Add GRU and corresponding operators by wchao1115 · Pull Request #83 · webmachinelearning/webnn

wchao1115 · 2020-08-16T20:46:02Z

Fill the operator gaps to support noise suppression first-wave models as outlined here.

Adding GRU and GRUcell operators to support GRU recurrent network. Defining a cell operator in addition to the network operator for added customization flexibility e.g. to support stacked cell recurrent network, etc.

#66

Preview | Diff

…utput to return the whole output sequence in addition to the last output.

…t atop the cell e.g. stacked cells, etc.

anssiko · 2020-08-17T18:48:19Z

 </div>

+### gru ### {#api-neuralnetworkcontext-gru}
+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 the network, as outlined in this [paper](https://arxiv.org/pdf/1406.1078.pdf).


I suggest you make this link a biblio reference (add into <pre class="biblio">) and point to it instead. This allows the reader to see all the references at https://webmachinelearning.github.io/webnn/#references and makes a clear distinction between normative ([[!foo]]) and informative references ([[foo]]).

(I notice this "in this paper" convention is used in batchNormalization prose as well, so that could be aligned similarly as well.)

Good idea. Fixed.

anssiko · 2020-08-17T18:53:04Z

Other than the referencing style comment, LGTM.

I'll defer to @huningxin for the API arguments and returns definitions.

wchao1115 · 2020-08-17T19:09:21Z

@pyu10055 please have a look also.

anssiko · 2020-08-18T10:12:09Z

@pyu10055 now has the superpower to assign himself as a reviewer for PRs. Also the editors and the chair can assign him. Thank you @pyu10055 for your help in reviewing changes to the API spec.

Note to my future self: in order to be able to assign folks as reviewers with the built-in feature, an individual must be invited as a member of the GH org (here webmachinelearning) and granted at minimum read permissions to the repository in question. See https://docs.github.com/en/github/managing-your-work-on-github/assigning-issues-and-pull-requests-to-other-github-users

anssiko · 2020-08-25T14:01:34Z

@pyu10055 I took the liberty to assign you as a reviewer to this PR without your explicit consent. Do you prefer to be added as a reviewer for the WebNN API changes going forward? I'd prefer not to burden you too much with review requests if you have limited bandwidth for reviews currently.

One option is to keep you in the review loop in a non-blocking manner i.e. if we don't hear any concerns from you we consider you're OK with the proposed changes. Would that work for you? Thanks for your help!

wchao1115 · 2020-08-27T05:16:50Z

@gramalingam

huningxin · 2020-08-27T07:40:27Z

+      let cellOutput = null;
+
+      for (let slot = 0; slot < slots; ++slot) {
+        cellHidden.push(nn.squeeze(nn.slice(hidden, [slot, 0, 0], [slot + 1, -1, -1]), [0]));


hidden is not defined. Is it hiddenState?

Super. I did miss renaming that one.

huningxin · 2020-08-27T07:40:59Z

+    The behavior of this operation with default argument values can be generically emulated from the usage of other operations as follow. However, user agents typically have a more efficient implementation for it, therefore its usage is encouraged from the performance standpoint.
+    <pre highlight="js">
+    let hiddenState = initialHiddenState;
+    let slots = (direction == "both" ? 2 : 1);


Would it be good to be const slots?

gramalingam · 2020-08-27T16:04:35Z

+
 ### gemm ### {#api-neuralnetworkcontext-gemm}
-Calculate the <a href="https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3">general matrixmultiplication of the Basic Liner Algebra Subprograms</a>. The calculation follows the expression `alpha * A * B + beta *C`, where `A`, `B`, and `C` are matrices, and `A` and `B` may optionally be transposed prior to the calculation.
+Calculate the [general matrixmultiplication of the Basic Liner Algebra Subprograms](https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3). The calculation follows the expression `alpha * A * B + beta *C`, where `A`, `B`, and `C` are matrices, and `A` and `B` may optionally be transposed prior to the calculation.


Is having matrixmultiplication as one word intentional? Two words seems better to me.

its a typo from the past. Fixed.

gramalingam · 2020-08-27T16:17:40Z

+
+partial interface NeuralNetworkContext {
+  sequence<Operand> gru(Operand input, Operand weight, 
+                        Operand recurrentWeight, Operand initialHiddenState,


Should we make the initialHiddenState optional (with a value 0, if not specified)?

good point. i'll look into it.

Please take a look. initialHiddenState is now optional.

gramalingam · 2020-08-27T18:38:11Z

+        - *recurrentWeight*: an {{Operand}}. The 3-D recurrent weight tensor of shape [num_directions, 3 * hidden_size, hidden_size]. The ordering of the weight vectors in the second dimension of the tensor shape is specified according to the *layout* argument.
+        - *initialHiddenState*: an {{Operand}}. The 3-D initial hidden state tensor of shape [num_directions, batch_size, hidden_size].
+        - *hiddenSize*: a {{long}} scalar. The value of the third dimension of the cell output tensor shape. It indicates the number of features in the hidden state.
+        - *steps*: a {{long}} scalar. The number of time steps in the recurrent network.


What about varying length sequences? Do we want to support them?

@gramalingam I did some research and I couldn't find variable length sequence supported in any other popular frameworks but ONNX. I looked at TensorFlow (and its JS sibling), PyTorch, CoreML and MXNet. No other framework supports this but ONNX. Do you remember why ONNX added this? Can we leave it for now for compactness?

It possibly comes from CNTK.

gramalingam · 2020-08-27T18:43:20Z

+        - *input*: an {{Operand}}. The input tensor.
+        - *starts*: a sequence of {{long}}. The starting indices of the corresponding axes of the input shape.
+        - *ends*: a sequence of {{long}}. The ending indices of the corresponding axes of the input shape.
+            The ending index value of -1 selects all the remaining tensor values from the starting index of the given axis. 


If we want to allow -1, it may be convenient to allow any negative value, interpreted as counting back from the end, and to allow it for starts as well.

@gramalingam I think we're going to have to choose which semantic we want here, the numpy's one or the TensorFlow's one. ONNX follows numpy's convention, which is what you suggested.

The TensorFlow's convention is (start, length) where start is 0-based and length is 1-based. Here length = -1 means the remaining number of elements to the end of the axis dimension.

The numpy's convention is (start, end) where both are 0-based indices, and both are allowed negative values. Here end = -1 means end = length - 1, so a slice of (0, -1) selects all the elements in an axis except the last one. The drawback of this convention is that to select to the end of the axis dimension without knowing the dimension size (length n), one would need to pass an artificially big number like INT_MAX i.e. (0, INT_MAX). That's fine but we now need to agree on what that big number maybe e.g. 32 or 64 bit, signed or unsigned, etc. Numpy has a benefit of a Python's missing argument idiom, so one could simply do (0,) as absence of end implies length n.

The (0,-1) use case of the TensorFlow's convention is quite handy and is used a lot in practice. I'm leaning more toward that convention. Note that PyTorch narrow operator follows the (start, length) convention as well but the spec does not explicitly state that it supports passing -1 as a length.

I agree using -1 to go until the end is convenient. If we treat (start,end) as being inclusive of end, we could support both features. But it could have the disadvantage of causing some confusion if people are used to particular interpretations.

Either way: if the convention is different from ONNX, it will complicate translating between the two. Sigh! I guess we can't make it work for both Tensorflow and ONNX. A compromise is unavoidable. So, I am okay with the proposal

gramalingam · 2020-08-27T18:44:16Z

+Produce a slice of the input tensor.
+<script type=idl>
+partial interface NeuralNetworkContext {
+  Operand slice(Operand input, sequence<long> starts, sequence<long> ends);


Adding a sequence<long> axes may be helpful.

gramalingam · 2020-08-27T20:01:05Z

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


nit: may be use "returnSequence ? [hiddenState, sequence] : [hiddenState]"? There will be a difference only in the case where steps=0.

I guess that does raise another question: the returned value should probably be an empty tensor (of size 0) rather than null.

steps == 0 should not be allowed. I'll add the clarification text on the argument.

…r slice, adding an optional axes argument.

huningxin

Thanks for the update. Looks good to me.

wchao1115 added 8 commits August 15, 2020 13:46

First draft gru with decomposed sample.

1e6cef8

Support PyTorch resetAfter flag for the new gate computation. Allow o…

b8c86a4

…utput to return the whole output sequence in addition to the last output.

Remove the extra first dimension of 1 of the output hidden state.

88c25a4

Support forward, backward and bidirectional RNN.

895d486

Refactor gruCell out of gru. This allows other type of RNN to be buil…

25269c8

…t atop the cell e.g. stacked cells, etc.

Back-fill all the needed operators -- sigmoid, slice, squeeze.

9568b14

Fix a typo

2734784

Generate HTML file.

7657db6

wchao1115 requested review from anssiko and huningxin August 16, 2020 20:46

anssiko requested changes Aug 17, 2020

View reviewed changes

anssiko requested a review from pyu10055 August 18, 2020 09:58

huningxin reviewed Aug 19, 2020

View reviewed changes

Comment thread index.bs Outdated

Comment thread index.bs

Comment thread index.bs

Comment thread index.bs Outdated

Comment thread index.bs

Comment thread index.bs Outdated

Comment thread index.bs

Comment thread index.bs Outdated

Addressing PR feedback

00e1d0c

huningxin reviewed Aug 25, 2020

View reviewed changes

Comment thread index.bs

Comment thread index.bs

anssiko approved these changes Aug 25, 2020

View reviewed changes

wchao1115 mentioned this pull request Aug 25, 2020

Custom operations #6

Open

huningxin reviewed Aug 27, 2020

View reviewed changes

gramalingam reviewed Aug 27, 2020

View reviewed changes

Make initialHiddenState optional. Follow the TensorFlow convention fo…

aa02309

…r slice, adding an optional axes argument.

gramalingam approved these changes Sep 1, 2020

View reviewed changes

huningxin approved these changes Sep 1, 2020

View reviewed changes

wchao1115 merged commit da18fb7 into master Sep 2, 2020

anssiko mentioned this pull request Sep 7, 2020

Noise suppression with DSP+DNN, WebNN and Web Audio API feature gaps w3c/machine-learning-workshop#100

Open

wchao1115 deleted the wchao/gru branch December 22, 2020 18:45

Conversation

wchao1115 commented Aug 16, 2020 • edited by pr-preview Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wchao1115 Aug 22, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

anssiko commented Aug 17, 2020

Uh oh!

wchao1115 commented Aug 17, 2020

Uh oh!

anssiko commented Aug 18, 2020

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

anssiko commented Aug 25, 2020

Uh oh!

wchao1115 commented Aug 27, 2020

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wchao1115 Aug 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wchao1115 Aug 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wchao1115 Aug 29, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gramalingam Aug 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wchao1115 commented Aug 16, 2020 •

edited by pr-preview Bot

Loading

wchao1115 Aug 22, 2020 •

edited

Loading

wchao1115 Aug 27, 2020 •

edited

Loading

wchao1115 Aug 27, 2020 •

edited

Loading

wchao1115 Aug 29, 2020 •

edited

Loading

gramalingam Aug 27, 2020 •

edited

Loading