Skip to content

Add function body for SoftmaxCrossEntropyLossGrad#10779

Merged
gramalingam merged 21 commits intomasterfrom
rama/SCELossGradFun
Apr 5, 2022
Merged

Add function body for SoftmaxCrossEntropyLossGrad#10779
gramalingam merged 21 commits intomasterfrom
rama/SCELossGradFun

Conversation

@gramalingam
Copy link
Contributor

Description:

Add function body for SoftmaxCrossEntropyLossGrad

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
// label : [B, d1, d2, ...]

// We decompose the forward propagation into two steps, for doing the backward prop.
// Step 1: loss = Neg(Logsoftmax(prediction-for-true-label))
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Neg(Logsoftmax(prediction-for-true-label))

Negative Log-likelihood(NLL)?

@codemzs
Copy link
Member

codemzs commented Mar 7, 2022

ONNX_CONTRIB_OPERATOR_SCHEMA(SoftmaxCrossEntropyLossGrad)

The node we see in the graph is SoftmaxCrossEntropyLossInternalGrad, do we need to make this change there as well? Same for SoftmaxCrossEntropyLossInternal


Refers to: orttraining/orttraining/core/graph/training_op_defs.cc:1419 in 0cfa560. [](commit_id = 0cfa560, deletion_comment = False)

gramalingam and others added 12 commits March 7, 2022 15:05
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
@snnn
Copy link
Contributor

snnn commented Mar 15, 2022

/azp run Linux GPU TensorRT CI Pipeline

@azure-pipelines
Copy link

Azure Pipelines successfully started running 1 pipeline(s).

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
souptc
souptc previously approved these changes Mar 16, 2022
Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>
@gramalingam gramalingam merged commit 2c24088 into master Apr 5, 2022
@gramalingam gramalingam deleted the rama/SCELossGradFun branch April 5, 2022 17:52
seddonm1 pushed a commit to seddonm1/onnxruntime that referenced this pull request May 15, 2022
* Add function definition for SoftmaxCrossEntropyLossGrad

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Cleanup

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Eliminate unused variable

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Fix index of weight tensor

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* A few fixes to handle typing and weight

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Fix for zero D dimensions

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Add function body to internal op also

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* A few fixes

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Fix type variable name

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Fix type constraint var

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Fix ignore_index handling in testcase

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>

* Add fun def for SoftmaxCrossEntropyLossInternal

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>

* Specify opset

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>

* Handle opset in NLL function

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Address PR feedback

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Modify onehot

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

* Eliminate duplicate statement

Co-authored-by: Ganesan Ramalingam <grama@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants