const bufferC = new Float32Array(4);
const inputs = {'A': bufferA, 'B': bufferB};
const outputs = {'C': bufferC};
const result = await context.compute(graph, inputs, outputs);
// The computed result of [[1, 1], [1, 1]] is in the buffer associated with
// the output operand.
console.log('Output value: ' + result.outputs.C);
// Note: the result.outputs.C buffer is different from the bufferC, but it
// shares the same backing memory allocation.

7.4. MLGraph interface

The MLGraph interface represents a compiled computational graph. A compiled graph once constructed is immutable and cannot be subsequently changed.
[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLGraph {};
MLGraph has the following internal slots:
[[context]] of type MLContext

The context of type MLContext associated with this MLGraph.

[[inputDescriptors]] of type record<DOMString, MLOperandDescriptor>

Maps the name of an input MLOperand to its MLOperandDescriptor for all input MLOperands of this MLGraph.

[[outputDescriptors]] of type record<DOMString, MLOperandDescriptor>

Maps the name of an output MLOperand to its MLOperandDescriptor for all output MLOperands of this MLGraph.

[[implementation]]

The underlying implementation provided by the User Agent.

7.5. MLOperandDescriptor dictionary

An MLOperandDescriptor describes the shape (dimensions) and data type of an operand. They are used to describe the inputs and constants for an MLGraph, and every MLOperand has an internal MLOperandDescriptor.

enum MLInputOperandLayout {
  "nchw",
  "nhwc"
};

enum MLOperandDataType {
  "float32",
  "float16",
  "int32",
  "uint32",
  "int64",
  "uint64",
  "int8",
  "uint8"
};

dictionary MLOperandDescriptor {
  required MLOperandDataType dataType;
  sequence<[EnforceRange] unsigned long> dimensions = [];
};
dataType, of type MLOperandDataType

The operand data type.

dimensions, of type sequence<[EnforceRange] unsigned long>, defaulting to []

The shape of the operand. It is empty for scalar operands, and non-empty for tensor operands.

The byte length of an MLOperandDescriptor desc is the value returned by the following steps:
  1. Let elementLength be 1.

  2. For each dimension of desc.dimensions:

    1. Set elementLength to elementLength * dimension.

  3. Let elementSize be the element size of one of the ArrayBufferView types that matches desc.dataType according to this table.

  4. Return elementLength * elementSize.

A valid dimension is an integer greater than zero in the range of unsigned long. Implementations may impose a smaller upper bound.

Should 0-size dimensions be supported? [Issue #391]

To check dimensions given MLOperandDescriptor descriptor, run the following steps:
  1. If any element of descriptor.dimensions is not a valid dimension, return false.

  2. If descriptor.dimensions's size is too large to be supported by the implementation, return false.

    The maximum number of operand dimensions is not defined, but native ML APIs usually have a maximum supported size. [Issue #456]

  3. If descriptor’s byte length is not supported by the implementation, then return false.

  4. Return true.

7.6. MLOperand interface

An MLOperand represents an intermediary graph being constructed as a result of compositing parts of an operation into a fully composed operation.

For instance, an MLOperand may represent a constant feeding to an operation or the result from combining multiple constants together into an operation. See also § 6 Programming Model.

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLOperand {
  MLOperandDataType dataType();
  sequence<unsigned long> shape();
};
MLOperand has the following internal slots:
[[builder]] of type MLGraphBuilder

The MLOperand's associated builder object.

[[descriptor]] of type MLOperandDescriptor

The MLOperand's descriptor.

[[name]] of type string

The MLOperand's name (only for input operands).

[[operator]] of type operator

Reference to MLOperand's corresponding operator.

An MLOperand's shape is its [[descriptor]].dimensions.

An MLOperand's rank is its shape's size.

An MLOperand's dataType is its [[descriptor]].dataType.

Since the [[builder]] object is bound by the MLGraphBuilder() constructor to an MLContext object, an MLOperand is also always bound to the same MLContext object.

7.6.1. Creating an MLOperand

The MLOperand objects are created by the methods of MLGraphBuilder, internally using the following algorithms.
To create an MLOperand given MLGraphBuilder builder and MLOperandDescriptor desc, run the following steps:
  1. Let operand be a new MLOperand.

  2. Set operand.[[builder]] to builder.

  3. Set operand.[[descriptor]] to desc.

  4. Return operand.

To copy an MLOperand given MLOperand operand, run the following steps:
  1. Let result be a new MLOperand.

  2. Set result.[[builder]] to operand.[[builder]].

  3. Set result.[[descriptor]] to operand.[[descriptor]].

  4. If operand.[[name]] exists, then set result.[[name]] to operand.[[name]].

  5. Return result.

To validate operand given MLGraphBuilder builder and MLOperand operand, return true if operand.[[builder]] is builder, and false otherwise.

7.6.2. dataType()

Return a data type of the MLOperand.
Returns: an MLOperandDataType. The data type of the operand.
The dataType() method steps are:
  1. Return this's dataType.

7.6.3. shape()

Return a shape of the MLOperand.
Returns: a sequence of unsigned long. The shape of the operand.
The shape() method steps are:
  1. Return this's shape.

7.7. MLActivation interface

Objects implementing the MLActivation interface represent activation function types.

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLActivation {};
MLActivation has the following internal slots:
[[name]] of type string

The MLActivation's name.

[[builder]] of type MLGraphBuilder

A dictionary containing MLActivation options.

[[operator]] of type operator

Reference to MLActivation's corresponding operator.

These activations function types are used to create other operations. One such use of this interface is for when an activation function is fused into another operation such as conv2d() or batchNormalization() during a graph construction session. Such fused activation functions can provide a significant performance improvement when supported natively by the underlying implementation. This is intended as an optimization opportunity for implementers.

7.7.1. Creating MLActivation

The MLActivation objects (including the ones passed as input to methods) are created by the methods of MLGraphBuilder and are identified by their name. The options dictionary is defined by those methods. The actual creation of the activation function e.g. a sigmoid() or relu() can then be deferred until when the rest of the graph is ready to connect with it such as during the construction of conv2d() for example.
To create an MLActivation given MLGraphBuilder builder, string name and optional ordered map options, run the following steps:
  1. Let activation be a new MLActivation.

  2. Set activation.[[builder]] to builder.

  3. Set activation.[[name]] to name.

  4. Let operator be an operator for the name operation, given options.

  5. Set activation.[[operator]] to operator.

  6. Return activation.

To validate activation given MLGraphBuilder builder and MLActivation activation, return true if activation.[[builder]] is builder, and false otherwise.

7.8. MLGraphBuilder interface

The MLGraphBuilder interface defines a set of operations as identified by the § 2 Use cases that can be composed into a computational graph. It also represents the intermediate state of a graph building session.

typedef record<DOMString, MLOperand> MLNamedOperands;

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLGraphBuilder {
  // Construct the graph builder from the context.
  constructor(MLContext context);

  // Create an operand for a graph input.
  MLOperand input(DOMString name, MLOperandDescriptor descriptor);

  // Create an operand for a graph constant.
  MLOperand constant(MLOperandDescriptor descriptor, ArrayBufferView bufferView);

  // Create a single-value operand from the specified number of the specified type.
  MLOperand constant(double value, optional MLOperandDataType type = "float32");

  // Compile the graph up to the specified output operands asynchronously.
  Promise<MLGraph> build(MLNamedOperands outputs);
};
The MLGraphBuilder.build() method compiles the graph builder state up to the specified output operands into a compiled graph according to the type of MLContext that creates it. When the [[contextType]] of the MLContext is set to "default", the compiled graph is initialized right before the MLGraph is returned. This graph initialization stage is important for optimal performance of the subsequent graph executions. It typically involves a process known as "weight preprocessing" where all the constant inputs to the graph are preprocessed and cached at the operating system level for subsequent graph execution calls. The initializing inputs are typically the constant weight data specified through the MLGraphBuilder/constant(value, type) method as constant operands during graph construction time.
MLGraphBuilder has the following internal slots:
[[context]] of type MLContext

The context of type MLContext associated with this MLGraphBuilder.

7.8.1. MLGraphBuilder constructor

The new MLGraphBuilder(context) constructor steps are:
  1. If this's relevant global object's associated Document is not allowed to use the webnn feature, then throw a "SecurityError" DOMException.

  2. Set this.[[context]] to context.

7.8.2. input operands

Create a named MLOperand based on a descriptor, that can be used as an input.

Arguments: Returns: an MLOperand object.
The input(name, descriptor) method steps are:
  1. If name is empty, then throw a TypeError.

  2. If checking dimensions given descriptor returns false, then throw a TypeError.

  3. Make graph connections:

    1. Let operand be the result of creating an MLOperand given this and descriptor.

    2. Set operand.[[name]] to name.

    3. Add operand to this's graph's inputs.

  4. Return operand.

The MLGraphBuilder API allows creating an MLGraph without input operands. If the underlying platform doesn’t support that, implementations may add a stub input or passing constants as inputs to the graph.

7.8.3. constant operands

Create a constant MLOperand that can be used in MLGraphBuilder methods.
7.8.3.1. constant(descriptor, bufferView)
Create a constant MLOperand of the specified data type and shape that contains the initializing data.
Arguments: Returns: an MLOperand. The constant output tensor.
The constant(descriptor, bufferView) method steps are:
  1. If checking dimensions given descriptor returns false, then throw a TypeError.

  2. If validating buffer with descriptor given bufferView and descriptor returns false, then throw a TypeError.

  3. Make graph connections:

    1. Let operand be the result of creating an MLOperand given this and descriptor.

    2. Let bytes be the result of getting a copy of the bytes held by the buffer source given bufferView.

    3. Add operand to this's graph's constants with bytes as value.

  4. Return operand.

7.8.3.2. constant(value, type)
Create a constant MLOperand of the specified value and data type.
Data truncation will occur when the specified value exceeds the range of the specified output data type e.g. when a float value is assigned to an "int8" data type, etc.
Arguments: Returns: an MLOperand. The constant output.
The constant(value, type) method steps are:
  1. Let descriptor be a new MLOperandDescriptor.

    1. Set descriptor.dataType to type.

    2. Set descriptor.dimensions to an empty list.

  2. Make graph connections:

    1. Let operand be the result of creating an MLOperand given this and descriptor.

    2. Add operand to this's graph's constants with value as value.

  3. Return operand.

7.8.3.3. constant(start, end, step, type)
Create a constant MLOperand of the specified data type and shape that contains the data as specified by the range.
Data truncation will occur when the values in the range exceed the range of the specified output data type e.g. when a float value is assigned to an "int8" data type, etc.
Arguments: Returns: an MLOperand. The constant 1-D output tensor of size max(0, ceil((end - start)/step)).
The constant(start, end, step, type) method steps are:
  1. Let descriptor be a new MLOperandDescriptor.

    1. Set descriptor.dataType to type.

    2. Let size be max(0, ceil((end - start)/step)).

    3. If size is not a valid dimension, then throw a TypeError.

    4. Set descriptor.dimensions to the list « size ».

  2. Make graph connections:

    1. Let operand be the result of creating an MLOperand given this and descriptor.

    2. Let buffer be an implementation-defined platform memory buffer the size of size multiplied by sizeof(descriptor.dataType).

    3. For each index in the range 0 to size, exclusive:

      1. Set buffer[index] to start + (index * step).

    4. Add operand to this's graph's constants with buffer as value.

  3. Return operand.

7.8.4. build method

Build a composed graph up to a given output operand into a computational graph asynchronously.
The build(outputs) method steps are:
  1. If outputs is empty, then return a new promise rejected with a TypeError.

  2. For each nameoperand of outputs:

    1. If name is empty, then return a new promise rejected with a TypeError.

    2. If validating operand given this and operand returns false, then return a new promise rejected with a TypeError.

    3. If operand is in this's graph's inputs or constants, then return a new promise rejected with a TypeError.

  3. Let operands be a new empty set.

  4. Let operators be a new empty set.

  5. Let inputs be a new empty set.

  6. Let queue be a new queue containing outputs’s values.

  7. While queue is not empty:

    1. Dequeue operand from queue.

    2. Append operand to operands.

    3. Append operand.[[operator]] to operators.

    4. If operand is in this's graph's inputs, append operand to inputs.

    5. For each input of operand.[[operator]]'s inputs:

      1. Enqueue input to queue.

  8. If any MLOperands in inputs have the same [[name]], then return a new promise rejected with a TypeError.

    If MLGraphBuilder can’t be re-used, then this can be simplified: enforce uniqueness in input() instead, and iteration can be done over all of the graph’s inputs instead of needing this traversal. [Issue #567]

  9. Let global be this's relevant global object.

  10. Let realm be this's relevant realm.

  11. Let graph be a new MLGraph with realm.

  12. Set graph.[[context]] to this.[[context]].

  13. For each operand in inputs:

    1. Set graph.[[inputDescriptors]][operand.[[name]]] to operand.[[descriptor]].

      If constants' ArrayBuffers are not transferred, make copies for graph's constants here. [Issue #566]

  14. For each nameoperand of outputs:

    1. Set graph.[[outputDescriptors]][name] to operand.[[descriptor]].

  15. Let promise be a new promise.

  16. Run the following steps in parallel:

    1. Let graphImpl be the result of converting this's graph with operands, operators, inputs, and outputs’s values into an implementation-defined format which can be interpreted by the underlying platform.

      1. If the underlying platform does not support a requested feature, then queue an ML task with global to reject promise with an "OperationError" DOMException, and abort these steps.

    2. Set graph.[[implementation]] to graphImpl.

    3. Queue an ML task with global to resolve promise with graph.

  17. Return promise.

7.8.5. argMin/argMax operations

Return the index location of the minimum or maxmium values of all the input values along the axes.
dictionary MLArgMinMaxOptions {
  sequence<[EnforceRange] unsigned long> axes;
  boolean keepDimensions = false;
  boolean selectLastIndex = false;
};

partial interface MLGraphBuilder {
  MLOperand argMin(MLOperand input, optional MLArgMinMaxOptions options = {});
  MLOperand argMax(MLOperand input, optional MLArgMinMaxOptions options = {});
};

MLArgMinMaxOptions has the following members:

axes, of type sequence<[EnforceRange] unsigned long>

The dimensions to reduce. The values must be in the range [0, N-1] where N is the rank of the input tensor. If not present, all dimensions are reduced. If empty, no dimensions are reduced, and the shape of the output tensor is the same as the shape of the input tensor.

keepDimensions, of type boolean, defaulting to false

If true, retains reduced dimensions with size 1.

selectLastIndex, of type boolean, defaulting to false

If true, select the last index instead of the first found along the axes.

Arguments:

Returns: an MLOperand. The N-D tensor of the reduced shape. The values must be of type "int64" in the range [0, N-1] where N is the corresponding size of each of the input dimensions specified by options.axes.

To create argMin/argMax operation given string op, MLOperand input and MLArgMinMaxOptions options, run the following steps:
  1. Assert: op is one of "argMin", "argMax".

  2. If validating operand with this and input returns false, then throw a TypeError.

  3. If options.axes exists, if any of its elements is not in the range 0 to input’s rank, exclusive, then throw a TypeError.

  4. Let outputShape be the result of calculating reduction output sizes given input’s shape, options.axes (if it exists), and options.keepDimensions.

  5. Let desc be a new MLOperandDescriptor.

  6. Set desc.dataType to "int64".

  7. Set desc.dimensions to outputShape.

  8. Make graph connections:

    1. Let operator be an operator for the operation op, given options.

    2. Let output be the result of creating an MLOperand given this and desc.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  9. Return output.

The following argMin/argMax algorithms are supported.
The argMin(input, options) method steps are:
  1. Let output be the result of running the create argMin/argMax operation given "argMin", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The argMax(input, options) method steps are:
  1. Let output be the result of running the create argMin/argMax operation given "argMax", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

7.8.6. batchNormalization

Normalize the values of the input tensor using [Batch-Normalization]. For each input feature, the mean and variance values of that feature are computed across all the samples in the batch dimension while the model is trained. These mean and variance values are then subsequently given to this operation during model inference.
dictionary MLBatchNormalizationOptions {
  MLOperand scale;
  MLOperand bias;
  [EnforceRange] unsigned long axis = 1;
  float epsilon = 1e-5;
  MLActivation activation;
};

partial interface MLGraphBuilder {
  MLOperand batchNormalization(MLOperand input, MLOperand mean, MLOperand variance,
                               optional MLBatchNormalizationOptions options = {});
};

MLBatchNormalizationOptions has the following members:

scale, of type MLOperand

The 1-D tensor of the scaling values whose size is equal to the size of the input dimension denoted by axis.

bias, of type MLOperand

The 1-D tensor of the bias values whose size is equal to the size of the input dimension denoted by axis.

axis, of type unsigned long, defaulting to 1

The index to the feature count dimension of the input shape for which the mean and variance values are. Its value must be in the range [0, N-1] where N is the rank of the input tensor. The default value is 1, corresponding to the channel ("c") dimension in the "nchw" data layout.

epsilon, of type float, defaulting to 1e-5

A small value to prevent computational error due to divide-by-zero.

activation, of type MLActivation

An optional activation function that immediately follows the normalization operation.

Arguments:

Returns: an MLOperand. The batch-normalized N-D tensor of the same shape as input.

The batchNormalization(input, mean, variance, options) method steps are:
  1. If validating operand with this and any of input, mean, variance, options.scale (if it exists), and options.bias (if it exists) returns false, then throw a TypeError.

  2. If options.activation exists, and validating activation with this and it returns false, then throw a TypeError.

  3. If options.axis is not in the range 0 to input’s rank, exclusive, then throw a TypeError.

  4. If mean’s rank is not 1, then throw a TypeError.

  5. If mean’s shape[0] is not equal to input’s shape[options.axis], then throw a TypeError.

  6. If variance’s rank is not 1, then throw a TypeError.

  7. If variance’s shape[0] is not equal to input’s shape[options.axis], then throw a TypeError.

  8. If options.scale exists:

    1. If its rank is not 1, then throw a TypeError.

    2. If its shape[0] is not equal to input’s shape[options.axis], then throw a TypeError.

  9. If options.bias exists:

    1. If its rank is not 1, then throw a TypeError.

    2. If its shape[0] is not equal to input’s shape[options.axis], then throw a TypeError.

  10. Make graph connections:

    1. Let operator be an operator for the batchNormalization operation, given input, mean, variance and options.

    2. Let output be the result of creating an MLOperand given this and input.[[descriptor]].

    3. If options.activation exists, then add it to operator’s activation functions.

    4. Set output.[[operator]] to operator.

    5. Set operator’s inputs to input, mean, and variance.

    6. If options.scale exists, then add it to operator’s inputs.

    7. If options.bias exists, then add it to operator’s inputs.

    8. Set operator’s output to output.

  11. Return output.

The behavior of this operation when the input tensor is 4-D of the "nchw" layout and the activation is relu() can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
const shape = [1,null,1,1];
return builder.relu(
  builder.add(
    builder.mul(
      builder.reshape(options.scale, shape),
      builder.div(
        builder.sub(input, builder.reshape(mean, shape)),
        builder.sqrt(builder.add(builder.reshape(variance, shape), builder.constant(options.epsilon)))
        )),
    builder.reshape(options.bias, shape)));

7.8.7. cast

Cast each element in the input tensor to the target data type.
partial interface MLGraphBuilder {
  MLOperand cast(MLOperand input, MLOperandDataType type);
};
Arguments:

Returns: an MLOperand. The N-D tensor of the same shape as input with each element casted to the target data type.

The cast(input, type) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let operator be an operator for the cast operation, given type.

    2. Let output be the result of copying an MLOperand given input.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.8. clamp

Clamp the input tensor element-wise within a range specified by the minimum and maximum values.
dictionary MLClampOptions {
  float minValue;
  float maxValue;
};

partial interface MLGraphBuilder {
  MLOperand clamp(MLOperand input, optional MLClampOptions options = {});
  MLActivation clamp(optional MLClampOptions options = {});
};
minValue, of type float

The minimum value of the range. When it is not specified, the clamping is not performed on the lower limit of the range.

maxValue, of type float

The maximum value of the range. When it is not specified, the clamping is not performed on the upper limit of the range.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
if (options.minValue === undefined) {
  if (options.maxValue === undefined) {
    return input;
  } else {
    return builder.min(input, builder.constant(options.maxValue));
  }
} else {
  if (options.maxValue === undefined) {
    return builder.max(input, builder.constant(options.minValue));
  } else {
    return builder.min(
        builder.max(input, builder.constant(options.minValue)),
        builder.constant(options.maxValue));
  }
}
To check clamp options given MLClampOptions options, run the following steps:
  1. If options.minValue is greater than options.maxValue, then return false.

    Not all implementations support minValue equal to maxValue. [Issue #396]

  2. Return true.

7.8.8.1. clamp(input, options)
Arguments: Returns:
The clamp(input, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. If checking clamp options given options returns false, then throw a TypeError.

  3. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the clamp operation, given options.minValue and options.maxValue.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  4. Return output.

7.8.8.2. clamp(options)
Arguments: Returns:
The clamp(options) method steps are:
  1. If checking clamp options given options returns false, then throw a TypeError.

  2. Let op be the result of creating an MLActivation given this, "clamp" and options.

  3. Return op.

7.8.9. concat

Concatenates the input tensors along a given axis.
partial interface MLGraphBuilder {
  MLOperand concat(sequence<MLOperand> inputs, [EnforceRange] unsigned long axis);
};
Arguments:

Returns: an MLOperand. The concatenated tensor of all the inputs along the axis. The output tensor has the same shape except on the dimension that all the inputs concatenated along. The size of that dimension is computed as the sum of all the input sizes of the same dimension.

The concat(inputs, axis) method steps are:
  1. If validating operand with this and any item in inputs returns false, then throw a TypeError.

  2. If inputs is empty, then throw a TypeError.

  3. Let first be inputs[0].

  4. If axis is greater than or equal to first’s rank, then throw a TypeError.

  5. Let desc be a new MLOperandDescriptor.

  6. Set desc.dataType to first’s dataType.

  7. Set desc.dimensions to a clone of first’s shape.

  8. Set desc.dimensions[axis] to first’s shape[axis].

  9. For each index in the range 1 to inputs’s size, exclusive:

    1. Let input be inputs[index].

    2. If input’s dataType is not equal to first’s dataType, then throw a TypeError.

    3. If input’s rank is not equal to first’s rank, then throw a TypeError.

    4. For each dim in the range 0 to input’s rank, exclusive:

      If the shape of each corresponding dimension and type of the operands, except for those of the dimension given by axis, is not the same, fail.
      1. If dim is not equal to axis and if input’s shape[dim] is not equal to first’s shape[dim], then throw a TypeError.

      2. If dim is equal to axis:

        1. Let size be the sum of desc.dimensions[axis] and input’s shape[dim].

        2. If size is not a valid dimension, then throw a TypeError.

        3. Set desc.dimensions[axis] to size.

  10. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the concat operation, given inputs and axis.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to inputs.

    5. Set operator’s output to output.

  11. Return output.

7.8.10. conv2d

Compute a 2-D convolution given 4-D input and filter tensors
enum MLConv2dFilterOperandLayout {
  "oihw",
  "hwio",
  "ohwi",
  "ihwo"
};

dictionary MLConv2dOptions {
  sequence<[EnforceRange] unsigned long> padding;
  sequence<[EnforceRange] unsigned long> strides;
  sequence<[EnforceRange] unsigned long> dilations;
  [EnforceRange] unsigned long groups = 1;
  MLInputOperandLayout inputLayout = "nchw";
  MLConv2dFilterOperandLayout filterLayout = "oihw";
  MLOperand bias;
  MLActivation activation;
};

partial interface MLGraphBuilder {
  MLOperand conv2d(MLOperand input,
                   MLOperand filter,
                   optional MLConv2dOptions options = {});
};

MLConv2dOptions has the following members:

padding, of type sequence<[EnforceRange] unsigned long>

A list of length 4: [beginningHeight, endingHeight, beginningWidth, endingWidth]. Specifies the additional rows and columns added to the beginning and ending of each spatial dimension of the convolution input. The default value is [0, 0, 0, 0].

strides, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [strideHeight, strideWidth]. Specifies the stride of the sliding window for each spatial dimension of the convolution input. The default value is [1, 1].

dilations, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [dilationHeight, dilationWidth]. Specifies the dilation factor for each spatial dimension applied on the convolution filter (kernel). The default value is [1, 1].

groups, of type unsigned long, defaulting to 1

The number of groups that input channels and output channels are divided into.

inputLayout, of type MLInputOperandLayout, defaulting to "nchw"

Specifies the layout format of the input and output tensor as follows:

  • "nchw"

    • input tensor: [batches, inputChannels, height, width]

    • output tensor: [batches, outputChannels, height, width]

  • "nhwc":

    • input tensor: [batches, height, width, inputChannels]

    • output tensor: [batches, height, width, outputChannels]

filterLayout, of type MLConv2dFilterOperandLayout, defaulting to "oihw"

Specifies the layout format of the filter tensor as follows:

  • "oihw": [outputChannels, inputChannels/groups, height, width]

  • "hwio": [height, width, inputChannels/groups, outputChannels]

  • "ohwi": [outputChannels, height, width, inputChannels/groups]

  • "ihwo": [inputChannels/groups, height, width, outputChannels]

bias, of type MLOperand

An additional 1-D tensor with the shape of [outputChannels] whose values are to be added to the convolution result.

activation, of type MLActivation

An optional activation function that immediately follows the convolution operation.

Arguments:

Returns: an MLOperand. The output 4-D tensor that contains the convolution result. The output shape is interpreted according to the options.inputLayout value. More specifically, the spatial dimensions or the sizes of the last two dimensions of the output tensor for the nchw input layout can be calculated as follow:

outputSize = 1 + (inputSize - (filterSize - 1) * dilation - 1 + beginningPadding + endingPadding) / stride

A depthwise conv2d operation is a variant of grouped convolution, used in models like the MobileNet, where the options.groups = inputChannels = outputChannels and the shape of filter tensor is [options.groups, 1, height, width] for "oihw" layout, [height, width, 1, options.groups] for "hwio" layout, [options.groups, height, width, 1] for "ohwi" layout and [1, height, width, options.groups] for "ihwo" layout.
To calculate conv output size given unsigned integers inputSize, filterSize, beginningPadding, endingPadding, stride and dilation, perform these steps. They return a number.
  1. Let effectiveFilterSize be ( filterSize - 1 ) * dilation + 1.

  2. Let outputSize be ( inputSize - effectiveFilterSize + beginningPadding + endingPadding ) / stride + 1.

  3. Return outputSize.

To calculate conv2d output sizes given unsigned integers inputHeight, inputWidth, filterHeight and filterWidth, list of 4 unsigned integers padding, list of 2 unsigned integers strides, and list of 2 unsigned integers dilations, perform these steps. They return a list of 2 numbers.
  1. Let outputHeight be the result of calculating conv output size given inputHeight, filterHeight, padding[0], padding[1], strides[0] and dilations[0].

  2. Let outputWidth be the result of calculating conv output size given inputWidth, filterWidth, padding[2], padding[3], strides[1] and dilations[1].

  3. Return « outputHeight, outputWidth ».

The conv2d(input, filter, options) method steps are:
  1. If validating operand with this and any of input, filter, and options.bias (if it exists) returns false, then throw a TypeError.

  2. If options.activation exists, and validating activation with this and it returns false, then throw a TypeError.

  3. If input’s rank is not 4, then throw a TypeError.

  4. If filter’s rank is not 4, then throw a TypeError.

  5. If filter’s dataType is not equal to input’s dataType, then throw a TypeError.

  6. If options.padding does not exist, set it to the list « 0, 0, 0, 0 ».

  7. Else if options.padding's size is not 4, then throw a TypeError.

  8. If options.strides does not exist, set it to the list « 1, 1 ».

  9. Else if options.strides's size is not 2, then throw a TypeError.

  10. If any element in options.strides is equal to 0, then throw a TypeError.

  11. If options.dilations does not exist, set it to the list « 1, 1 ».

  12. Else if options.dilations's size is not 2, then throw a TypeError.

  13. If options.groups is 0, then throw a TypeError.

  14. Calculate the output shape:

    1. Switch on options.inputLayout:

      "nchw"
      1. Let batches be inputShape[0].

      2. Let inputChannels be inputShape[1].

      3. Let inputHeight be inputShape[2].

      4. Let inputWidth be inputShape[3].

      "nhwc"
      1. Let batches be inputShape[0].

      2. Let inputHeight be inputShape[1].

      3. Let inputWidth be inputShape[2].

      4. Let inputChannels be inputShape[3].

    2. Let filterShape be filter’s shape.

    3. Switch on options.filterLayout:

      "hwio"
      1. Let filterHeight be filterShape[0].

      2. Let filterWidth be filterShape[1].

      3. Let filterInputChannels be filterShape[2].

      4. Let outputChannels be filterShape[3].

      "ohwi"
      1. Let outputChannels be filterShape[0].

      2. Let filterHeight be filterShape[1].

      3. Let filterWidth be filterShape[2].

      4. Let filterInputChannels be filterShape[3].

      "ihwo"
      1. Let filterInputChannels be filterShape[0].

      2. Let filterHeight be filterShape[1].

      3. Let filterWidth be filterShape[2].

      4. Let outputChannels be filterShape[3].

      "oihw"
      1. Let outputChannels be filterShape[0].

      2. Let filterInputChannels be filterShape[1].

      3. Let filterHeight be filterShape[2].

      4. Let filterWidth be filterShape[3].

    4. If inputChannels % options.groups is not 0, then throw a TypeError.

    5. Else if inputChannels / options.groups is not equal to filterInputChannels, then throw a TypeError.

    6. If options.bias exists:

      1. If its rank is not 1, then throw a TypeError.

      2. If its shape[0] is not equal to outputChannels, then throw a TypeError.

      3. If its dataType is not equal to input’s dataType, then throw a TypeError.

    7. Let outputSizes be the result of calculating conv2d output sizes given inputHeight, inputWidth, filterHeight, filterWidth, options.padding, options.strides, and options.dilations.

    8. Switch on options.inputLayout:

      "nchw"

      Let outputShape be « batches, outputChannels, floor( outputSizes[0] ), floor( outputSizes[1] ) ».

      "nhwc"

      Let outputShape be « batches, floor( outputSizes[0] ), floor( outputSizes[1] ), outputChannels ».

    9. If any item in outputShape is not a valid dimension, then throw a TypeError.

    10. Let desc be a new MLOperandDescriptor.

    11. Set desc.dataType to input’s dataType.

    12. Set desc.dimensions to outputShape.

  15. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the conv2d operation, given options and filter.

    3. If options.activation exists, then add it to operator’s activations.

    4. Set output.[[operator]] to operator.

    5. Set operator’s inputs to input and filter.

    6. If options.bias exists, then add it to operator’s inputs.

    7. Set operator’s output to output.

  16. Return output.

7.8.11. convTranspose2d

Compute a 2-D transposed convolution given 4-D input and filter tensors
enum MLConvTranspose2dFilterOperandLayout {
  "iohw",
  "hwoi",
  "ohwi"
};

dictionary MLConvTranspose2dOptions {
  sequence<[EnforceRange] unsigned long> padding;
  sequence<[EnforceRange] unsigned long> strides;
  sequence<[EnforceRange] unsigned long> dilations;
  sequence<[EnforceRange] unsigned long> outputPadding;
  sequence<[EnforceRange] unsigned long> outputSizes;
  [EnforceRange] unsigned long groups = 1;
  MLInputOperandLayout inputLayout = "nchw";
  MLConvTranspose2dFilterOperandLayout filterLayout = "iohw";
  MLOperand bias;
  MLActivation activation;
};

partial interface MLGraphBuilder {
  MLOperand convTranspose2d(MLOperand input, MLOperand filter,
                            optional MLConvTranspose2dOptions options = {});
};

MLConvTranspose2dOptions has the following members:

padding, of type sequence<[EnforceRange] unsigned long>

A list of length 4: [beginningHeight, endingHeight, beginningWidth, endingWidth]. Specifies the additional rows and columns added to the beginning and ending of each spatial dimension of the convolution input. The default value is [0, 0, 0, 0].

strides, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [strideHeight, strideWidth]. Specifies the stride of the sliding window for each spatial dimension of the convolution input. The default value is [1, 1].

dilations, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [dilationHeight, dilationWidth]. Specifies the dilation factor for each spatial dimension applied on the convolution filter (kernel). The default value is [1, 1].

outputPadding, of type sequence<[EnforceRange] unsigned long>

A list of length 2. Specifies the padding values applied to each spatial dimension of the output tensor. The explicit padding values are needed to disambiguate the output tensor shape for transposed convolution when the value of the options.strides is greater than 1.

Note that these values are only used to disambiguate output shape when needed; it does not necessarily cause any padding value to be written to the output tensor.

The default value is [0, 0].

outputSizes, of type sequence<[EnforceRange] unsigned long>

A list of length 2. Specifies the sizes of the last two dimensions of the output tensor. When the output sizes are explicitly specified, the output padding values in outputPadding are ignored.

If not specified, the output sizes are automatically computed.

groups, of type unsigned long, defaulting to 1

The number of groups that input channels and output channels are divided into.

inputLayout, of type MLInputOperandLayout, defaulting to "nchw"

Specifies the layout format of the input and output tensor as follows:

  • "nchw"

    • input tensor: [batches, inputChannels, height, width]

    • output tensor: [batches, outputChannels, height, width]

  • "nhwc":

    • input tensor: [batches, height, width, inputChannels]

    • output tensor: [batches, height, width, outputChannels]

filterLayout, of type MLConvTranspose2dFilterOperandLayout, defaulting to "iohw"

Specifies the layout format of the filter tensor as follow:

  • "iohw": [inputChannels, outputChannels/groups, height, width]

  • "hwoi": [height, width, outputChannels/groups, inputChannels]

  • "ohwi": [outputChannels/groups, height, width, inputChannels]

bias, of type MLOperand

An additional 1-D tensor with the shape of [outputChannels] whose values are to be added to the convolution result.

activation, of type MLActivation

An optional activation function that immediately follows the convolution operation.

Arguments:

Returns: an MLOperand. The output 4-D tensor that contains the transposed convolution result. The output shape is interpreted according to the options.inputLayout value. More specifically, unless the options.outputSizes values are explicitly specified, the options.outputPadding may be needed to compute the spatial dimension values of the output tensor as follow:

outputSize = (inputSize - 1) * stride + (filterSize - 1) * dilation + 1 - beginningPadding - endingPadding + outputPadding

To calculate convtranspose output size given unsigned integers inputSize, filterSize, beginningPadding, endingPadding, stride, dilation, and outputPadding, perform these steps. They return a number.
  1. Let effectiveFilterSize be ( filterSize - 1 ) * dilation + 1.

  2. Let outputSize be ( inputSize - 1 ) * stride + effectiveFilterSize - beginningPadding - endingPadding + outputPadding.

  3. Return outputSize.

To calculate convtranspose2d output sizes given unsigned integers inputHeight, inputWidth, filterHeight and filterWidth, list of 4 unsigned integers padding, list of 2 unsigned integers strides, list of 2 unsigned integers dilations, and list of 2 unsigned integers outputPadding, perform these steps. They return a list of 2 numbers.
  1. Let outputHeight be the result of calculating convtranspose output size given inputHeight, filterHeight, padding[0], padding[1], strides[0], dilations[0], and outputPadding[0].

  2. Let outputWidth be the result of calculating convtranspose output size given inputWidth, filterWidth, padding[2], padding[3], strides[1], dilations[1] and outputPadding[1].

  3. Return « outputHeight, outputWidth ».

The convTranspose2d(input, filter, options) method steps are:
  1. If validating operand with this and any of input, filter, and options.bias (if it exists) returns false, then throw a TypeError.

  2. If options.activation exists, and validating activation with this and it returns false, then throw a TypeError.

  3. If input’s rank is not 4, then throw a TypeError.

  4. If filter’s rank is not 4, then throw a TypeError.

  5. If filter’s dataType is not equal to input’s dataType, then throw a TypeError.

  6. If options.padding does not exist, set it to the list « 0, 0, 0, 0 ».

  7. Else if options.padding's size is not 4, then throw a TypeError.

  8. If options.strides does not exist, set it to the list « 1, 1 ».

  9. Else if options.strides's size is not 2, then throw a TypeError.

  10. If any element in options.strides is equal to 0, then throw a TypeError.

  11. If options.dilations does not exist, set it to the list « 1, 1 ».

  12. Else if options.dilations's size is not 2, then throw a TypeError.

  13. If options.outputPadding does not exist, set it to the list « 0, 0 ».

  14. Else if options.outputPadding's size is not 2, then throw a TypeError.

  15. If options.outputSizes exists:

    1. If its size is not 2, then throw a TypeError.

  16. Otherwise:

    1. If options.outputPadding[0] is greater than or equal to options.strides[0], or options.outputPadding[1] is greater than or equal to options.strides[1], then throw a TypeError.

  17. Calculate the output shape:

    1. Switch on options.inputLayout:

      "nchw"
      1. Let batches be inputShape[0].

      2. Let inputChannels be inputShape[1].

      3. Let inputHeight be inputShape[2].

      4. Let inputWidth be inputShape[3].

      "nhwc"
      1. Let batches be inputShape[0].

      2. Let inputHeight be inputShape[1].

      3. Let inputWidth be inputShape[2].

      4. Let inputChannels be inputShape[3].

    2. Let filterShape be filter’s shape.

    3. Switch on options.filterLayout:

      "iohw"
      1. Let filterInputChannels be filterShape[0].

      2. Let filterOutputChannels be |filterShape[1].

      3. Let filterHeight be filterShape[2].

      4. Let filterWidth be filterShape[3].

      "hwoi"
      1. Let filterHeight be filterShape[0].

      2. Let filterWidth be filterShape[1].

      3. Let filterOutputChannels be |filterShape[2].

      4. Let filterInputChannels be filterShape[3].

      "ohwi"
      1. Let filterOutputChannels be |filterShape[0].

      2. Let filterHeight be filterShape[1].

      3. Let filterWidth be filterShape[2].

      4. Let filterInputChannels be filterShape[3].

    4. If inputChannels is not equal to filterInputChannels, then throw a TypeError.

    5. Let outputChannels be filterOutputChannels * options.groups

    6. If options.bias exists:

      1. If its rank is not 1, then throw a TypeError.

      2. If its shape[0] is not equal to outputChannels, then throw a TypeError.

      3. If its dataType is not equal to input’s dataType, then throw a TypeError.

    7. If options.outputSizes exists, let outputSizes be options.outputSizes.

    8. Else let outputSizes be the result of calculating convtranspose2d output sizes given inputHeight, inputWidth, filterHeight, filterWidth, options.padding, options.strides, options.dilations, and options.outputPadding.

    9. Switch on options.inputLayout:

      "nchw"

      Let outputShape be « batches, outputChannels, floor( outputSizes[0] ), floor( outputSizes[1] ) ».

      "nhwc"

      Let outputShape be « batches, floor( outputSizes[0] ), floor( outputSizes[1] ), outputChannels ».

    10. If any item in outputShape is not a valid dimension, then throw a TypeError.

    11. Let desc be a new MLOperandDescriptor.

    12. Set desc.dataType to input’s dataType.

    13. Set desc.dimensions to outputShape.

  18. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the convTranspose2d operation, given options and filter.

    3. If options.activation exists, then add it to operator’s activations.

    4. Set output.[[operator]] to operator.

    5. Set operator’s inputs to input and filter.

    6. If options.bias exists, then add it to operator’s inputs.

    7. Set operator’s output to output.

  19. Return output.

7.8.12. Element-wise binary operations

Compute the element-wise binary addition, subtraction, multiplication, division, power, maximum and minimum of the two input tensors.

The element-wise binary operations will be broadcasted according to [numpy-broadcasting-rule]. The rank of the output tensor is the maximum rank of the input tensors. For each dimension of the output tensor, its size is the maximum size along that dimension of the input tensors.

partial interface MLGraphBuilder {
  MLOperand add(MLOperand a, MLOperand b);
  MLOperand sub(MLOperand a, MLOperand b);
  MLOperand mul(MLOperand a, MLOperand b);
  MLOperand div(MLOperand a, MLOperand b);
  MLOperand max(MLOperand a, MLOperand b);
  MLOperand min(MLOperand a, MLOperand b);
  MLOperand pow(MLOperand a, MLOperand b);
};
Arguments:

Returns: an MLOperand. The output tensor that contains the result of element-wise binary operation of the two input tensors.

Operation types:
To create element-wise binary operation given string op, MLOperand a and MLOperand b, run the following steps:
  1. Assert: op is one of "add", "sub", "mul", "div", "max", "min", "pow".

  2. If validating operand with this and any of a and b returns false, then throw a TypeError.

  3. If a’s dataType is not equal to b’s dataType, then throw a TypeError.

  4. Let descriptor be a new MLOperandDescriptor.

  5. Set descriptor.dataType to a’s dataType.

  6. Set descriptor.dimensions to the result of bidirectionally broadcasting the shapes a’s shape and b’s shape.

    1. If that returns failure, then throw a TypeError.

  7. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and descriptor.

    2. Let operator be an operator for the binary operation op, given a and b.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to a and b.

    5. Set operator’s output to output.

  8. Return output.

The element-wise binary operation algorithms invoke the create element-wise binary operation steps as follows.
The add(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "add", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The sub(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "sub", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The mul(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "mul", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The div(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "div", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The max(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "max", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The min(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "min", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The pow(a, b) method steps are:
  1. Let output be the result of running the create element-wise binary operation given "pow", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

7.8.13. Element-wise logical operations

Compare input tensors element-wise and return a uint8 tensor of values 0 or 1 for the comparisons. For single-operand operations, return the logical results of the operation.

The input tensor will be broadcasted according to [numpy-broadcasting-rule]. The rank of the output tensor is the maximum rank of the input tensors.

partial interface MLGraphBuilder {
  MLOperand equal(MLOperand a, MLOperand b);
  MLOperand greater(MLOperand a, MLOperand b);
  MLOperand greaterOrEqual(MLOperand a, MLOperand b);
  MLOperand lesser(MLOperand a, MLOperand b);
  MLOperand lesserOrEqual(MLOperand a, MLOperand b);
  MLOperand not(MLOperand a);
};
Arguments:

Returns: an MLOperand. The output tensor that contains the result of element-wise comparison of the two input tensors.

Operation types:
Although operations greaterOrEqual() and lesserOrEqual() can each be implemented in terms of operations not(), lesser(), and greater() in other words builder.greaterOrEqual(a, b) is builder.not(builder.lesser(a, b)), they are specifically defined to handle NaN cases and for performance reason to avoid double comparisons.
To create element-wise logical operation given string op, MLOperand a and an optional MLOperand b, run the following steps:
  1. Assert: op is one of "equal", "greater", "greaterOrEqual", "lesser", "lesserOrEqual", "not".

  2. If op is "not".

    1. If validating operand with this and a returns false, then throw a TypeError.

    2. If a’s dataType is not "uint8", then throw a TypeError.

  3. If op is anything else but "not".

    1. If validating operand with this and any of a and b returns false, then throw a TypeError.

    2. If a’s dataType is not equal to b’s dataType, then throw a TypeError.

  4. Let descriptor be a new MLOperandDescriptor.

  5. Set descriptor.dataType to "uint8".

  6. Set descriptor.dimensions to the result of bidirectionally broadcasting the shapes a’s shape and b’s shape.

    1. If that returns failure, then throw a TypeError.

  7. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and descriptor.

    2. Let operator be an operator for the logical operation op, given a and (if op is not "not") b.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to a and (if op is anything other than "not") b.

    5. Set operator’s output to output.

  8. Return output.

The element-wise logical operation algorithms invoke the create element-wise logical operation steps as follows.
The equal(a, b) method steps are:
  1. Let output be the result of running the create element-wise logical operation given "equal", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The greater(a, b) method steps are:
  1. Let output be the result of running the create element-wise logical operation given "greater", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The greaterOrEqual(a, b) method steps are:
  1. Let output be the result of running the create element-wise logical operation given "greaterOrEqual", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The lesser(a, b) method steps are:
  1. Let output be the result of running the create element-wise logical operation given "lesser", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The lesserOrEqual(a, b) method steps are:
  1. Let output be the result of running the create element-wise logical operation given "lesserOrEqual", a and b.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The not(a) method steps are:
  1. Let output be the result of running the create element-wise logical operation given "not" and a.

    1. If that throws an error, then re-throw the error.

  2. Return output.

7.8.14. Element-wise unary operations

Compute the element-wise unary operation for input tensor.
partial interface MLGraphBuilder {
  MLOperand abs(MLOperand input);
  MLOperand ceil(MLOperand input);
  MLOperand cos(MLOperand input);
  MLOperand erf(MLOperand input);
  MLOperand exp(MLOperand input);
  MLOperand floor(MLOperand input);
  MLOperand identity(MLOperand input);
  MLOperand log(MLOperand input);
  MLOperand neg(MLOperand input);
  MLOperand reciprocal(MLOperand input);
  MLOperand sin(MLOperand input);
  MLOperand sqrt(MLOperand input);
  MLOperand tan(MLOperand input);
};
Arguments:

Returns: an MLOperand. The output tensor that contains the result of element-wise unary operation of the input tensor. The shape of the output tensor is the same as the shape of input tensor.

Operation types:
To create element-wise unary operation given string op and MLOperand input, run the following steps:
  1. Assert: op is one of "abs", "ceil", "cos", "erf", "exp", "floor", "identity", "log", "neg", "reciprocal", "sin", "sqrt", "tan".

  2. If validating operand with this and input returns false, then throw a TypeError.

  3. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the unary operation op.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  4. Return output.

The element-wise unary operation algorithms invoke the create element-wise unary operation steps as follows.
The abs(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "abs" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The ceil(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "ceil" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The cos(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "cos" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The erf(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "erf" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The exp(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "exp" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The floor(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "floor" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The identity(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "identity" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The log(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "log" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The neg(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "neg" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reciprocal(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "reciprocal" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The sin(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "sin" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The sqrt(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "sqrt" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The tan(input) method steps are:
  1. Let output be the result of running the create element-wise unary operation given "tan" and input.

    1. If that throws an error, then re-throw the error.

  2. Return output.

7.8.15. elu

Calculate the exponential linear unit function (ELU) on the input tensor element-wise. The calculation follows the expression max(0, x) + alpha * (exp(min(0, x)) - 1).
dictionary MLEluOptions {
  float alpha = 1;
};

partial interface MLGraphBuilder {
  MLOperand elu(MLOperand input, optional MLEluOptions options = {});
  MLActivation elu(optional MLEluOptions options = {});
};

MLEluOptions has the following members:

alpha, of type float, defaulting to 1

A scalar multiplier.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.add(
          builder.max(builder.constant(0), x),
          builder.mul(
            builder.constant(options.alpha),
            builder.sub(
              builder.exp(builder.min(builder.constant(0), x)),
              builder.constant(1))));
7.8.15.1. elu(input, options)
Arguments:

Returns:

The elu(input, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the ELU operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.15.2. elu(options)
Arguments:

Returns:

The elu(options) method steps are:
  1. Let op be the result of creating an MLActivation given this, "elu" and options.

  2. Return op.

7.8.16. expand

Expand any dimension of size 1 of the input tensor to a larger size according to the new shape. The expansion is consistent with [numpy-broadcasting-rule]. The input dimensions must have the size of 1 or match the sizes of the corresponding output dimensions according to the new shape.
partial interface MLGraphBuilder {
  MLOperand expand(MLOperand input, sequence<[EnforceRange] unsigned long> newShape);
};
Arguments:

Returns: an MLOperand. The tensor with expanded size dimensions.

The expand(input, newShape) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Let outputDescriptor be a new MLOperandDescriptor.

  3. Set outputDescriptor.dataType to input’s dataType.

  4. Set outputDescriptor.dimensions to the result of unidirectionally broadcasting the shapes input’s shape and newShape.

    1. If that returns failure, then throw a TypeError.

  5. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and outputDescriptor.

    2. Let operator be an operator for the expand operation, given input and newShape.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  6. Return output.

7.8.17. gather

Gather values of the input tensor along an axis according to the indices.
dictionary MLGatherOptions {
  [EnforceRange] unsigned long axis = 0;
};

partial interface MLGraphBuilder {
  MLOperand gather(MLOperand input,
                   MLOperand indices,
                   optional MLGatherOptions options = {});
};

MLGatherOptions has the following members:

axis, of type unsigned long, defaulting to 0

The axis along which the gathered values are obtained. Its value must be in the range [0, N-1] where N is the rank of the input tensor.

Arguments:

Returns: an MLOperand. The output N-D tensor of rank equal to the rank of input + the rank of indices - 1.

The indices parameter to gather() can not be clamped to the allowed range when the graph is built because the inputs are not known until execution. Implementations can introduce clamp() in the compiled graph if the required clamping behavior is not provided by the underlying platform. Similarly, if the underlying platform does not support negative indices, the implementation can introduce operations in the compiled graph to transform a negative index from the end of the dimension into a positive index.
The gather(input, indices, options) method steps are:
  1. If validating operand with this and any of input and indices returns false, then throw a TypeError.

  2. If indices’s dataType is not "uint32" or "int64", then throw a TypeError.

  3. Let shapeInput be input’s shape and rankInput be shapeInput’s rank.

  4. Let shapeIndices be indices’s shape.

  5. Let axis be options.axis.

  6. If axis is greater than or equal to rankInput, then throw a TypeError.

  7. Let dimCount be zero.

  8. Let rankOutput be zero.

  9. Let shapeOutput be an empty list.

  10. For each size of shapeInput:

    1. If dimCount is equal to axis then break.

    2. Set shapeOutput[dimCount] to size.

    3. Increment dimCount by one.

  11. Set rankOutput to dimCount.

  12. Let dimCount be zero.

  13. For each size of shapeIndices:

    1. Set shapeOutput[rankOutput + dimCount] to size.

    2. Increment dimCount by one.

  14. Set rankOutput to rankOutput + dimCount.

  15. Let dimCount be zero.

  16. For each size of shapeInput:

    1. If dimCount is less than or equal to axis then continue.

    2. Set shapeOutput[rankOutput + dimCount - axis - 1] to size.

    3. Increment dimCount by one.

  17. Let desc be a new MLOperandDescriptor.

  18. Set desc.dimensions to shapeOutput.

  19. Set desc.dataType to input’s dataType.

  20. Make graph connections:

    1. Let output be the result of creating an MLOperand given desc.

    2. Let operator be an operator for the Gather operation, given input, indices, and options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to input and indices.

    5. Set operator’s output to output.

  21. Return output.

Examples of how gather works in different slicing schemes.
// input of shape [4,3]:
//   [[ 0,  1,  2],
//    [10, 11, 12], 
//    [20, 21, 22], 
//    [30, 31, 32]]
const input = builder.constant(
{ dimensions: [4,3] }, new Float32Array([0,1,2,10,11,12,20,21,22,30,31,32]));

const indices1 = builder.constant(
{ dataType: 'uint32', dimensions: [2] }, new Uint32Array([3,1]));

const indices2 = builder.constant(
{ dataType: 'uint32', dimensions: [3] }, new Uint32Array([2,1,1]));

const indices3 = builder.constant(
{ dataType: 'uint32', dimensions: [2,2] }, new Uint32Array([0,1,1,2]));

// axis = 0 (default)
// indices of shape [2]: 
//   [3,1]
// output of shape [2,3]:
//   [[30, 31, 32], 
//    [10, 11, 12]]
const output1 = builder.gather(input, indices1);

// axis = 1
// indices of shape [3]:
//   [2,1,1]
// output of shape [4,3]:
//   [[ 2,  1,  1],
//    [12, 11, 11], 
//    [22, 21, 21],
//    [32, 31, 31]]
const output2 = builder.gather(input, indices2, { axis: 1 });

// axis = 1
// indices of shape [2,2]: 
//   [[0, 1], 
//    [1, 2]]
// output of shape [4,2,2]:
//   [[[ 0,  1], [ 1,  2]],
//    [[10, 11], [11, 12]],
//    [[20, 21], [21, 22]],
//    [[30, 31], [31, 32]]]
const output3 = builder.gather(input, indices3, { axis: 1 });

7.8.18. gelu

Compute the gaussian error linear unit function (GELU) of the input tensor. The calculation follows the expression 0.5 * x * (1 + erf(x / sqrt(2))).
partial interface MLGraphBuilder {
  MLOperand gelu(MLOperand input);
  MLActivation gelu();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.mul(
           builder.mul(x, builder.constant(0.5)),
           builder.add(
               builder.constant(1),
               builder.erf(
                   builder.div(
                       x,
                       builder.sqrt(builder.constant(2))))));
7.8.18.1. gelu(input)
Arguments:

Returns:

The gelu(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the gelu operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.18.2. gelu()
Arguments:

Returns:

The gelu() method steps are:
  1. Let op be the result of creating an MLActivation given this and "gelu".

  2. Return op.

7.8.19. gemm

Calculate the general matrix multiplication of the Basic Linear Algebra Subprograms. The calculation follows the expression alpha * A * B + beta * C, where A is a 2-D tensor with shape [M, K] or [K, M], B is a 2-D tensor with shape [K, N] or [N, K], and C is unidirectionally broadcastable to the shape [M, N]. A and B may optionally be transposed prior to the calculation.
dictionary MLGemmOptions {
  MLOperand c;
  float alpha = 1.0;
  float beta = 1.0;
  boolean aTranspose = false;
  boolean bTranspose = false;
};

partial interface MLGraphBuilder {
  MLOperand gemm(MLOperand a, MLOperand b, optional MLGemmOptions options = {});
};

MLGemmOptions has the following members:

c, of type MLOperand

The third input tensor. It is either a scalar, or of the shape that is unidirectionally broadcastable to the shape [M, N]. When it is not specified, the computation is done as if c is a scalar 0.0.

alpha, of type float, defaulting to 1.0

A multiplier for the first input.

beta, of type float, defaulting to 1.0

A multiplier for the third input c.

aTranspose, of type boolean, defaulting to false

Indicates if the first input should be transposed prior to calculating the output.

bTranspose, of type boolean, defaulting to false

Indicates if the second input should be transposed prior to calculating the output.

Arguments:

Returns: an MLOperand. The output 2-D tensor of shape [M, N] that contains the calculated product of all the inputs.

The gemm(a, b, options) method steps are:
  1. If validating operand with this and any of a and b returns false, then throw a TypeError.

  2. If a’s rank is not 2 or b’s rank is not 2, then throw a TypeError.

  3. Let shapeA be a clone of a’s shape.

  4. Let shapeB be a clone of b’s shape.

  5. If options.aTranspose is true, then reverse the order of the items in shapeA.

  6. If options.bTranspose is true, then reverse the order of the items in shapeB.

  7. If shapeA[1] is not equal to shapeB[0], then throw a TypeError.

  8. If options.c exists and is not unidirectionally broadcastable to the shape « shapeA[0], shapeB[1] », then throw a TypeError.

    Type compatibility between a, b and options.c can be also checked.
  9. Let desc be a new MLOperandDescriptor.

  10. Set desc.dimensions to the list « shapeA[0], shapeB[1] ».

  11. Set desc.dataType to a’s dataType.

  12. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the GEMM operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to a and b.

    5. If options.c exists, then add it to operator’s inputs.

    6. Set operator’s output to output.

  13. Return output.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
if (options.aTranspose)
  a = builder.transpose(a);

if (options.bTranspose)
  b = builder.transpose(b);

let ab = builder.matmul(builder.mul(builder.constant(options.alpha), a), b);
return (c ? builder.add(ab, builder.mul(builder.constant(options.beta), c)) : ab);

7.8.20. gru

Gated Recurrent Unit [GRU] recurrent network uses an update, reset, and new gate to compute the output state that rolls into the output across the temporal sequence of the network.
enum MLGruWeightLayout {
  "zrn",  // update-reset-new gate ordering
  "rzn"   // reset-update-new gate ordering
};

enum MLRecurrentNetworkDirection {
  "forward",
  "backward",
  "both"
};

dictionary MLGruOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  MLOperand initialHiddenState;
  boolean resetAfter = true;
  boolean returnSequence = false;
  MLRecurrentNetworkDirection direction = "forward";
  MLGruWeightLayout layout = "zrn";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> gru(MLOperand input,
                          MLOperand weight,
                          MLOperand recurrentWeight,
                          [EnforceRange] unsigned long steps,
                          [EnforceRange] unsigned long hiddenSize,
                          optional MLGruOptions options = {});
};

MLGruOptions has the following members:

bias, of type MLOperand

The 2-D input bias tensor of shape [numDirections, 3 * hiddenSize]. The ordering of the bias vectors in the second dimension of the tensor shape is specified according to the layout argument.

recurrentBias, of type MLOperand

The 2-D recurrent bias tensor of shape [numDirections, 3 * hiddenSize]. The ordering of the bias vectors in the second dimension of the tensor shape is specified according to the layout argument.

initialHiddenState, of type MLOperand

The 3-D initial hidden state tensor of shape [numDirections, batchSize, hiddenSize]. When not specified, implementations SHOULD use a tensor filled with zero.

resetAfter, of type boolean, defaulting to true

Indicates whether to apply the reset gate after or before matrix multiplication.

returnSequence, of type boolean, defaulting to false

Indicates whether to also return the entire sequence with every output from each time step in it in addition to the output of the last time step.

direction, of type MLRecurrentNetworkDirection, defaulting to "forward"

The processing direction of the input sequence. When set to "both", the size of the first dimension of the weight and the bias tensor shapes must be 2, and the input is processed in both directions.

layout, of type MLGruWeightLayout, defaulting to "zrn"

The ordering of the weight and bias vectors for the internal gates of GRU, specifically the update (z), reset (r), and new (n) gate, as indicated in the second dimension of the weight and bias tensor shape.

activations, of type sequence<MLActivation>

Specifies a pair of activation functions with the first function used for the update and reset gate, and the second used for the new gate. When not specified, implementations SHOULD use the pair of sigmoid ("sigmoid") and the hyperbolic tangent ("tanh") functions, respectively.

Arguments:

Returns: a sequence of MLOperand. The first element of the sequence is a 3-D tensor of shape [numDirections, batchSize, hiddenSize], the cell output from the last time step of the network. Additionally, if options.returnSequence is set to true, the second element is the 4-D output tensor of shape [steps, numDirections, batchSize, hiddenSize] containing every cell outputs from each time step in the temporal sequence.

The gru(input, weight, recurrentWeight, steps, hiddenSize, options) method steps are:
  1. If validating operand with this and any of input, weight, recurrentWeight, options.bias (if it exists), options.recurrentBias (if it exists), and options.initialHiddenState (if it exists) returns false, then throw a TypeError.

  2. If options.activations exists, and validating activation with this and any item in it returns false, then throw a TypeError.

  3. If the rank of any of input, weight or recurrentWeight is not 3, then throw a TypeError.

  4. If options.bias exists:

    1. If its shape[1] is not equal to 3 * hiddenSize, then throw a TypeError.

  5. If options.recurrentBias exists:

    1. If its shape[1] is not equal to 3 * hiddenSize, then throw a TypeError.

  6. If options.initialHiddenState exists:

    1. If its rank is not 3, then throw a TypeError.

  7. If options.activations exists and its size is not 2, then throw a TypeError.

  8. If steps is not equal to input’s shape[0], then throw a TypeError.

  9. Let batchSize be input’s shape[1].

  10. Let numDirections be 2 if options.direction is "both", or 1 otherwise.

  11. Calculate the output shape:

    1. Let desc0 be a new MLOperandDescriptor.

    2. Set desc0.dimensions to the list « numDirections, batchSize, hiddenSize ».

    3. Set desc0.dataType to input’s dataType.

    4. If options.returnSequence is true:

      1. Let desc1 be a new MLOperandDescriptor.

      2. Set desc1.dataType to input’s dataType.

      3. Set desc1.dimensions to the list « steps, numDirections, batchSize, hiddenSize ».

  12. Make graph connections:

    1. Let operator be an operator for "gru", given weight, recurrentWeight, steps, hiddenSize and options as parameters.

    2. Let output0 be the result of creating an MLOperand given this and desc0.

    3. If options.returnSequence is true:

      1. Let output1 be the result of creating an MLOperand given this and desc1.

      2. Let output be the list « output0, output1 ».

      3. Set output0.[[operator]] and output1.[[operator]] to operator.

    4. Otherwise:

      1. Let output be the list « output0 ».

      2. Set output0.[[operator]] to operator.

    5. Set operator’s inputs to input, weight, and recurrentWeight.

    6. If options.bias exists, then add it to operator’s inputs.

    7. If options.recurrentBias exists, then add it to operator’s inputs.

    8. If options.initialHiddenState exists, then add it to operator’s inputs.

    9. If options.activations exists, then add its items to operator’s activation functions.

    10. Set operator’s output to output.

  13. Return output.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
function squeeze(builder, op) {
  return builder.reshape(op, op.shape().remove(0));
}

const numDirections = (options.direction == "both" ? 2 : 1);
let hiddenState = options.initialHiddenState;

if (!hiddenState) {
  const desc = { dataType: 'float32', dimensions: [numDirections, 1, hiddenSize] };
  const totalSize = numDirections * hiddenSize;
  hiddenState = builder.constant(desc, new Float32Array(totalSize).fill(0));
}

let sequence = null;
let currentWeight = [];
let currentRecurrentWeight = [];
let currentBias = [];
let currentRecurrentBias = [];

for (let dir = 0; dir < numDirections; ++dir) {
  currentWeight.push(squeeze(builder, builder.slice(weight, [dir, 0, 0], [1, 3 * hiddenSize, inputSize])));
  currentRecurrentWeight.push(squeeze(builder, builder.slice(recurrentWeight, [dir, 0, 0], [1, 3 * hiddenSize, hiddenSize])));
  currentBias.push(options.bias ? (squeeze(builder, builder.slice(options.bias, [dir, 0], [1, 3 * hiddenSize]))) : null);
  currentRecurrentBias.push(options.recurrentBias ?
    (squeeze(builder, builder.slice(options.recurrentBias, [dir, 0], [1, 3 * hiddenSize]))) : null);
}

for (let step = 0; step < steps; ++step) {
  let currentHidden = [];
  let currentOutput = null;

  for (let dir = 0; dir < numDirections; ++dir) {
    currentHidden.push(squeeze(builder, builder.slice(hiddenState, [dir, 0, 0], [1, batchSize, hiddenSize])));
  }

  for (let dir = 0; dir < numDirections; ++dir) {
    let slice = (dir == 1 || options.direction == "backward" ? steps - step - 1 : step);
    let currentInput = squeeze(builder, builder.slice(input, [slice, 0, 0], [1, batchSize, inputSize]));

    let result = builder.reshape(
      builder.gruCell(
        currentInput, currentWeight[dir], currentRecurrentWeight[dir],
        currentHidden[dir], hiddenSize, { bias: currentBias[dir],
        recurrentBias: currentRecurrentBias[dir], resetAfter: options.resetAfter,
        layout: options.layout, activations: options.activations }),
      [1, null, hiddenSize]);

    currentOutput = (currentOutput ? builder.concat([currentOutput, result], 0) : result);
  }

  hiddenState = currentOutput;

  if (options.returnSequence) {
    currentOutput = builder.reshape(currentOutput, [1, numDirections, null, hiddenSize]);
    sequence = (sequence ? builder.concat([sequence, currentOutput], 0) : currentOutput);
  }
}

return (sequence ? [hiddenState, sequence] : [hiddenState]);

7.8.21. gruCell

A single time step of the Gated Recurrent Unit [GRU] recurrent network using an update gate and a reset gate to compute the hidden state that rolls into the output across the temporal sequence of a recurrent network.
dictionary MLGruCellOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  boolean resetAfter = true;
  MLGruWeightLayout layout = "zrn";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  MLOperand gruCell(MLOperand input,
                    MLOperand weight,
                    MLOperand recurrentWeight,
                    MLOperand hiddenState,
                    [EnforceRange] unsigned long hiddenSize,
                    optional MLGruCellOptions options = {});
};

MLGruCellOptions has the following members:

bias, of type MLOperand

The 1-D input bias tensor of shape [3 * hiddenSize]. The ordering of the bias vectors in the second dimension of the tensor shape is specified according to the layout argument.

recurrentBias, of type MLOperand

The 1-D recurrent bias tensor of shape [3 * hiddenSize]. The ordering of the bias vectors in the second dimension of the tensor shape is specified according to the layout argument.

resetAfter, of type boolean, defaulting to true

Indicates whether to apply the reset gate after or before matrix multiplication.

layout, of type MLGruWeightLayout, defaulting to "zrn"

The ordering of the weight and bias vectors for the internal gates of GRU, specifically the update (z), reset (r), and new (n) gate, as indicated in the second dimension of the weight and bias tensor shape.

activations, of type sequence<MLActivation>

Specifies a pair of activation functions with the first function used for the update and reset gate, and the second used for the new gate. When not specified, implementations SHOULD use the pair of sigmoid ("sigmoid") and the hyperbolic tangent ("tanh") functions, respectively.

Arguments:

Returns: an MLOperand. The 2-D tensor of shape [batchSize, hiddenSize], the cell output hidden state of a single time step of the recurrent network.

The gruCell(input, weight, recurrentWeight, hiddenState, hiddenSize, options) method steps are:
  1. If validating operand with this and any of input, weight, recurrentWeight, hiddenState, options.bias (if it exists), and options.recurrentBias (if it exists) returns false, then throw a TypeError.

  2. If options.activations exists, and validating activation with this and any item in it returns false, then throw a TypeError.

  3. If the rank of any of input, weight, recurrentWeight or hiddenState is not 2, then throw a TypeError.

  4. If weight’s shape[0] is not equal to 3 * hiddenSize, then throw a TypeError.

  5. If recurrentWeight’s shape[0] is not equal to 3 * hiddenSize, then throw a TypeError.

  6. If options.bias exists:

    1. If its rank is not equal to 3 * hiddenSize, then throw a TypeError.

  7. If options.recurrentBias exists:

    1. If its rank is not equal to 3 * hiddenSize, then throw a TypeError.

  8. If options.activations exists and its size is not 2, then throw a TypeError.

  9. Let desc be a new MLOperandDescriptor.

  10. Set desc.dimensions to the list « input’s shape[0], hiddenSize ».

  11. Set desc.dataType to input’s dataType.

  12. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for "gruCell", given weight, recurrentWeight, hiddenState, hiddenSize and options as parameters.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to input, weight, recurrentWeight, and hiddenState.

    5. If options.bias exists, then add it to operator’s inputs.

    6. If options.recurrentBias exists, then add it to operator’s inputs.

    7. If options.activations exists, then add its items to operator’s activation functions.

    8. Set operator’s output to output.

  13. Return output.

The behavior of this operation when the weight layout is the default "zrn" layout, and the activation functions of the update/reset gate and new gate are sigmoid() and tanh() respectively can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
const one = builder.constant(1);
const zero = builder.constant(0);

// update gate (z)
let z = builder.sigmoid(
  builder.add(
    builder.add(
      (options.bias ? builder.slice(options.bias, [0], [hiddenSize]) : zero),
      (options.recurrentBias ? builder.slice(options.recurrentBias, [0], [hiddenSize]) : zero)
      ),
    builder.add(
      builder.matmul(
        input,
        builder.transpose(builder.slice(weight, [0, 0], [hiddenSize, inputSize]))
        ),
      builder.matmul(
        hiddenState,
        builder.transpose(builder.slice(recurrentWeight, [0, 0], [hiddenSize, hiddenSize]))
        )
      )
    )
  );

// reset gate (r)
let r = builder.sigmoid(
  builder.add(
    builder.add(
      (options.bias ? builder.slice(options.bias, [hiddenSize], [hiddenSize]) : zero),
      (options.recurrentBias ? builder.slice(options.recurrentBias, [hiddenSize], [hiddenSize]) : zero)
      ),
    builder.add(
      builder.matmul(
        input,
        builder.transpose(builder.slice(weight, [hiddenSize, 0], [hiddenSize, inputSize]))
        ),
      builder.matmul(
        hiddenState,
        builder.transpose(builder.slice(recurrentWeight, [hiddenSize, 0], [hiddenSize, hiddenSize]))
        )
      )
    )
  );

// new gate (n)
let n;
if (resetAfter) {
  n = builder.tanh(
    builder.add(
      (options.bias ? builder.slice(options.bias, [2 * hiddenSize], [hiddenSize]) : zero),
      builder.add(
        builder.matmul(
          input,
          builder.transpose(builder.slice(weight, [2 * hiddenSize, 0], [hiddenSize, inputSize]))
          ),
        builder.mul(
          r,
          builder.add(
            (options.recurrentBias ? builder.slice(options.recurrentBias, [2 * hiddenSize], [hiddenSize]) : zero),
            builder.matmul(
              hiddenState,
              builder.transpose(builder.slice(recurrentWeight, [2 * hiddenSize, 0], [hiddenSize, hiddenSize]))
              )
            )
          )
        )
      )
    );
}
else {
  n = builder.tanh(
    builder.add(
      builder.add(
        (options.bias ? builder.slice(options.bias, [2 * hiddenSize], [hiddenSize]) : zero),
        (options.recurrentBias ? builder.slice(options.recurrentBias, [2 * hiddenSize], [hiddenSize]) : zero)
        ),
      builder.add(
        builder.matmul(
          input,
          builder.transpose(builder.slice(weight, [2 * hiddenSize, 0], [hiddenSize, inputSize]))
          ),
        builder.matmul(
          builder.mul(r, hiddenState),
          builder.transpose(builder.slice(recurrentWeight, [2 * hiddenSize, 0], [hiddenSize, hiddenSize]))
          )
        )
      )
    );
}

// compute the new hidden state
return builder.add(builder.mul(z, hiddenState), builder.mul(n, builder.sub(one, z)));

7.8.22. hardSigmoid

Calculate the non-smooth hard sigmoid function on the input tensor, used instead of the sigmoid function for faster computation.
dictionary MLHardSigmoidOptions {
  float alpha = 0.2;
  float beta = 0.5;
};

partial interface MLGraphBuilder {
  MLOperand hardSigmoid(MLOperand input, optional MLHardSigmoidOptions options = {});
  MLActivation hardSigmoid(optional MLHardSigmoidOptions options = {});
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.max(
           builder.min(
               builder.add(
                   builder.mul(builder.constant(options.alpha), x),
                   builder.constant(options.beta)),
               builder.constant(1)),
           builder.constant(0));

MLHardSigmoidOptions has the following members:

alpha, of type float, defaulting to 0.2

A scalar multiplier.

beta, of type float, defaulting to 0.5

A scalar addition.

7.8.22.1. hardSigmoid(input, options)
Arguments:

Returns:

The hardSigmoid(input, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the hard sigmoid operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.22.2. hardSigmoid(options)
Arguments:

Returns:

The hardSigmoid(options) method steps are:
  1. Let op be the result of creating an MLActivation given this, "hardSigmoid" and options.

  2. Return op.

7.8.23. hardSwish

Computes the nonlinear function y = x * max(0, min(6, (x + 3))) / 6 that is introduced by [MobileNetV3] on the input tensor element-wise.
partial interface MLGraphBuilder {
  MLOperand hardSwish(MLOperand input);
  MLActivation hardSwish();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.div(
           builder.mul(
               x,
               builder.max(
                   builder.constant(0),
                   builder.min(
                       builder.constant(6),
                       builder.add(x, builder.constant(3))))),
           builder.constant(6));
7.8.23.1. hardSwish(input)
Arguments:

Returns:

The hardSwish(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the hard-swish operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.23.2. hardSwish()
Arguments:

Returns:

The hardSwish() method steps are:
  1. Let op be the result of creating an MLActivation given this and "hardSwish".

  2. Return op.

7.8.24. instanceNormalization

Normalize the input using [Instance-Normalization]. Unlike batchNormalization() where the mean and variance values used in the normalization are computed across all the samples in the batch dimension while the model is trained, the mean and variance values used in the instance normalization are computed on the fly for each input feature of each individual sample in the batch.
dictionary MLInstanceNormalizationOptions {
  MLOperand scale;
  MLOperand bias;
  float epsilon = 1e-5;
  MLInputOperandLayout layout = "nchw";
};

partial interface MLGraphBuilder {
  MLOperand instanceNormalization(MLOperand input,
                                  optional MLInstanceNormalizationOptions options = {});
};

MLInstanceNormalizationOptions has the following members:

scale, of type MLOperand

The 1-D tensor of the scaling values whose size is equal to the number of channels, i.e. the size of the feature dimension of the input. For example, for an input tensor with "nchw" layout, the size is equal to input’s shape[1].

bias, of type MLOperand

The 1-D tensor of the bias values whose size is equal to the size of the feature dimension of the input. For example, for an input tensor with "nchw" layout, the size is equal to input’s shape[1].

epsilon, of type float, defaulting to 1e-5

A small value to prevent computational error due to divide-by-zero.

layout, of type MLInputOperandLayout, defaulting to "nchw"

The layout format of the input.

Arguments:

Returns: an MLOperand. The instance-normalized 4-D tensor of the same shape as input.

The instanceNormalization(input, options) method steps are:
  1. If validating operand with this and any of input, options.scale (if it exists), and options.bias (if it exists) returns false, then throw a TypeError.

  2. If input’s rank is not 4, then throw a TypeError.

  3. If options.scale exists:

    1. If its rank is not equal to 1, then throw a TypeError.

  4. If options.bias exists:

    1. If its rank is not equal to 1, then throw a TypeError.

  5. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the instance normalization operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. If options.scale exists, then add it to operator’s inputs.

    6. If options.bias exists, then add it to operator’s inputs.

    7. Set operator’s output to output.

  6. Return output.

The behavior of this operation when the input tensor is 4-D of the "nchw" layout can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
// The reduction of the mean and variance values happens over the spatial dimensions of the input
// e.g. axis 2 and 3 of the input tensor.
const reduceOptions = { axes: [2,3], keepDimensions: true };
const mean = builder.reduceMean(input, reduceOptions);
const variance = builder.reduceMean(
  builder.pow(
    builder.sub(input, mean),
    buider.constant(2)),
  reduceOptions
  );

// The scale and bias values are applied per input feature
// e.g. axis 1 of the input tensor.
const shape = [1,null,1,1];
return builder.add(
  builder.mul(
    builder.reshape(options.scale, shape),
    builder.div(
      builder.sub(input, mean),
      buidler.sqrt(builder.add(variance, options.epsilon))
      )
    ),
  builder.reshape(options.bias, shape)
  );

7.8.25. layerNormalization

Normalize the input using [Layer-Normalization]. Unlike batchNormalization() where the mean and variance values are computed across all the samples in the batch dimension while the model is trained, and in instanceNormalization() where the mean and variance values are computed on the fly for each input feature of each individual sample in the batch, the means and variance values of the layer normalization are computed on the fly across all the input features of each individual sample in the batch.
dictionary MLLayerNormalizationOptions {
  MLOperand scale;
  MLOperand bias;
  sequence<[EnforceRange] unsigned long> axes;
  float epsilon = 1e-5;
};

partial interface MLGraphBuilder {
  MLOperand layerNormalization(MLOperand input,
                               optional MLLayerNormalizationOptions options = {});
};

MLLayerNormalizationOptions has the following members:

scale, of type MLOperand

The N-D tensor of the scaling values whose shape is determined by the axes member in that each value in axes indicates the dimension of the input tensor with scaling values. For example, for an axes values of [1,2,3], the shape of this tensor is the list of the corresponding sizes of the input dimension 1, 2 and 3. When this member is not present, the scaling value is assumed to be 1.

bias, of type MLOperand

The N-D tensor of the bias values whose shape is determined by the axes member in that each value in axes indicates the dimension of the input tensor with bias values. For example, for an axes values of [1,2,3], the shape of this tensor is the list of the corresponding sizes of the input dimension 1, 2 and 3. When this member is not present, the bias value is assumed to be 0.

axes, of type sequence<[EnforceRange] unsigned long>

The indices to the input dimensions to reduce. When this member is not present, it is treated as if all dimensions except the first were given (e.g. for a 4-D input tensor, axes = [1,2,3]). That is, the reduction for the mean and variance values are calculated across all the input features for each independent batch. If empty, no dimensions are reduced.

epsilon, of type float, defaulting to 1e-5

A small value to prevent computational error due to divide-by-zero.

Arguments:

Returns: an MLOperand. The layer-normalized N-D tensor of the same shape as input.

The layerNormalization(input, options) method steps are:
  1. If validating operand with this and any of input, options.scale (if it exists), and options.bias (if it exists) returns false, then throw a TypeError.

  2. If options.axes does not exist, then set options.axes to a new list, either equal to the range from 1 to input’s rank, exclusive, if input’s rank is greater than 1, or an empty list otherwise.

  3. If options.scale exists:

    1. If its rank is not equal to options.axes's size, then throw a TypeError.

  4. If options.bias exists:

    1. If its rank is not equal to options.axes's size, then throw a TypeError.

  5. For each index in the range 0 to options.axes's size, exclusive:

    1. Let axis be options.axes[index].

    2. If axis is greater or equal to input’s rank, then throw a TypeError.

    3. Let size be input’s shape[axis].

    4. If options.scale exists:

      1. If its shape[index] is not equal to size, then throw a TypeError.

    5. If options.bias exists:

      1. If its shape[index] is not equal to size, then throw a TypeError.

  6. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the layer normalization operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. If options.scale exists, then add it to operator’s inputs.

    6. If options.bias exists, then add it to operator’s inputs.

    7. Set operator’s output to output.

  7. Return output.

The behavior of this operation when the axes parameter is set to [1,2,3] can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
// The reduction of the mean and variance values happens over the spatial dimensions 
// across all the input features (i.e. all channels) of the input tensor.
const reduceOptions = { axes: [1,2,3], keepDimensions: true };
const mean = builder.reduceMean(input, reduceOptions);
const variance = builder.reduceMean(
  builder.pow(
    builder.sub(input, mean),
    buider.constant(2)),
  reduceOptions
  );

// The scale and bias tensors are of the shape of the input dimensions specified 
// by the values in the axes parameter (i.e. [1,2,3]).
return builder.add(
  builder.mul(
    options.scale,
    builder.div(
      builder.sub(input, mean),
      buidler.sqrt(builder.add(variance, options.epsilon))
      )
    ),
  options.bias
  );

7.8.26. leakyRelu

Calculate the leaky version of rectified linear function on the input tensor element-wise. The calculation follows the expression max(0, x) + alpha * min(0, x).
dictionary MLLeakyReluOptions {
  float alpha = 0.01;
};

partial interface MLGraphBuilder {
  MLOperand leakyRelu(MLOperand input, optional MLLeakyReluOptions options = {});
  MLActivation leakyRelu(optional MLLeakyReluOptions options = {});
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.add(builder.max(builder.constant(0), x),
          builder.mul(builder.constant(options.alpha), builder.min(builder.constant(0), x)));

MLLeakyReluOptions has the following members:

alpha, of type float, defaulting to 0.01

A scalar multiplier.

7.8.26.1. leakyRelu(input, options)
Arguments:

Returns:

The leakyRelu(input, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the Leaky RELU operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.26.2. leakyRelu(options)
Arguments:

Returns:

The leakyRelu(options) method steps are:
  1. Let op be the result of creating an MLActivation given this, "leakyRelu" and options.

  2. Return op.

7.8.27. linear

Calculate a linear function y = alpha * x + beta on the input tensor.
dictionary MLLinearOptions {
  float alpha = 1;
  float beta = 0;
};

partial interface MLGraphBuilder {
  MLOperand linear(MLOperand input, optional MLLinearOptions options = {});
  MLActivation linear(optional MLLinearOptions options = {});
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.add(
          builder.mul(x, builder.constant(options.alpha)),
          builder.constant(options.beta));

MLLinearOptions has the following members:

alpha, of type float, defaulting to 1

A scalar multiplier.

beta, of type float, defaulting to 0

A scalar addition.

7.8.27.1. linear(input, options)
Arguments:

Returns:

The linear(input, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the linear operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.27.2. linear(options)
Arguments:

Returns:

The linear(options) method steps are:
  1. Let op be the result of creating an MLActivation given this, "linear" and options.

  2. Return op.

7.8.28. lstm

Long Short-Term Memory [LSTM] recurrent network uses an input, output, forget, and cell gate to compute the output state that rolls into the output across the temporal sequence of the network.
enum MLLstmWeightLayout {
  "iofg", // input-output-forget-cell gate ordering
  "ifgo"  // input-forget-cell-output gate ordering
};

dictionary MLLstmOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  MLOperand peepholeWeight;
  MLOperand initialHiddenState;
  MLOperand initialCellState;
  boolean returnSequence = false;
  MLRecurrentNetworkDirection direction = "forward";
  MLLstmWeightLayout layout = "iofg";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> lstm(MLOperand input,
                           MLOperand weight,
                           MLOperand recurrentWeight,
                           [EnforceRange] unsigned long steps,
                           [EnforceRange] unsigned long hiddenSize,
                           optional MLLstmOptions options = {});
};

MLLstmOptions has the following members:

bias, of type MLOperand

The 2-D input bias tensor of shape [numDirections, 4 * hiddenSize]. The ordering of the bias vectors in the second dimension of the tensor shape is specified according to layout.

recurrentBias, of type MLOperand

The 2-D recurrent bias tensor of shape [numDirections, 4 * hiddenSize]. The ordering of the bias vectors in the first dimension of the tensor shape is specified according to layout.

peepholeWeight, of type MLOperand

The 2-D weight tensor for peepholes of shape [numDirections, 3 * hiddenSize]. The pack ordering of the weight vectors is for the input (i), output (o), and forget (f) gate, respectively.

initialHiddenState, of type MLOperand

The 3-D initial hidden state tensor of shape [numDirections, batchSize, hiddenSize]. When not specified, implementations SHOULD use a tensor filled with zero.

initialCellState, of type MLOperand

The 3-D initial hidden state tensor of shape [numDirections, batchSize, hiddenSize]. When not specified, implementations SHOULD use a tensor filled with zero.

returnSequence, of type boolean, defaulting to false

Indicates whether to also return the entire sequence with every output from each time step in it in addition to the output of the last time step.

direction, of type MLRecurrentNetworkDirection, defaulting to "forward"

The processing direction of the input sequence. When set to "both", the size of the first dimension of the weight and the bias tensor shapes must be 2, and the input is processed in both directions.

layout, of type MLLstmWeightLayout, defaulting to "iofg"

The ordering of the weight and bias vectors for the internal gates of LSTM, specifically the input (i), output (o), forget (f), and cell (g) gate, as indicated in the first dimension of the weight and bias tensor shapes.

activations, of type sequence<MLActivation>

A list of three activation functions, the first one is used for the input (i), forget (f), and output (o) gate, the second one is used for the cell (g) gate, and the last used for filtering the output cell state before combining it with the result of the output gate to form the output hidden state. When not specified, implementations SHOULD use the sequence of the sigmoid function ("sigmoid") followed by two hyperbolic tangent functions ("tanh") respectively.

Arguments:

Returns: a sequence of MLOperand. The first element of the sequence is a 3-D tensor of shape [numDirections, batchSize, hiddenSize], the output hidden state from the last time step of the network. The second element is a 3-D tensor of shape [numDirections, batchSize, hiddenSize], the output cell state from the last time step of the network. Additionally, if options.returnSequence is set to true, the third element is the 4-D output tensor of shape [steps, numDirections, batchSize, hiddenSize] containing every output from each time step in the temporal sequence.

The lstm(input, weight, recurrentWeight, steps, hiddenSize, options) method steps are:
  1. If validating operand with this and any of input, weight, recurrentWeight, options.bias (if it exists), options.recurrentBias (if it exists), options.peepholeWeight (if it exists), options.initialHiddenState (if it exists), and options.initialCellState (if it exists) returns false, then throw a TypeError.

  2. If options.activations exists, and validating activation with this and any item in it returns false, then throw a TypeError.

  3. Let numDirections be 2 if options.direction is "both", or 1 otherwise.

  4. If the rank of any of input, weight or recurrentWeight is not 3, then throw a TypeError.

  5. If input’s shape[0] is not equal to steps, then throw a TypeError.

  6. Let batchSize be input’s shape[1].

  7. If options.bias exists:

    1. If its rank is not 2, then throw a TypeError.

    2. If its shape[0] is not numDirections, then throw a TypeError.

    3. If its shape[1] is not 4 * hiddenSize, then throw a TypeError.

  8. If options.recurrentBias exists:

    1. If its rank is not 2, then throw a TypeError.

    2. If its shape[0] is not numDirections, then throw a TypeError.

    3. If its shape[1] is not 4 * hiddenSize, then throw a TypeError.

  9. If options.peepholeWeight exists:

    1. If its rank is not 2, then throw a TypeError.

    2. If its shape[0] is not numDirections, then throw a TypeError.

    3. If its shape[1] is not 4 * hiddenSize, then throw a TypeError.

  10. If options.initialHiddenState exists:

    1. If its rank is not 3, then throw a TypeError.

    2. If its shape[0] is not numDirections, then throw a TypeError.

    3. If its shape[1] is not equal to batchSize, then throw a TypeError.

    4. If its shape[2] is not hiddenSize, then throw a TypeError.

  11. If options.initialCellState exists:

    1. If its rank is not 3, then throw a TypeError.

    2. If its shape[0] is not numDirections, then throw a TypeError.

    3. If its shape[1] is not equal to batchSize, then throw a TypeError.

    4. If its shape[2] is not hiddenSize, then throw a TypeError.

  12. If options.activations exists:

    1. If its size is not 3, then throw a TypeError.

  13. Calculate the output shape:

    1. Let desc be a new MLOperandDescriptor.

    2. Set desc.dimensions to the list « numDirections, batchSize, hiddenSize ».

    3. Set desc.dataType to input’s dataType.

    4. If options.returnSequence is true:

      1. Let desc2 be a new MLOperandDescriptor.

      2. Set desc2.dataType to input’s dataType.

      3. Set desc2.dimensions to the list « steps, numDirections, batchSize, hiddenSize ».

  14. Make graph connections:

    1. Let operator be an operator for the LSTM operation, given weight, recurrentWeight, steps, hiddenSize and options.

    2. Let output0 be the result of creating an MLOperand given this and desc.

    3. Let output1 be the result of creating an MLOperand given this and desc.

    4. If options.returnSequence is true:

      1. Let output2 be the result of creating an MLOperand given this and desc2.

      2. Let output be the list « output0, output1, output2 ».

      3. Set output0.[[operator]], output1.[[operator]] and output2.[[operator]] to operator.

    5. Otherwise:

      1. Let output be the list « output0, output1 ».

      2. Set output0.[[operator]] and output1.[[operator]] to operator.

    6. Set operator’s inputs to input, weight, and recurrentWeight.

    7. If options.bias exists, then add it to operator’s inputs.

    8. If options.recurrentBias exists, then add it to operator’s inputs.

    9. If options.peepholeWeight exists, then add it to operator’s inputs.

    10. If options.initialHiddenState exists, then add it to operator’s inputs.

    11. If options.initialCellState exists, then add it to operator’s inputs.

    12. If options.activations exists, then add its items to operator’s activation functions.

    13. Set operator’s output to output.

  15. Return output.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
function squeeze(builder, op) {
  return builder.reshape(op, op.shape().remove(0));
}

const numDirections = (options.direction == "both" ? 2 : 1);
let hiddenState = options.initialHiddenState;
let cellState = options.initialCellState;

if (!hiddenState) {
  const desc = { dataType: 'float32', dimensions: [numDirections, 1, hiddenSize] };
  const totalSize = numDirections * hiddenSize;
  hiddenState = builder.constant(desc, new Float32Array(totalSize).fill(0));
}

if (!cellState) {
  const desc = { dataType: 'float32', dimensions: [numDirections, 1, hiddenSize] };
  const totalSize = numDirections * hiddenSize;
  cellState = builder.constant(desc, new Float32Array(totalSize).fill(0));
}

let sequence = null;
let currentWeight = [];
let currentRecurrentWeight = [];
let currentBias = [];
let currentRecurrentBias = [];
let currentPeepholeWeight = [];

for (let dir = 0; dir < numDirections; ++dir) {
  currentWeight.push(squeeze(builder, builder.slice(weight, [dir, 0, 0], [1, 4 * hiddenSize, inputSize])));
  currentRecurrentWeight.push(squeeze(builder, builder.slice(recurrentWeight, [dir, 0, 0], [1, 4 * hiddenSize, hiddenSize])));
  currentBias.push(options.bias ? (squeeze(builder, builder.slice(options.bias, [dir, 0], [1, 4 * hiddenSize]))) : null);
  currentRecurrentBias.push(options.recurrentBias ?
    (squeeze(builder, builder.slice(options.recurrentBias, [dir, 0], [1, 4 * hiddenSize]))) : null);
  currentPeepholeWeight.push(options.peepholeWeight ?
    (squeeze(builder, builder.slice(options.peepholeWeight, [dir, 0], [1, 3 * hiddenSize]))) : null);
}

for (let step = 0; step < steps; ++step) {
  let currentHidden = [];
  let currentCell = [];
  let nextHidden = null;
  let nextCell = null;

  for (let dir = 0; dir < numDirections; ++dir) {
    currentHidden.push(squeeze(builder, builder.slice(hiddenState, [dir, 0, 0], [1, batchSize, hiddenSize])));
    currentCell.push(squeeze(builder, builder.slice(cellState, [dir, 0, 0], [1, batchSize, hiddenSize])));
  }

  for (let dir = 0; dir < numDirections; ++dir) {
    let slice = (dir == 1 || options.direction == "backward" ? steps - step - 1 : step);
    let currentInput = squeeze(builder, builder.slice(input, [slice, 0, 0], [1, batchSize, inputSize]));

    let results = builder.lstmCell(
      currentInput, currentWeight[dir], currentRecurrentWeight[dir],
      currentHidden[dir], currentCell[dir], hiddenSize, { bias: currentBias[dir],
      recurrentBias: currentRecurrentBias[dir], peepholeWeight: currentPeepholeWeight[dir],
      layout: options.layout, activations: options.activations });

    let output = builder.reshape(results[0], [1, null, hiddenSize]);
    let cell = builder.reshape(results[1], [1, null, hiddenSize]);

    nextHidden = (nextHidden ? builder.concat([nextHidden, output], 0) : output);
    nextCell = (nextCell ? builder.concat([nextCell, cell], 0) : cell);
  }

  hiddenState = nextHidden;
  cellState = nextCell;

  if (options.returnSequence) {
    nextHidden = builder.reshape(nextHidden, [1, numDirections, null, hiddenSize]);
    sequence = (sequence ? builder.concat([sequence, nextHidden], 0) : nextHidden);
  }
}

return (sequence ? [hiddenState, cellState, sequence] : [hiddenState, cellState]);

7.8.29. lstmCell

A single time step of the Long Short-Term Memory [LSTM] recurrent network using a cell state, an input, output, and forget gate to compute the cell state and the hidden state of the next time step that rolls into the output across the temporal sequence of the network.
dictionary MLLstmCellOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  MLOperand peepholeWeight;
  MLLstmWeightLayout layout = "iofg";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> lstmCell(MLOperand input,
                               MLOperand weight,
                               MLOperand recurrentWeight,
                               MLOperand hiddenState,
                               MLOperand cellState,
                               [EnforceRange] unsigned long hiddenSize,
                               optional MLLstmCellOptions options = {});
};

MLLstmCellOptions has the following members:

bias, of type MLOperand

The 1-D input bias tensor of shape [4 * hiddenSize]. The ordering of the bias vectors in the first dimension of the tensor shape is specified according to the layout argument.

recurrentBias, of type MLOperand

The 1-D recurrent bias tensor of shape [4 * hiddenSize]. The ordering of the bias vectors in the first dimension of the tensor shape is specified according to the layout argument.

peepholeWeight, of type MLOperand

The 1-D weight tensor for peepholes of shape [3 * hiddenSize]. The pack ordering of the weight vectors is for the input (i), output (o), and forget (f) gate, respectively.

layout, of type MLLstmWeightLayout, defaulting to "iofg"

The ordering of the weight and bias vectors for the internal gates of LSTM, specifically the input (i), output (o), forget (f), and cell (g) gate, as indicated in the first dimension of the weight and bias tensor shapes.

activations, of type sequence<MLActivation>

A list of three activation functions, the first one is used for the input (i), forget (f), and output (o) gate, the second one is used for the cell (g) gate, and the last used for filtering the output cell state before combining it with the result of the output gate to form the output hidden state. When not specified, they are assumed to be of the sigmoid function ("sigmoid") followed by two hyperbolic tangent functions ("tanh") respectively.

Arguments:

Returns: a sequence of MLOperand. The first element of the sequence is the output hidden state of the current time step of the recurrent network. The following element is the output cell state. Both elements are 2-D tensors of shape [batchSize, hiddenSize].

The lstmCell(input, weight, recurrentWeight, hiddenState, cellState, hiddenSize, options) method steps are:
  1. If validating operand with this and any of input, weight, recurrentWeight, hiddenState, cellState, options.bias (if it exists), options.recurrentBias (if it exists), and options.peepholeWeight (if it exists) returns false, then throw a TypeError.

  2. If options.activations exists, and validating activation with this and any item in it returns false, then throw a TypeError.

  3. If the rank of any of input, weight, recurrentWeight, hiddenState or cellState is not 2, then throw a TypeError.

  4. Let batchSize be input’s shape[0].

  5. If options.bias exists:

    1. If its rank is not 1, then throw a TypeError.

    2. If its shape[0] is not 4 * hiddenSize, then throw a TypeError.

  6. If options.recurrentBias exists:

    1. If its rank is not 1, then throw a TypeError.

    2. If its shape[0] is not 4 * hiddenSize, then throw a TypeError.

  7. If options.peepholeWeight exists:

    1. If its rank is not 1, then throw a TypeError.

    2. If its shape[0] is not 3 * hiddenSize, then throw a TypeError.

  8. If options.activations exists:

    1. If its size is not 3, then throw a TypeError.

  9. Let desc be a new MLOperandDescriptor.

  10. Set desc.dimensions to the list « batchSize, hiddenSize ».

  11. Set desc.dataType to input’s dataType.

  12. Make graph connections:

    1. Let output0 be the result of creating an MLOperand given this and desc.

    2. Let output1 be the result of creating an MLOperand given this and desc.

    3. Let output be the list « output0, output1 ».

    4. Let operator be an operator for the LSTM cell operation, given weight, recurrentWeight, hiddenState, cellState, hiddenSize and options.

    5. Set output0.[[operator]] and output1.[[operator]] to operator.

    6. Set operator’s inputs to input, weight, recurrentWeight, hiddenState, and cellState.

    7. If options.bias exists, then add it to operator’s inputs.

    8. If options.recurrentBias exists, then add it to operator’s inputs.

    9. If options.peepholeWeight exists, then add it to operator’s inputs.

    10. If options.activations exists, then add its items to operator’s activation functions.

    11. Set operator’s output to output.

  13. Return output.

The behavior of this operation when the weight layout is the default "iofg" layout, and the activation functions of the input/forget/output gate and the cell gate/the cell state’s filter for the output hidden state are sigmoid() and tanh() respectively can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
const zero = builder.constant(0);

// input gate (i)
let i = builder.sigmoid(
  builder.add(
    builder.mul(
      cellState,
      (options.peepholeWeight ? builder.slice(options.peepholeWeight, [0], [hiddenSize]) : zero)
    ),
    builder.add(
      builder.add(
        (options.bias ? builder.slice(options.bias, [0], [hiddenSize]) : zero),
        (options.recurrentBias ? builder.slice(options.recurrentBias, [0], [hiddenSize]) : zero)
      ),
      builder.add(
        builder.matmul(
          input,
          builder.transpose(builder.slice(weight, [0, 0], [hiddenSize, inputSize]))
        ),
        builder.matmul(
          hiddenState,
          builder.transpose(builder.slice(recurrentWeight, [0, 0], [hiddenSize, hiddenSize]))
        )
      )
    )
  )
);

// forget gate (f)
let f = builder.sigmoid(
  builder.add(
    builder.mul(
      cellState,
      (options.peepholeWeight ? builder.slice(options.peepholeWeight, [2 * hiddenSize], [hiddenSize]) : zero)
    ),
    builder.add(
      builder.add(
        (options.bias ? builder.slice(options.bias, [2 * hiddenSize], [hiddenSize]) : zero),
        (options.recurrentBias ? builder.slice(options.recurrentBias, [2 * hiddenSize], [hiddenSize]) : zero)
      ),
      builder.add(
        builder.matmul(
          input,
          builder.transpose(builder.slice(weight, [2 * hiddenSize, 0], [hiddenSize, inputSize]))
        ),
        builder.matmul(
          hiddenState,
          builder.transpose(builder.slice(recurrentWeight, [2 * hiddenSize, 0], [hiddenSize, hiddenSize]))
        )
      )
    )
  )
);

// cell gate (g)
let g = builder.tanh(
  builder.add(
    builder.add(
      (options.bias ? builder.slice(options.bias, [3 * hiddenSize], [hiddenSize]) : zero),
      (options.recurrentBias ? builder.slice(options.recurrentBias, [3 * hiddenSize], [hiddenSize]) : zero)
    ),
    builder.add(
      builder.matmul(
        input,
        builder.transpose(builder.slice(weight, [3 * hiddenSize, 0], [hiddenSize, inputSize]))
      ),
      builder.matmul(
        hiddenState,
        builder.transpose(builder.slice(recurrentWeight, [3 * hiddenSize, 0], [hiddenSize, hiddenSize]))
      )
    )
  )
);

// output gate (o)
let o = builder.sigmoid(
  builder.add(
    builder.mul(
      cellState,
      (options.peepholeWeight ? builder.slice(options.peepholeWeight, [hiddenSize], [hiddenSize]) : zero)
    ),
    builder.add(
      builder.add(
        (options.bias ? builder.slice(options.bias, [hiddenSize], [hiddenSize]) : zero),
        (options.recurrentBias ? builder.slice(options.recurrentBias, [hiddenSize], [hiddenSize]) : zero)
      ),
      builder.add(
        builder.matmul(
          input,
          builder.transpose(builder.slice(weight, [hiddenSize, 0], [hiddenSize, inputSize]))
        ),
        builder.matmul(
          hiddenState,
          builder.transpose(builder.slice(recurrentWeight, [hiddenSize, 0], [hiddenSize, hiddenSize]))
        )
      )
    )
  )
);

// output cell state (ct)
let ct = builder.add(builder.mul(f, cellState), builder.mul(i, g));

// output hidden state (ht)
let ht = builder.mul(o, builder.tanh(ct));

return [ht, ct];

7.8.30. matmul

Compute the matrix product of two input tensors.
partial interface MLGraphBuilder {
  MLOperand matmul(MLOperand a, MLOperand b);
};
Arguments:

Returns: an MLOperand. The output tensor that contains the matrix product of two input tensors.

Computes the matrix product of two input tensors as follows:
To calculate matmul output sizes, given MLOperand a and MLOperand b run the following steps:
  1. Let shapeA be a clone of a’s shape

  2. Let rankA be a’s rank.

  3. Let shapeB be a clone of b’s shape

  4. Let rankB be b’s rank.

  5. If either rankA or rankB is less than 2, then throw a TypeError.

  6. Let colsA be shapeA[rankA - 1].

  7. Let rowsA be shapeA[rankA - 2].

  8. Let colsB be shapeB[rankB - 1].

  9. Let rowsB be shapeB[rankB - 2].

  10. If colsA is not equal to rowsB, then throw a TypeError.

  11. Let batchShapeA be a clone of shapeA with the spatial dimensions (last 2 items) removed.

  12. Let batchShapeB be a clone of shapeB with the spatial dimensions (last 2 items) removed.

  13. Let outputShape be the result of bidirectionally broadcasting the shapes batchShapeA and batchShapeB. If that returns failure, then throw a TypeError.

  14. Append « rowsA, colsB » to outputShape.

  15. Return outputShape.

The matmul(a, b) method steps are:
  1. If validating operand with this and any of a and b returns false, then throw a TypeError.

  2. Let desc be a new MLOperandDescriptor.

  3. Set desc.dimensions to the result of calculating matmul output sizes given a and b.

  4. If that throws an error, re-throw the error.

  5. Set desc.dataType to a’s dataType.

  6. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the matrix multiplication operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to a and b.

    5. Set operator’s output to output.

  7. Return output.

7.8.31. pad

Inflate the tensor with constant or mirrored values on the edges.
enum MLPaddingMode {
  "constant",
  "edge",
  "reflection",
  "symmetric"
};

dictionary MLPadOptions {
  MLPaddingMode mode = "constant";
  float value = 0;
};

partial interface MLGraphBuilder {
  MLOperand pad(MLOperand input,
                sequence<[EnforceRange] unsigned long> beginningPadding,
                sequence<[EnforceRange] unsigned long> endingPadding,
                optional MLPadOptions options = {});
};

MLPadOptions has the following members:

mode, of type MLPaddingMode, defaulting to "constant"

The different ways to pad the tensor.

value, of type float, defaulting to 0

The padding value when mode is set to "constant".

Arguments:

Returns: an MLOperand. The padded output tensor. Each dimension of the output tensor can be calculated as follow:

output size = beginning padding + input size + ending padding

To calculate padding output sizes, given input, beginningPadding and endingPadding, run the following steps:
  1. Let shape be a copy of input’s shape.

  2. For index in the range 0 to shape’s rank, exclusive:

    1. Add to shape[index] the value of beginningPadding[index].

    2. Add to shape[index] the value of endingPadding[index].

  3. Return shape.

The pad(input, beginningPadding, endingPadding, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. If beginningPadding’s size and endingPadding’s size are not both equal to input’s rank, then throw a "TypeError".

  3. Let desc be a copy of input.[[descriptor]].

  4. Let outputShape be the result of calculating padding output sizes given input, beginningPadding and endingPadding.

  5. If any item in outputShape is not a valid dimension, then throw a TypeError.

  6. Set desc.dimensions to outputShape.

  7. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the padding operation, given beginningPadding, endingPadding and options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  8. Return output.

Examples for constant, edge, reflection and symmetric padding:
// input: [[1,2,3], [4,5,6]]
const input = builder.constant(
  { dataType: 'float32', dimensions: [2,3] }, new Float32Array([1,2,3,4,5,6]));

const beginningPadding = [1,2];
const endingPadding = [1,2];

// "constant" padded:
//    [[0,0,0,0,0,0,0],
//     [0,0,1,2,3,0,0],
//     [0,0,4,5,6,0,0],
//     [0,0,0,0,0,0,0]]
builder.pad(input, beginningPadding, endingPadding);

// "edge" padded:
//    [[1,1,1,2,3,3,3],
//     [1,1,1,2,3,3,3],
//     [4,4,4,5,6,6,6],
//     [4,4,4,5,6,6,6]]
builder.pad(input, beginningPadding, endingPadding, { mode: "edge" });

// "reflection" padded:
//    [[6,5,4,5,6,5,4],
//     [3,2,1,2,3,2,1],
//     [6,5,4,5,6,5,4],
//     [3,2,1,2,3,2,1]]
builder.pad(input, beginningPadding, endingPadding, { mode: "reflection" });

// "symmetric" padded:
//    [[2,1,1,2,3,3,2],
//     [2,1,1,2,3,3,2],
//     [5,4,4,5,6,6,5],
//     [5,4,4,5,6,6,5]]
builder.pad(input, beginningPadding, endingPadding, { mode: "symmetric" });

7.8.32. Pooling operations

Compute a pooling operation across all the elements within the moving window over the input tensor.
enum MLRoundingType {
  "floor",
  "ceil"
};

dictionary MLPool2dOptions {
  sequence<[EnforceRange] unsigned long> windowDimensions;
  sequence<[EnforceRange] unsigned long> padding;
  sequence<[EnforceRange] unsigned long> strides;
  sequence<[EnforceRange] unsigned long> dilations;
  MLInputOperandLayout layout = "nchw";
  MLRoundingType roundingType = "floor";
  sequence<[EnforceRange] unsigned long> outputSizes;
};

partial interface MLGraphBuilder {
  MLOperand averagePool2d(MLOperand input, optional MLPool2dOptions options = {});
  MLOperand l2Pool2d(MLOperand input, optional MLPool2dOptions options = {});
  MLOperand maxPool2d(MLOperand input, optional MLPool2dOptions options = {});
};

MLPool2dOptions has the following members:

windowDimensions, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [windowHeight, windowWidth]. Specifies the dimensions of the sliding window. The default value for the window dimensions are the height and width dimensions of the input shape.

padding, of type sequence<[EnforceRange] unsigned long>

A list of length 4: [beginningHeight, endingHeight, beginningWidth, endingWidth]. Specifies the additional rows and columns added to the beginning and ending of each spatial dimension of the convolution input. The default value is [0,0,0,0].

strides, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [strideHeight, strideWidth]. Specifies the stride of the sliding window for each spatial dimension of the convolution input. The default value is [1,1].

dilations, of type sequence<[EnforceRange] unsigned long>

A list of length 2: [dilationHeight, dilationWidth]. Specifies the dilation factor for each spatial dimension applied on the convolution filter (kernel). The default value is [1,1].

layout, of type MLInputOperandLayout, defaulting to "nchw"

Specifies the layout format of the input and output tensor as follows:

  • "nchw"

    • input tensor: [batches, inputChannels, height, width]

    • output tensor: [batches, outputChannels, height, width]

  • "nhwc":

    • input tensor: [batches, height, width, inputChannels]

    • output tensor: [batches, height, width, outputChannels]

roundingType, of type MLRoundingType, defaulting to "floor"

The rounding function used to compute the output shape.

outputSizes, of type sequence<[EnforceRange] unsigned long>

A list of length 2. Specifies the sizes of the two spacial dimensions of the output tensor. When the output sizes are explicitly specified, the roundingType is ignored.

If not specified, the output sizes are automatically computed.

Arguments:

Returns: an MLOperand. The output 4-D tensor that contains the result of the reduction. The logical shape is interpreted according to the value of layout. More specifically, if the options.roundingType is "floor", the spatial dimensions of the output tensor can be calculated as follow:

output size = floor(1 + (input size - filter size + beginning padding + ending padding) / stride)

or if options.roundingType is "ceil":

output size = ceil(1 + (input size - filter size + beginning padding + ending padding) / stride)

A global pooling operation such as one for the max pooling operation is a variant of pooling where the window dimensions is the spatial dimensions (last two dimensions) of the input shape, as follows.
// 'global' max pooling
builder.maxPool2d(input);
To calculate pool2d output sizes given MLInputOperandLayout layout, list of 4 unsigned integers inputShape, MLRoundingType roundingType, list of 2 unsigned integers windowDimensions, list of 4 unsigned integers padding, list of 2 unsigned integers strides, list of 2 unsigned integers dilations, and optional list of 2 unsigned integers outputSizes, perform these steps. They return a list of 4 unsigned integers.
  1. Switch on layout:

    "nchw"
    1. Let batches be inputShape[0].

    2. Let channels be inputShape[1].

    3. Let inputHeight be inputShape[2].

    4. Let inputWidth be inputShape[3].

    "nhwc"
    1. Let batches be inputShape[0].

    2. Let inputHeight be inputShape[1].

    3. Let inputWidth be inputShape[2].

    4. Let channels be inputShape[3].

  2. If outputSizes is not given, then:

    1. Let outputHeight be outputSizes[0].

    2. Let outputWidth be outputSizes[1].

  3. Otherwise:

    1. Let outputSizes be the result of calculating conv2d output sizes given inputHeight, inputWidth, windowDimensions[0], windowDimensions[1], padding, strides, and dilations.

    2. Let outputHeight be outputSizes[0].

    3. Let outputWidth be outputSizes[1].

    4. Switch on roundingType

      "floor"
      1. Set outputWidth to floor(outputWidth).

      2. Set outputHeight to floor(outputHeight).

      "ceil"
      1. Set outputWidth to ceiling(outputWidth).

      2. Set outputHeight to ceiling(outputHeight).

  4. Switch on layout:

    "nchw"

    Return « batches, channels, outputHeight, outputWidth ».

    "nhwc"

    Return « batches, outputHeight, outputWidth, channels ».

To create pooling operation given string op, MLOperand input and MLPool2dOptions options, run the following steps:
  1. Assert: op is one of "averagePool2d", "l2Pool2d", "maxPool2d".

  2. If validating operand with this and input returns false, then throw a TypeError.

  3. If input’s rank is not 4, then throw a TypeError.

  4. If options.windowDimensions exists and its size is not 2, then throw a TypeError.

  5. Otherwise, set options.windowDimensions to the height and width dimensions of the shape of input.

  6. If options.outputSizes exists, or if options.padding does not exist, set options.padding to the list « 0, 0, 0, 0 ».

  7. If options.padding's size is not 4, then throw a TypeError.

  8. If options.strides does not exist, set options.strides to the list « 1, 1 ».

  9. If options.strides's size is not 2, then throw a TypeError.

  10. If any value in options.strides is not greater than 0, then throw a TypeError.

  11. If options.outputSizes exists:

    1. If its size is not 2, then throw a TypeError.

    2. If its elements are not smaller than the elements at the same dimension (index) for options.strides, then throw a TypeError.

  12. If options.dilations does not exist, set options.dilations to the list « 1, 1 ».

  13. If options.dilations's size is not 2, then throw a TypeError.

  14. If any value in options.dilations is not greater than 0, then throw a TypeError.

  15. Let desc be a copy of input.[[descriptor]].

  16. Let outputShape be the result of calculating pool2d output sizes given options.layout, input’s shape, options.roundingType, options.windowDimensions, options.padding, options.strides, options.dilations, and options.outputSizes (if it exists).

  17. If any item in outputShape is not a valid dimension, then throw a TypeError.

  18. Set desc.dimensions to outputShape.

  19. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the op pooling operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  20. Return output.

The following pooling algorithms are supported.
The averagePool2d(input, options) method steps are:
  1. Let output be the result of running the create pooling operation given "averagePool2d", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The l2Pool2d(input, options) method steps are:
  1. Let output be the result of running the create pooling operation given "l2Pool2d", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The maxPool2d(input, options) method steps are:
  1. Let output be the result of running the create pooling operation given "maxPool2d", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

7.8.32.1. averagePool2d
Calculate the average value for patches of a feature map, and use it to create a pooled feature map. See § 7.8.32 Pooling operations for more detail.
7.8.32.2. l2Pool2d
Apply the L2 norm function to a region of the input feature map. The L2 norm is the square root of the sum of the squares of its elements. See § 7.8.32 Pooling operations for more detail.
7.8.32.3. maxPool2d
Calculate the maximum value for patches of a feature map, and use it to create a pooled feature map. See § 7.8.32 Pooling operations for more detail.

7.8.33. prelu

Calculate the parametric version of rectified linear function (Parametric ReLU) on the input tensor element-wise. Parametric ReLU is a type of leaky ReLU that, instead of having a scalar slope like 0.01, making the slope (coefficient of leakage) into a parameter that is learned during the model training phase of this operation. The calculation follows the expression max(0, x) + slope * min(0, x).
partial interface MLGraphBuilder {
  MLOperand prelu(MLOperand input, MLOperand slope);
};
Arguments:

Returns:

The prelu(input, slope) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Let descriptor be a new MLOperandDescriptor.

  3. Set descriptor.dataType to input’s dataType.

  4. Set descriptor.dimensions to the result of unidirectionally broadcasting the shapes slope’s shape and input’s shape.

    1. If that returns failure, then throw a TypeError.

  5. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and descriptor.

    2. Let operator be an operator for the PReLU operation, given slope.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to input and slope.

    5. Set operator’s output to output.

  6. Return output.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.add(builder.max(builder.constant(0), x),
                   builder.mul(slope, builder.min(builder.constant(0), x)));

7.8.34. Reduction operations

Reduce the input tensor along all dimensions, or along the axes specified in the axes array parameter. For each specified axis, the dimension with that index is reduced, i.e. the resulting tensor will not contain it, unless the keepDimensions option is specified. The values of the resulting tensor are calculated using the specified reduction function that takes as parameters all the values across the reduced dimension.
dictionary MLReduceOptions {
  sequence<[EnforceRange] unsigned long> axes;
  boolean keepDimensions = false;
};

partial interface MLGraphBuilder {
  MLOperand reduceL1(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceL2(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceLogSum(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceLogSumExp(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceMax(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceMean(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceMin(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceProduct(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceSum(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceSumSquare(MLOperand input, optional MLReduceOptions options = {});
};

MLReduceOptions has the following members:

axes, of type sequence<[EnforceRange] unsigned long>

The dimensions to reduce. The values in the list must be in the range [0, N-1] where N is the rank of the input tensor. If not present, all dimensions are reduced. If empty, no dimensions are reduced, and the shape of the output tensor is the same as the shape of the input tensor.

keepDimensions, of type boolean, defaulting to false

If true, the output has the same rank as the input, setting any reduced dimensions to size 1.

Arguments:

Returns: an MLOperand. The reduced output tensor.

Reduction types:
To calculate reduction output sizes, given a list of unsigned integers inputShape, a optional list of unsigned integers axes, and boolean keepDimensions, perform the following steps. They return a new list of unsigned integers.
  1. Let inputRank be inputShape’s size.

  2. If axes is not given, let axes be the range 0 to inputRank, exclusive.

  3. If keepDimensions is true, then:

    1. Let outputShape be a clone of inputShape.

    2. For each axis of axes:

      1. Set outputShape[axis] to 1.

  4. Otherwise:

    1. Let outputShape be an empty list.

    2. For each index in the range 0 to inputRank, exclusive:

      1. If axes does not contain index, then append inputShape[index].

  5. Return outputShape.

To create reduce operation given string op, MLOperand input and MLReduceOptions options, run the following steps:
  1. Assert: op is one of "reduceL1", "reduceL2", "reduceLogSum", "reduceLogSumExp", "reduceMax", "reduceMean", "reduceMin", "reduceProduct", "reduceSum", "reduceSumSquare".

  2. If validating operand with this and input returns false, then throw a TypeError.

  3. If options.axes exists, if any of its elements is not in the range 0 to input’s rank, exclusive, then throw a TypeError.

  4. Let outputShape be the result of calculating reduction output sizes given input’s shape, options.axes (if it exists), and options.keepDimensions.

  5. Let desc be a new MLOperandDescriptor.

  6. Set desc.dataType to input’s dataType.

  7. Set desc.dimensions to outputShape.

  8. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the op reduce operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  9. Return output.

The following reduce algorithms are supported.
The reduceL1(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceL1", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceL2(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceL2", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceLogSum(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceLogSum", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceLogSumExp(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceLogSumExp", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceMax(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceMax", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceMean(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceMean", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceMin(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceMin", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceProduct(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceProduct", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceSum(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceSum", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The reduceSumSquare(input, options) method steps are:
  1. Let output be the result of running the create reduce operation given "reduceSumSquare", input and options.

    1. If that throws an error, then re-throw the error.

  2. Return output.

The behavior of several reduction operations can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
// reduceLogSum(input, options)
return builder.log(builder.reduceSum(input, options));

// reduceLogSumExp(input, options)
return builder.log(builder.reduceSum(builder.exp(input), options));

// reduceSumSquare(input, options)
return builder.reduceSum(builder.pow(input, 2), options);
Some underlying platforms do not support an option like keepDimensions directly. This does not affect the underlying tensor data, only the shape. For example, if the input shape is [2, 3, 4], the axis is 1, and keepDimensions is true, the expected output shape is [2, 1 ,4]. If the underlying platform never keeps reduced dimensions it will produce an output shape of [2, 4]. The implementation can introduce a no-op reshape to [2, 1, 4]. A similar no-op reshape can be introduced if keepDimensions is false but the underlying platform always keeps reduced dimensions.

7.8.35. relu

Compute the rectified linear function of the input tensor.
partial interface MLGraphBuilder {
  MLOperand relu(MLOperand input);
  MLActivation relu();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.max(builder.constant(0), x);
7.8.35.1. relu(input)
Arguments:

Returns:

The relu(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the ReLU operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.35.2. relu()
Arguments:

Returns:

The relu() method steps are:
  1. Let op be the result of creating an MLActivation given this and "relu".

  2. Return op.

7.8.36. resample2d

Resample the tensor values from the source to the destination spatial dimensions according to the scaling factors.
enum MLInterpolationMode {
  "nearest-neighbor",
  "linear"
};

dictionary MLResample2dOptions {
  MLInterpolationMode mode = "nearest-neighbor";
  sequence<float> scales;
  sequence<[EnforceRange] unsigned long> sizes;
  sequence<[EnforceRange] unsigned long> axes;
};

partial interface MLGraphBuilder {
  MLOperand resample2d(MLOperand input, optional MLResample2dOptions options = {});
};
Arguments:

Returns: an MLOperand. The output 4-D tensor.

MLResample2dOptions has the following members:

mode, of type MLInterpolationMode, defaulting to "nearest-neighbor"

The interpolation algorithm used to fill the output tensor values.

scales, of type sequence<float>

A list of length 2. Specifies the scaling factor in each spatial dimensions of the input: [scaleHeight, scaleWidth]. The default value is [1.0, 1.0].

sizes, of type sequence<[EnforceRange] unsigned long>

A list of length 2. Specifies the target sizes for each spatial dimensions of the input: [sizeHeight, sizeWidth]. When the target sizes are specified, the scales argument is ignored, since the scaling factor values are derived from the target sizes of each spatial dimension of the input.

axes, of type sequence<[EnforceRange] unsigned long>

A list of length 2. Specifies the two consecutive dimensions of the input tensor to which the interpolation algorithm applies. The valid values in the sequence are [0, 1], [1, 2] or [2, 3]. The default value is [2, 3].

To check resample options given options, run the following steps:
  1. If options.scales does not exist, set it to the list « 1.0, 1.0 ».

  2. Otherwise, if any of its values is not greater than 0, or if its size is not 2, return false.

  3. If options.sizes exists, and if its size is not 2, or if any of its values is not greater than 0, return false.

  4. If options.axes does not exists, set it to the list « 2, 3 ».

  5. Otherwise, if its value is not one of « 0, 1», « 1, 2», « 2, 3 », return false.

  6. Return true.

To calculate resample output sizes given MLOperand input and MLResample2dOptions options, run the following steps:
  1. Let desc be a new MLOperandDescriptor initialized to input.[[descriptor]].

  2. For index in the range 0 to options.axes's size, exclusive:

    1. If options.sizes exists, then let size be options.sizes[index].

    2. Otherwise, let size be floor(input’s shape[options.axes[index]] * options.scales[index]).

    3. If size is not a valid dimension, then return failure.

    4. Set desc.dimensions[options.axes[index]] to size.

  3. Return desc.

The resample2d(input, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. If input’s rank is not 4, then throw a TypeError.

  3. If checking resample options given options returns false, then throw a TypeError.

  4. Let desc be the result of calculating resample output sizes given input and options. If that returns failure, then throw a TypeError.

  5. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the resample 2D operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  6. Return output.

7.8.37. reshape

Alter the shape of a tensor to a new shape. Reshape does not copy or change the content of the tensor. It just changes the tensor’s logical dimensions for the subsequent operations.
partial interface MLGraphBuilder {
  MLOperand reshape(MLOperand input, sequence<[EnforceRange] unsigned long> newShape);
};
Arguments:

Returns: an MLOperand. The output tensor. The values of the output tensor are the same as values of the input tensor. The shape of the output tensor is specified by the newShape argument.

The reshape(input, newShape) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Let outputShape be an empty array of unsigned long.

  3. If newShape’s size is 0, set outputShape to an empty list for a scalar.

  4. If any item in newShape is not a valid dimension, then throw a TypeError.

  5. Let inputElementCount be the product of all elements in input’s shape. Empty dimensions yield an inputElementCount of 1.

  6. If product of all values in newShape is not equal to inputElementCount, then throw a TypeError.

  7. Let desc be a copy of input.[[descriptor]].

  8. Set desc.dimensions to newShape.

  9. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and desc.

    2. Let operator be an operator for the reshape operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  10. Return output.

Many shape-related operations such as squeeze, unsqueeze, and flatten can be generically implemented using the reshape() operation as follows:
// Returns a tensor with all specified dimensions of input of size 1 removed.
function squeeze(input, axes) {
  if (!axes) axes = [];
  if (!axes.length)
    input.shape().forEach((item, i) => { axes.push(i); });
  shape = Array.from(input.shape());
  for (let axis in axes.sort().reverse())
    if (axis < shape.length && shape[axis] == 1)
      shape.splice(axis, 1);
  return builder.reshape(input, shape);
}

// Returns a new tensor with a dimension of size one inserted at the specified position.
function unsqueeze(input, axes) {
  shape = Array.from(input.shape());
  for(let axis in axes.sort())
    shape.splice(axis, 0, 1);
  return builder.reshape(input, shape);
}

// Flattens input by reshaping it into a one-dimensional tensor. 
function flatten(input, axis) {
  if (axis > input.shape().length) return input;
  let before = axis.slice(0, axis).reduce((a, b) => { a * b; });
  let after = axis.slice(axis, input.shape().length).reduce((a, b) => { a * b; });
  return builder.reshape(input, [before, after]);
}

7.8.38. sigmoid

Compute the sigmoid function of the input tensor. The calculation follows the expression 1 / (exp(-x) + 1).
partial interface MLGraphBuilder {
  MLOperand sigmoid(MLOperand input);
  MLActivation sigmoid();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.div(
          builder.constant(1),
          builder.add(
            builder.exp(builder.neg(x)),
            builder.constant(1)));
7.8.38.1. sigmoid(input)
Arguments:

Returns:

The sigmoid(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the sigmoid operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.38.2. sigmoid()
Arguments:

Returns:

The sigmoid() method steps are:
  1. Let op be the result of creating an MLActivation given this and "sigmoid".

  2. Return op.

7.8.39. slice

Produce a slice of the input tensor.
partial interface MLGraphBuilder {
  MLOperand slice(MLOperand input,
                  sequence<[EnforceRange] unsigned long> starts,
                  sequence<[EnforceRange] unsigned long> sizes);
};
Arguments:

Returns: an MLOperand. The output tensor of the same rank as the input tensor with tensor values stripped to the specified starting and ending indices in each dimension.

The slice(input, starts, sizes) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. If any of sizes’s items are 0, then throw a TypeError.

  3. If starts’s size and sizes’s size are not both equal to input’s rank, then throw a TypeError.

  4. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the slice operation, given starts and sizes.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  5. Return output.

7.8.40. softmax

Compute the softmax values of the 2-D input tensor along axis 1.
partial interface MLGraphBuilder {
  MLOperand softmax(MLOperand input);
  MLActivation softmax();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
// This sample deploys a well-known implementation trick [1] to compute the
// exponentials of the distances to the max value, instead of the exponentials
// of the input values itself, in order to increase the numerical stability of
// the result.
// [1]: https://cs231n.github.io/linear-classify/#softmax
const max_x = builder.reduceMax(x, { axes: [1], keepDimensions: true });
const exp_x = builder.exp(builder.sub(x, max_x));
return builder.div(exp_x, builder.reduceSum(exp_x, { axes: [1], keepDimensions: true }));
7.8.40.1. softmax(input)
Arguments:

Returns:

The softmax(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. If input’s rank is not 2, then throw a TypeError.

  3. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the softmax operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  4. Return output.

7.8.40.2. softmax()
Arguments:

Returns:

The softmax() method steps are:
  1. Let op be the result of creating an MLActivation given this and "softmax".

  2. Return op.

7.8.41. softplus

Compute the softplus function of the input tensor. The calculation follows the expression ln(1 + exp(x)).
partial interface MLGraphBuilder {
  MLOperand softplus(MLOperand input);
  MLActivation softplus();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.log(
         builder.add(
           builder.exp(x),
           builder.constant(1)));
7.8.41.1. softplus(input)
Arguments:

Returns:

The softplus(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the softplus operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.41.2. softplus()
Arguments:

Returns:

The softplus() method steps are:
  1. Let op be the result of creating an MLActivation given this and "softplus".

  2. Return op.

7.8.42. softsign

Compute the softsign function of the input tensor. The calculation follows the expression x / (1 + |x|).
partial interface MLGraphBuilder {
  MLOperand softsign(MLOperand input);
  MLActivation softsign();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.div(x, builder.add(builder.constant(1), builder.abs(x)));
7.8.42.1. softsign(input)
Arguments:

Returns:

The softsign(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the softsign operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.42.2. softsign()
Arguments:

Returns:

The softsign() method steps are:
  1. Let op be the result of creating an MLActivation given this and "softsign".

  2. Return op.

7.8.43. split

Split the input tensor into a number of sub tensors along the given axis.
dictionary MLSplitOptions {
  [EnforceRange] unsigned long axis = 0;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> split(
      MLOperand input,
      ([EnforceRange] unsigned long or sequence<[EnforceRange] unsigned long>) splits,
      optional MLSplitOptions options = {});
};
Arguments:

Returns: a sequence of MLOperand. The splitted output tensors. If splits is an unsigned long, the size of the output sequence equals to splits. The shape of each output tensor is the same as input except the dimension size of axis equals to the quotient of dividing the dimension size of input along axis by splits. If splits is a sequence of unsigned long, the size of the output sequence equals to the size of splits. The shape of the i-th output tensor is the same as input except along axis where the dimension size is splits[i].

MLSplitOptions has the following members:

axis, of type unsigned long, defaulting to 0

The dimension along which to split. Its value must be in the range [0, N-1] where N is the rank of the input tensor.

The split(input, splits, options) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Let axis be options.axis.

  3. If splits is an unsigned long:

    1. If input’s shape[axis] % splits is not 0, then throw a TypeError.

    2. Otherwise, let splitCount be splits.

  4. If splits is a sequence of unsigned long:

    1. If the sum of its elements is not equal to input’s shape[axis], then throw a TypeError.

    2. Otherwise, let splitCount be splits’s size.

  5. Make graph connections:

    1. Let operator be an operator for the split operation, given splits and options.

    2. Let outputs be a new list.

    3. For each index in the range 0 to splitCount, exclusive:

      1. Let operand be the result of copying an MLOperand given input.

      2. If splits is an unsigned long, then let newDimension be operand’s shape[axis] / splits.

      3. Otherwise, let newDimension be splits[index].

      4. Set operand’s shape[axis] to newDimension.

      5. Set operand.[[operator]] to operator.

      6. Append operand to outputs.

    4. Set operator’s input to input.

    5. Set operator’s outputs to outputs.

  6. Return outputs.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
// This sample shows the case that the splits parameter is an array.
const outputs = [];
let starts = Array(input_rank).fill(0);
let sizes = input_shape;
let start = 0;
for (const size of splits) {
  starts[options.axis] = start;
  sizes[options.axis] = size;
  outputs.push(builder.slice(input, starts, sizes));
  start += size;
}
return outputs;

7.8.44. tanh

Compute the hyperbolic tangent function of the input tensor. The calculation follows the expression (exp(2 * x) - 1) / (exp(2 * x) + 1).
partial interface MLGraphBuilder {
  MLOperand tanh(MLOperand input);
  MLActivation tanh();
};
The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
return builder.div(
          builder.sub(builder.exp(builder.mul(builder.constant(2), x)), builder.constant(1)),
          builder.add(builder.exp(builder.mul(builder.constant(2), x)), builder.constant(1)));
7.8.44.1. tanh(input)
Arguments:

Returns:

The tanh(input) method steps are:
  1. If validating operand with this and input returns false, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the hyperbolic tangent operation.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

7.8.44.2. tanh()
Arguments:

Returns:

The tanh() method steps are:
  1. Let op be the result of creating an MLActivation given this and "tanh".

  2. Return op.

7.8.45. transpose

Permute the dimensions of the input tensor according to the permutation argument.
dictionary MLTransposeOptions {
  sequence<[EnforceRange] unsigned long> permutation;
};

partial interface MLGraphBuilder {
  MLOperand transpose(MLOperand input, optional MLTransposeOptions options = {});
};

MLTransposeOptions has the following members:

permutation, of type sequence<[EnforceRange] unsigned long>

The values used to permute the output shape. The default value is [N-1, ..., 0], where N is the rank of the input tensor, e.g. [2,1,0] for a 3-D tensor. These default values cause the output to become a transposed tensor of the input. When specified, the number of values in the sequence must be the same as the rank of the input tensor, and the values in the sequence must be within the range from 0 to N-1 with no two or more same values found in the sequence.

Arguments:

Returns: an MLOperand. The permuted or transposed N-D tensor.

The transpose(input, options) method steps are:
  1. If options.permutation does not exist, let options.permutation be the reversed sequence of all indices for input’s shape.

  2. Otherwise if options.permutation exists:

    1. If its rank is not equal to input’s rank, then throw a TypeError.

    2. If its values are not in the range 0 to input’s rank exclusive, then throw a TypeError.

    3. If it contains duplicate values, then throw a TypeError.

  3. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the transpose operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  4. Return output.

7.8.46. triangular

Given a 2-D tensor (matrix), return a 2-D tensor containing either the upper or lower triangular part of the input tensor. If the input tensor has greater than 2 dimensions it is treated as a batch of matrices and the result has the same shape.
dictionary MLTriangularOptions {
  boolean upper = true;
  [EnforceRange] long diagonal = 0;
};

partial interface MLGraphBuilder {
  MLOperand triangular(MLOperand input, optional MLTriangularOptions options = {});
};

MLTriangularOptions has the following members:

upper, of type boolean, defaulting to true

Indicates whether the output the upper or the lower part of the input matrix is retained. True indicates that the upper part is retained.

diagonal, of type long, defaulting to 0

Specifies how many diagonals above or below the main diagonals of the input matrix are retained or excluded. A value of 0 means no diagonals other than the main diagonals are affected.

Arguments:

Returns: an MLOperand. The output tensor representing a triangular matrix, or batch of matrices which is the same shape as the input.

The triangular(input, options) method steps are:
  1. If input’s rank is less than 2, then throw a TypeError.

  2. Make graph connections:

    1. Let output be the result of copying an MLOperand given input.

    2. Let operator be an operator for the triangular operation, given options.

    3. Set output.[[operator]] to operator.

    4. Set operator’s input to input.

    5. Set operator’s output to output.

  3. Return output.

Examples of how triangular works in different diagonal settings.
// input:
//   [[7, 1, 2],
//    [9, 4, 8],
//    [2, 6, 3]]
const input = builder.constant(
{ dimensions: [3,3] }, new Float32Array([7,1,2,9,4,8,2,6,3]));

// upper triangular matrix:
//   [[7, 1, 2], 
//    [0, 4, 8],
//    [0, 0, 3]]
const upper = builder.triangular(input);

// upper triangular matrix with one additional set of diagonals excluded:
//   [[0, 1, 2], 
//    [0, 0, 8],
//    [0, 0, 0]]
const upperPositive = builder.triangular(input, { diagonal: 1 });

// upper triangular matrix with one additional set of diagonals retained:
//   [[7, 1, 2], 
//    [9, 4, 8],
//    [0, 6, 3]]
const upperNegative = builder.triangular(input, { diagonal: -1 });

// lower triangular matrix:
//   [[7, 0, 0],
//    [9, 4, 0],
//    [2, 6, 3]]
const lower = builder.triangular(input, { upper: false });

// lower triangular matrix with one additional set of diagonals retained:
//   [[7, 1, 0],
//    [9, 4, 8],
//    [2, 6, 3]]
const lowerPositive = builder.triangular(input, { upper: false, diagonal: 1 });

// lower triangular matrix with one additional set of diagonals excluded:
//   [[0, 0, 0],
//    [9, 0, 0],
//    [2, 6, 0]]
const lowerNegative = builder.triangular(input, { upper: false, diagonal: -1 })

// lower triangular matrix with two batches:
//   [[[7, 0, 0],
//     [9, 4, 0],
//     [2, 6, 3]],
//    [[1, 0, 0],
//     [4, 5, 0],
//     [7, 8, 9]]]
const lowerWithBatches = builder.triangular(input, { upper: false });

7.8.47. where

Select the values from the input or the other tensor depending on the corresponding boolean values of the condition tensor. The condition tensor is often the output of one of the element-wise logical operations.

The input tensors will be broadcasted according to [numpy-broadcasting-rule] to the final output shape. The rank of the output tensor is the maximum rank of the input tensors. For each dimension of the output tensor, its size is the maximum size along that dimension of the input tensors.

partial interface MLGraphBuilder {
  MLOperand where(MLOperand condition, MLOperand input, MLOperand other);
};
Arguments:

Returns: an MLOperand. The output tensor that contains the values selected element-wise from either the input or the other tensor.

The where(condition, input, other) method steps are:
  1. If condition’s dataType is not equal to "uint8", then throw a TypeError.

  2. If input’s dataType is not equal to other’s dataType, then throw a TypeError.

  3. Let descriptor be a new MLOperandDescriptor.

  4. Set descriptor.dataType to input’s dataType.

  5. Set descriptor.dimensions to the result of bidirectionally broadcasting the shapes input’s shape and other’s shape.

    1. If that returns failure, then throw a TypeError.

  6. If condition is not bidirectionally broadcastable to descriptor.dimensions, then throw a TypeError.

  7. Make graph connections:

    1. Let output be the result of creating an MLOperand given this and descriptor.

    2. Let operator be an operator for the where operation, given condition, input and other.

    3. Set output.[[operator]] to operator.

    4. Set operator’s inputs to condition, input and other.

    5. Set operator’s output to output.

  8. Return output.

The behavior of this operation can be generically emulated from the usage of other operations as follows, although user agents typically have a more efficient implementation. In cases where the underlying platform does not directly support an operation, this decomposition can be used as a template to guide the implementation.
const c = builder.clamp(condition, {'minValue': 0, 'maxValue': 1});
builder.add(
  builder.mul(
    input,
    builder.cast(c, input.dataType())),
  builder.mul(
    other,
    builder.cast(builder.not(c), other.dataType())));

8. Algorithms

8.1. Broadcasting

Broadcasting refers to how operations treat tensors with different shapes, and follow the precedent set by [numpy-broadcasting-rule].

To unidirectionally broadcast the shapes shapeA and shapeB, perform the following steps. shapeA and shapeB are lists of positive integers, representing the dimensions of tensors, and the steps return a new list of positive integers, or failure.
  1. Let sizeA be shapeA’s size.

  2. Let sizeB be shapeB’s size.

  3. If sizeB > sizeA, then return failure.

  4. Let paddedB be a clone of shapeB.

  5. While paddedB’s size is less than sizeA, prepend 1 to paddedB.

  6. Let outputShape be a new list.

  7. For each index in the range 0 to sizeA, exclusive:

    1. Let dimA be shapeA[index].

    2. Let dimB be paddedB[index].

    3. If dimA is not equal to dimB and dimA is not equal to 1, then return failure.

    4. Append dimA to outputShape.

  8. Return outputShape.

shapeA is unidirectionally broadcastable to shapeB if unidirectionally broadcasting the shapes shapeA and shapeB does not result in failure.

To bidirectionally broadcast the shapes shapeA and shapeB, perform the following steps. shapeA and shapeB are lists of positive integers, representing the dimensions of tensors, and the steps return a new list of positive integers, or failure.
  1. Let sizeA be shapeA’s size.

  2. Let sizeB be shapeB’s size.

  3. Let outputSize be the maximum of sizeA and sizeB.

  4. Let paddedA be a clone of shapeA.

  5. While paddedA’s size is less than outputSize, prepend 1 to paddedA.

  6. Let paddedB be a clone of shapeB.

  7. While paddedB’s size is less than outputSize, prepend 1 to paddedB.

  8. Let outputShape be a new list.

  9. For each index in the range 0 to outputSize, exclusive:

    1. Let dimA be paddedA[index].

    2. Let dimB be paddedB[index].

    3. If dimA is not equal to dimB, and dimA is not equal to 1, and dimB is not equal to 1, then return failure.

    4. Append the maximum of dimA and dimB to outputShape.

  10. Return outputShape.

shapeA is bidirectionally broadcastable to shapeB if bidirectionally broadcasting the shapes shapeA and shapeB does not result in failure.

9. Examples

The following code gets the MLContext object.
const context = await navigator.ml.createContext({powerPreference: 'low-power'});
Given the following build graph:
constant1 ---+
             +--- Add ---> intermediateOutput1 ---+
input1    ---+                                    |
                                                  +--- Mul---> output
constant2 ---+                                    |
             +--- Add ---> intermediateOutput2 ---+
input2    ---+
The following code implements the graph:
// Use tensors in 4 dimensions.
const TENSOR_DIMS = [1, 2, 2, 2];
const TENSOR_SIZE = 8;

const builder = new MLGraphBuilder(context);

// Create MLOperandDescriptor object.
const desc = {dataType: 'float32', dimensions: TENSOR_DIMS};

// constant1 is a constant MLOperand with the value 0.5.
const constantBuffer1 = new Float32Array(TENSOR_SIZE).fill(0.5);
const constant1 = builder.constant(desc, constantBuffer1);

// input1 is one of the input MLOperands. Its value will be set before execution.
const input1 = builder.input('input1', desc);

// constant2 is another constant MLOperand with the value 0.5.
const constantBuffer2 = new Float32Array(TENSOR_SIZE).fill(0.5);
const constant2 = builder.constant(desc, constantBuffer2);

// input2 is another input MLOperand. Its value will be set before execution.
const input2 = builder.input('input2', desc);

// intermediateOutput1 is the output of the first Add operation.
const intermediateOutput1 = builder.add(constant1, input1);

// intermediateOutput2 is the output of the second Add operation.
const intermediateOutput2 = builder.add(constant2, input2);

// output is the output MLOperand of the Mul operation.
const output = builder.mul(intermediateOutput1, intermediateOutput2);
Compile the graph up to the output operand.
// Compile the constructed graph.
const graph = await builder.build({'output': output});
The following code executes the compiled graph.
// Setup the input buffers with value 1.
const inputBuffer1 = new Float32Array(TENSOR_SIZE).fill(1);
const inputBuffer2 = new Float32Array(TENSOR_SIZE).fill(1);
const outputBuffer = new Float32Array(TENSOR_SIZE);

// Execute the compiled graph with the specified inputs.
const inputs = {
'input1': inputBuffer1,
'input2': inputBuffer2,
};
const outputs = {'output': outputBuffer};
const result = await context.compute(graph, inputs, outputs);

console.log('Output value: ' + result.outputs.output);
// Output value: 2.25,2.25,2.25,2.25,2.25,2.25,2.25,2.25

10. Appendices

10.1. MLOperandDataType and ArrayBufferView compatibility

MLOperandDataType ArrayBufferView
float32 Float32Array
float16 Float16Array
int32 Int32Array
uint32 Uint32Array
int8 Int8Array
uint8 Uint8Array

Float16Array is at ECMA Stage 3 signaling its design is finished. Implementers wanting to enable this type ahead native implementations can emulate the type by passing raw bits via Uint16Array. [Issue webnn#373]

11. Acknowledgements

This specification follows the concepts of the Android Neural Networks API C API.

Thanks to Tomoyuki Shimizu, Ningxin Hu, Zhiqiang Yu and Belem Zhang for the use cases.

Thanks to Nikhil Thorat, Daniel Smilkov, Ganesan Ramalingam, Rafael Cintron and Benjamin Poulain for their contributions to the API specification.

Thanks to Sangwhan Moon and the W3C Technical Architecture Group for review of this specification for web architecture fit, design consistency and developer ergonomics.

Thanks to Zoltan Kis for adding algorithms and making navigating this specification a delightful experience. Thanks to Joshua Bell for aligning the specification with modern editorial conventions. Thanks to Ningxin Hu, Lisha Guo, Shiyi Zou, Mingming Xu, Junwei Fu, Bruce Dai and Bin Miao for careful review and comments.

Thanks to W3C Privacy Interest Group for privacy and security review and feedback.

Thanks to Alex Gough and the Chrome Security team for security review and questions.

Thanks to Michal Karzynski for sharing practical guidelines and learnings from ONNX.

Thanks to Kaustubha Govind and Chrome privacy reviewers for feedback and privacy considerations.

Thanks to Jiewei Qian for Chromium implementation review and feedback.

Thanks to Dwayne Robinson, Joshua Lochner and Wanming Lin for their work investigating and providing recommendation for transformer support. Additional thanks to Dwayne and Wanming for providing reviews of operator conformance and web-platform-tests implementation.

Thanks to Feng Dai for his continuous contributions that keep web-platform-tests evolving alongside the specification.

Conformance

Document conventions

Conformance requirements are expressed with a combination of descriptive assertions and RFC 2119 terminology. The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in the normative parts of this document are to be interpreted as described in RFC 2119. However, for readability, these words do not appear in all uppercase letters in this specification.

All of the text of this specification is normative except sections explicitly marked as non-normative, examples, and notes. [RFC2119]

Examples in this specification are introduced with the words “for example” or are set apart from the normative text with class="example", like this:

This is an example of an informative example.

Informative notes begin with the word “Note” and are set apart from the normative text with class="note", like this:

Note, this is an informative note.

Conformant Algorithms

Requirements phrased in the imperative as part of algorithms (such as "strip any leading space characters" or "return false and abort these steps") are to be interpreted with the meaning of the key word ("must", "should", "may", etc) used in introducing the algorithm.

Conformance requirements phrased as algorithms or specific steps can be implemented in any manner, so long as the end result is equivalent. In particular, the algorithms defined in this specification are intended to be easy to understand and are not intended to be performant. Implementers are encouraged to optimize.

Index

Terms defined by this specification

Terms defined by reference

References

Normative References

[ECMASCRIPT]
ECMAScript Language Specification. URL: https://tc39.es/ecma262/multipage/
[HTML]
Anne van Kesteren; et al. HTML Standard. Living Standard. URL: https://html.spec.whatwg.org/multipage/
[INFRA]
Anne van Kesteren; Domenic Denicola. Infra Standard. Living Standard. URL: https://infra.spec.whatwg.org/
[NUMPY-BROADCASTING-RULE]
The SciPy community. General Broadcasting Rules of NumPy. July 2019. URL: https://docs.scipy.org/doc/numpy/user/basics.broadcasting.html#general-broadcasting-rules
[PERMISSIONS-POLICY-1]
Ian Clelland. Permissions Policy. 18 December 2023. WD. URL: https://www.w3.org/TR/permissions-policy-1/
[RFC2119]
S. Bradner. Key words for use in RFCs to Indicate Requirement Levels. March 1997. Best Current Practice. URL: https://datatracker.ietf.org/doc/html/rfc2119
[VC-DATA-MODEL-2.0]
Manu Sporny; et al. Verifiable Credentials Data Model v2.0. 16 April 2024. CR. URL: https://www.w3.org/TR/vc-data-model-2.0/
[WEBGPU]
Kai Ninomiya; Brandon Jones; Jim Blandy. WebGPU. 24 April 2024. WD. URL: https://www.w3.org/TR/webgpu/
[WEBIDL]
Edgar Chen; Timothy Gu. Web IDL Standard. Living Standard. URL: https://webidl.spec.whatwg.org/

Informative References

[Batch-Normalization]
Sergey Ioffe; Christian Szegedy. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. March 2015. URL: https://arxiv.org/abs/1502.03167
[ContextualLoss]
Roey Mechrez; Itamar Talmi; Lihi Zelnik-Manor. The Contextual Loss for Image Transformation with Non-Aligned Data. July 2018. URL: https://arxiv.org/abs/1803.02077
[DeepLabv3+]
Liang-Chieh Chen; et al. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. August 2018. URL: https://arxiv.org/abs/1802.02611
[DeepMoji]
Bjarke Felbo; et al. Using millions of emoji occurrences to learn any-domain representations for detecting sentiment, emotion and sarcasm. October 2017. URL: https://arxiv.org/abs/1708.00524
[ELU]
Djork-Arné Clevert; Thomas Unterthiner; Sepp Hochreiter. Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs). February 2016. URL: https://arxiv.org/abs/1511.07289
[Error-Function]
Larry C. Andrews. Special functions of mathematics for engineers. 1998. URL: https://books.google.com/books?id=2CAqsF-RebgC&pg=PA110
[FaceForensics++]
Andreas Rössler; et al. FaceForensics++. January 2019. URL: https://github.com/ondyari/FaceForensics
[FaceNet]
Florian Schroff; Dmitry Kalenichenko; James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. June 2015. URL: https://arxiv.org/abs/1503.03832
[FAN]
Adrian Bulat; Georgios Tzimiropoulos. How far are we from solving the 2D & 3D Face Alignment problem? (and a dataset of 230,000 3D facial landmarks). September 2017. URL: https://arxiv.org/abs/1703.07332
[GNMT]
Minh-Thang Luong; Eugene Brevdo; Rui Zhao. Neural Machine Translation (seq2seq) Tutorial. May 2017. URL: https://github.com/tensorflow/nmt
[GPT2]
Alec Radford; et al. Language Models are Unsupervised Multitask Learners. February 2019. URL: https://d4mucfpksywv.cloudfront.net/better-language-models/language-models.pdf
[GRU]
Kyunghyun Cho; et al. Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation. September 2014. URL: https://arxiv.org/pdf/1406.1078.pdf
[HR-TIME-3]
Yoav Weiss. High Resolution Time. 19 July 2023. WD. URL: https://www.w3.org/TR/hr-time-3/
[IM2TXT]
Oriol Vinyals; et al. Show and Tell: Lessons learned from the 2015 MSCOCO Image Captioning Challenge. September 2016. URL: https://arxiv.org/abs/1609.06647
[Instance-Normalization]
Dmitry Ulyanov; Andrea Vedaldi; Victor Lempitsky. Instance Normalization: The Missing Ingredient for Fast Stylization. July 2016. URL: https://arxiv.org/abs/1607.08022
[Layer-Normalization]
Jimmy Lei Ba; Jamie Ryan Kiros; Geoffrey E. Hinton. Layer Normalization. July 2016. URL: https://arxiv.org/abs/1607.06450
[LDM]
Robin Rombach; et al. High-Resolution Image Synthesis with Latent Diffusion Models. April 2022. URL: https://arxiv.org/abs/2112.10752
[LeakyReLU]
Andrew L. Maas; Awni Y. Hannun; Andrew Y. Ng. Rectifier Nonlinearities Improve Neural Network Acoustic Models. June 2013. URL: https://pdfs.semanticscholar.org/367f/2c63a6f6a10b3b64b8729d601e69337ee3cc.pdf
[LLAMA-2-7B]
Hugo Touvron; et al. Llama 2: Open Foundation and Fine-Tuned Chat Models. July 2023. URL: https://arxiv.org/abs/2307.09288
[LSTM]
Sepp Hochreiter; Jürgen Schmidhuber. Long Short-Term Memory. November 1997. URL: https://doi.org/10.1162/neco.1997.9.8.1735
[m2m100_418M]
Angela Fan; et al. Beyond English-Centric Multilingual Machine Translation. October 2020. URL: https://arxiv.org/abs/2010.11125
[MaskR-CNN]
Kaiming He; et al. Mask R-CNN. January 2018. URL: https://arxiv.org/abs/1703.06870
[MobileNetV3]
Andrew Howard; et al. Searching for MobileNetV3. November 2019. URL: https://arxiv.org/pdf/1905.02244
[MODELS]
Machine Learning for the Web Community Group. The first-wave models. 2020. URL: https://github.com/webmachinelearning/webnn/blob/master/op_compatibility/first_wave_models.md
[OpenNMT]
Guillaume Klein; et al. OpenNMT: Open-Source Toolkit for Neural Machine Translation. March 2017. URL: https://arxiv.org/abs/1701.02810
[PairedCycleGAN]
Huiwen Chang; et al. PairedCycleGAN: Asymmetric Style Transfer for Applying and Removing Makeup. June 2018. URL: http://openaccess.thecvf.com/content_cvpr_2018/html/Chang_PairedCycleGAN_Asymmetric_Style_CVPR_2018_paper.html
[PoseNet]
Dan Oved. Real-time Human Pose Estimation in the Browser with TensorFlow.js. May 2018. URL: https://medium.com/tensorflow/real-time-human-pose-estimation-in-the-browser-with-tensorflow-js-7dd0bc881cd5
[POWERFUL-FEATURES]
Mike West. Secure Contexts. 10 November 2023. CR. URL: https://www.w3.org/TR/secure-contexts/
[RNNoise]
Jean-Marc Valin. Recurrent neural network for audio noise reduction. September 2017. URL: https://github.com/xiph/rnnoise
[SECURITY-PRIVACY-QUESTIONNAIRE]
Theresa O'Connor; Peter Snyder. Self-Review Questionnaire: Security and Privacy. 16 December 2021. NOTE. URL: https://www.w3.org/TR/security-privacy-questionnaire/
[SegAny]
Alexander Kirillov; et al. Segment Anything. April 2023. URL: https://arxiv.org/abs/2304.02643
[SRGAN]
Christian Ledig; et al. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. May 2017. URL: https://arxiv.org/abs/1609.04802
[SSD]
Wei Liu; et al. SSD: Single Shot MultiBox Detector. December 2016. URL: https://arxiv.org/abs/1512.02325
[T5-SMALL]
Colin Raffel; et al. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. June 2020. URL: https://jmlr.org/papers/volume21/20-074/20-074.pdf
[Video-Summarization-with-LSTM]
Ke Zhang; et al. Video summarization with long short-term memory. October 2016. URL: http://www-scf.usc.edu/~zhan355/ke_eccv2016.pdf
[WEBMACHINELEARNING-ETHICS]
Anssi Kostiainen. Ethical Principles for Web Machine Learning. 8 January 2024. NOTE. URL: https://www.w3.org/TR/webmachinelearning-ethics/
[Whisper]
Alec Radford; et al. Robust Speech Recognition via Large-Scale Weak Supervision. December 2022. URL: https://arxiv.org/abs/2212.04356
[YOLO]
Joseph Redmon; et al. You Only Look Once: Unified, Real-Time Object Detection. May 2016. URL: https://arxiv.org/abs/1506.02640

IDL Index

interface mixin NavigatorML {
  [SecureContext, SameObject] readonly attribute ML ml;
};
Navigator includes NavigatorML;
WorkerNavigator includes NavigatorML;

enum MLDeviceType {
  "cpu",
  "gpu"
};

enum MLPowerPreference {
  "default",
  "high-performance",
  "low-power"
};

dictionary MLContextOptions {
  MLDeviceType deviceType = "cpu";
  MLPowerPreference powerPreference = "default";
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface ML {
  Promise<MLContext> createContext(optional MLContextOptions options = {});
  Promise<MLContext> createContext(GPUDevice gpuDevice);
};

typedef record<DOMString, ArrayBufferView> MLNamedArrayBufferViews;

dictionary MLComputeResult {
  MLNamedArrayBufferViews inputs;
  MLNamedArrayBufferViews outputs;
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLContext {
  Promise<MLComputeResult> compute(
      MLGraph graph, MLNamedArrayBufferViews inputs, MLNamedArrayBufferViews outputs);
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLGraph {};

enum MLInputOperandLayout {
  "nchw",
  "nhwc"
};

enum MLOperandDataType {
  "float32",
  "float16",
  "int32",
  "uint32",
  "int64",
  "uint64",
  "int8",
  "uint8"
};

dictionary MLOperandDescriptor {
  required MLOperandDataType dataType;
  sequence<[EnforceRange] unsigned long> dimensions = [];
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLOperand {
  MLOperandDataType dataType();
  sequence<unsigned long> shape();
};

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLActivation {};

typedef record<DOMString, MLOperand> MLNamedOperands;

[SecureContext, Exposed=(Window, DedicatedWorker)]
interface MLGraphBuilder {
  // Construct the graph builder from the context.
  constructor(MLContext context);

  // Create an operand for a graph input.
  MLOperand input(DOMString name, MLOperandDescriptor descriptor);

  // Create an operand for a graph constant.
  MLOperand constant(MLOperandDescriptor descriptor, ArrayBufferView bufferView);

  // Create a single-value operand from the specified number of the specified type.
  MLOperand constant(double value, optional MLOperandDataType type = "float32");

  // Compile the graph up to the specified output operands asynchronously.
  Promise<MLGraph> build(MLNamedOperands outputs);
};

dictionary MLArgMinMaxOptions {
  sequence<[EnforceRange] unsigned long> axes;
  boolean keepDimensions = false;
  boolean selectLastIndex = false;
};

partial interface MLGraphBuilder {
  MLOperand argMin(MLOperand input, optional MLArgMinMaxOptions options = {});
  MLOperand argMax(MLOperand input, optional MLArgMinMaxOptions options = {});
};

dictionary MLBatchNormalizationOptions {
  MLOperand scale;
  MLOperand bias;
  [EnforceRange] unsigned long axis = 1;
  float epsilon = 1e-5;
  MLActivation activation;
};

partial interface MLGraphBuilder {
  MLOperand batchNormalization(MLOperand input, MLOperand mean, MLOperand variance,
                               optional MLBatchNormalizationOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand cast(MLOperand input, MLOperandDataType type);
};

dictionary MLClampOptions {
  float minValue;
  float maxValue;
};

partial interface MLGraphBuilder {
  MLOperand clamp(MLOperand input, optional MLClampOptions options = {});
  MLActivation clamp(optional MLClampOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand concat(sequence<MLOperand> inputs, [EnforceRange] unsigned long axis);
};

enum MLConv2dFilterOperandLayout {
  "oihw",
  "hwio",
  "ohwi",
  "ihwo"
};

dictionary MLConv2dOptions {
  sequence<[EnforceRange] unsigned long> padding;
  sequence<[EnforceRange] unsigned long> strides;
  sequence<[EnforceRange] unsigned long> dilations;
  [EnforceRange] unsigned long groups = 1;
  MLInputOperandLayout inputLayout = "nchw";
  MLConv2dFilterOperandLayout filterLayout = "oihw";
  MLOperand bias;
  MLActivation activation;
};

partial interface MLGraphBuilder {
  MLOperand conv2d(MLOperand input,
                   MLOperand filter,
                   optional MLConv2dOptions options = {});
};

enum MLConvTranspose2dFilterOperandLayout {
  "iohw",
  "hwoi",
  "ohwi"
};

dictionary MLConvTranspose2dOptions {
  sequence<[EnforceRange] unsigned long> padding;
  sequence<[EnforceRange] unsigned long> strides;
  sequence<[EnforceRange] unsigned long> dilations;
  sequence<[EnforceRange] unsigned long> outputPadding;
  sequence<[EnforceRange] unsigned long> outputSizes;
  [EnforceRange] unsigned long groups = 1;
  MLInputOperandLayout inputLayout = "nchw";
  MLConvTranspose2dFilterOperandLayout filterLayout = "iohw";
  MLOperand bias;
  MLActivation activation;
};

partial interface MLGraphBuilder {
  MLOperand convTranspose2d(MLOperand input, MLOperand filter,
                            optional MLConvTranspose2dOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand add(MLOperand a, MLOperand b);
  MLOperand sub(MLOperand a, MLOperand b);
  MLOperand mul(MLOperand a, MLOperand b);
  MLOperand div(MLOperand a, MLOperand b);
  MLOperand max(MLOperand a, MLOperand b);
  MLOperand min(MLOperand a, MLOperand b);
  MLOperand pow(MLOperand a, MLOperand b);
};

partial interface MLGraphBuilder {
  MLOperand equal(MLOperand a, MLOperand b);
  MLOperand greater(MLOperand a, MLOperand b);
  MLOperand greaterOrEqual(MLOperand a, MLOperand b);
  MLOperand lesser(MLOperand a, MLOperand b);
  MLOperand lesserOrEqual(MLOperand a, MLOperand b);
  MLOperand not(MLOperand a);
};

partial interface MLGraphBuilder {
  MLOperand abs(MLOperand input);
  MLOperand ceil(MLOperand input);
  MLOperand cos(MLOperand input);
  MLOperand erf(MLOperand input);
  MLOperand exp(MLOperand input);
  MLOperand floor(MLOperand input);
  MLOperand identity(MLOperand input);
  MLOperand log(MLOperand input);
  MLOperand neg(MLOperand input);
  MLOperand reciprocal(MLOperand input);
  MLOperand sin(MLOperand input);
  MLOperand sqrt(MLOperand input);
  MLOperand tan(MLOperand input);
};

dictionary MLEluOptions {
  float alpha = 1;
};

partial interface MLGraphBuilder {
  MLOperand elu(MLOperand input, optional MLEluOptions options = {});
  MLActivation elu(optional MLEluOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand expand(MLOperand input, sequence<[EnforceRange] unsigned long> newShape);
};

dictionary MLGatherOptions {
  [EnforceRange] unsigned long axis = 0;
};

partial interface MLGraphBuilder {
  MLOperand gather(MLOperand input,
                   MLOperand indices,
                   optional MLGatherOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand gelu(MLOperand input);
  MLActivation gelu();
};

dictionary MLGemmOptions {
  MLOperand c;
  float alpha = 1.0;
  float beta = 1.0;
  boolean aTranspose = false;
  boolean bTranspose = false;
};

partial interface MLGraphBuilder {
  MLOperand gemm(MLOperand a, MLOperand b, optional MLGemmOptions options = {});
};

enum MLGruWeightLayout {
  "zrn",  // update-reset-new gate ordering
  "rzn"   // reset-update-new gate ordering
};

enum MLRecurrentNetworkDirection {
  "forward",
  "backward",
  "both"
};

dictionary MLGruOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  MLOperand initialHiddenState;
  boolean resetAfter = true;
  boolean returnSequence = false;
  MLRecurrentNetworkDirection direction = "forward";
  MLGruWeightLayout layout = "zrn";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> gru(MLOperand input,
                          MLOperand weight,
                          MLOperand recurrentWeight,
                          [EnforceRange] unsigned long steps,
                          [EnforceRange] unsigned long hiddenSize,
                          optional MLGruOptions options = {});
};

dictionary MLGruCellOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  boolean resetAfter = true;
  MLGruWeightLayout layout = "zrn";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  MLOperand gruCell(MLOperand input,
                    MLOperand weight,
                    MLOperand recurrentWeight,
                    MLOperand hiddenState,
                    [EnforceRange] unsigned long hiddenSize,
                    optional MLGruCellOptions options = {});
};

dictionary MLHardSigmoidOptions {
  float alpha = 0.2;
  float beta = 0.5;
};

partial interface MLGraphBuilder {
  MLOperand hardSigmoid(MLOperand input, optional MLHardSigmoidOptions options = {});
  MLActivation hardSigmoid(optional MLHardSigmoidOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand hardSwish(MLOperand input);
  MLActivation hardSwish();
};

dictionary MLInstanceNormalizationOptions {
  MLOperand scale;
  MLOperand bias;
  float epsilon = 1e-5;
  MLInputOperandLayout layout = "nchw";
};

partial interface MLGraphBuilder {
  MLOperand instanceNormalization(MLOperand input,
                                  optional MLInstanceNormalizationOptions options = {});
};

dictionary MLLayerNormalizationOptions {
  MLOperand scale;
  MLOperand bias;
  sequence<[EnforceRange] unsigned long> axes;
  float epsilon = 1e-5;
};

partial interface MLGraphBuilder {
  MLOperand layerNormalization(MLOperand input,
                               optional MLLayerNormalizationOptions options = {});
};

dictionary MLLeakyReluOptions {
  float alpha = 0.01;
};

partial interface MLGraphBuilder {
  MLOperand leakyRelu(MLOperand input, optional MLLeakyReluOptions options = {});
  MLActivation leakyRelu(optional MLLeakyReluOptions options = {});
};

dictionary MLLinearOptions {
  float alpha = 1;
  float beta = 0;
};

partial interface MLGraphBuilder {
  MLOperand linear(MLOperand input, optional MLLinearOptions options = {});
  MLActivation linear(optional MLLinearOptions options = {});
};

enum MLLstmWeightLayout {
  "iofg", // input-output-forget-cell gate ordering
  "ifgo"  // input-forget-cell-output gate ordering
};

dictionary MLLstmOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  MLOperand peepholeWeight;
  MLOperand initialHiddenState;
  MLOperand initialCellState;
  boolean returnSequence = false;
  MLRecurrentNetworkDirection direction = "forward";
  MLLstmWeightLayout layout = "iofg";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> lstm(MLOperand input,
                           MLOperand weight,
                           MLOperand recurrentWeight,
                           [EnforceRange] unsigned long steps,
                           [EnforceRange] unsigned long hiddenSize,
                           optional MLLstmOptions options = {});
};

dictionary MLLstmCellOptions {
  MLOperand bias;
  MLOperand recurrentBias;
  MLOperand peepholeWeight;
  MLLstmWeightLayout layout = "iofg";
  sequence<MLActivation> activations;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> lstmCell(MLOperand input,
                               MLOperand weight,
                               MLOperand recurrentWeight,
                               MLOperand hiddenState,
                               MLOperand cellState,
                               [EnforceRange] unsigned long hiddenSize,
                               optional MLLstmCellOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand matmul(MLOperand a, MLOperand b);
};

enum MLPaddingMode {
  "constant",
  "edge",
  "reflection",
  "symmetric"
};

dictionary MLPadOptions {
  MLPaddingMode mode = "constant";
  float value = 0;
};

partial interface MLGraphBuilder {
  MLOperand pad(MLOperand input,
                sequence<[EnforceRange] unsigned long> beginningPadding,
                sequence<[EnforceRange] unsigned long> endingPadding,
                optional MLPadOptions options = {});
};

enum MLRoundingType {
  "floor",
  "ceil"
};

dictionary MLPool2dOptions {
  sequence<[EnforceRange] unsigned long> windowDimensions;
  sequence<[EnforceRange] unsigned long> padding;
  sequence<[EnforceRange] unsigned long> strides;
  sequence<[EnforceRange] unsigned long> dilations;
  MLInputOperandLayout layout = "nchw";
  MLRoundingType roundingType = "floor";
  sequence<[EnforceRange] unsigned long> outputSizes;
};

partial interface MLGraphBuilder {
  MLOperand averagePool2d(MLOperand input, optional MLPool2dOptions options = {});
  MLOperand l2Pool2d(MLOperand input, optional MLPool2dOptions options = {});
  MLOperand maxPool2d(MLOperand input, optional MLPool2dOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand prelu(MLOperand input, MLOperand slope);
};

dictionary MLReduceOptions {
  sequence<[EnforceRange] unsigned long> axes;
  boolean keepDimensions = false;
};

partial interface MLGraphBuilder {
  MLOperand reduceL1(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceL2(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceLogSum(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceLogSumExp(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceMax(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceMean(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceMin(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceProduct(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceSum(MLOperand input, optional MLReduceOptions options = {});
  MLOperand reduceSumSquare(MLOperand input, optional MLReduceOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand relu(MLOperand input);
  MLActivation relu();
};

enum MLInterpolationMode {
  "nearest-neighbor",
  "linear"
};

dictionary MLResample2dOptions {
  MLInterpolationMode mode = "nearest-neighbor";
  sequence<float> scales;
  sequence<[EnforceRange] unsigned long> sizes;
  sequence<[EnforceRange] unsigned long> axes;
};

partial interface MLGraphBuilder {
  MLOperand resample2d(MLOperand input, optional MLResample2dOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand reshape(MLOperand input, sequence<[EnforceRange] unsigned long> newShape);
};

partial interface MLGraphBuilder {
  MLOperand sigmoid(MLOperand input);
  MLActivation sigmoid();
};

partial interface MLGraphBuilder {
  MLOperand slice(MLOperand input,
                  sequence<[EnforceRange] unsigned long> starts,
                  sequence<[EnforceRange] unsigned long> sizes);
};

partial interface MLGraphBuilder {
  MLOperand softmax(MLOperand input);
  MLActivation softmax();
};

partial interface MLGraphBuilder {
  MLOperand softplus(MLOperand input);
  MLActivation softplus();
};

partial interface MLGraphBuilder {
  MLOperand softsign(MLOperand input);
  MLActivation softsign();
};

dictionary MLSplitOptions {
  [EnforceRange] unsigned long axis = 0;
};

partial interface MLGraphBuilder {
  sequence<MLOperand> split(
      MLOperand input,
      ([EnforceRange] unsigned long or sequence<[EnforceRange] unsigned long>) splits,
      optional MLSplitOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand tanh(MLOperand input);
  MLActivation tanh();
};

dictionary MLTransposeOptions {
  sequence<[EnforceRange] unsigned long> permutation;
};

partial interface MLGraphBuilder {
  MLOperand transpose(MLOperand input, optional MLTransposeOptions options = {});
};

dictionary MLTriangularOptions {
  boolean upper = true;
  [EnforceRange] long diagonal = 0;
};

partial interface MLGraphBuilder {
  MLOperand triangular(MLOperand input, optional MLTriangularOptions options = {});
};

partial interface MLGraphBuilder {
  MLOperand where(MLOperand condition, MLOperand input, MLOperand other);
};

Issues Index

Document operations susceptible to out-of-bounds access as a guidance to implementers.
Investigate side channel attack feasibility considering the current state where CPU is shared between processes running renderers.
Hinting partially mitigates the concern. Investigate additional mitigations.
Should 0-size dimensions be supported? [Issue #391]
The maximum number of operand dimensions is not defined, but native ML APIs usually have a maximum supported size. [Issue #456]
If MLGraphBuilder can’t be re-used, then this can be simplified: enforce uniqueness in input() instead, and iteration can be done over all of the graph’s inputs instead of needing this traversal. [Issue #567]
If constants' ArrayBuffers are not transferred, make copies for graph's constants here. [Issue #566]
Not all implementations support minValue equal to maxValue. [Issue #396]