This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 #17872
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
There's still some failure in fusedlstm tests. Please take a check. |
|
@zixuanweeei Thanks for your contribution, could you also cherry-pick the commit to 1.7? DJL LSTM model depends on this commit. Thanks! |
Sure, please add this requirement in 1.7 roadmap #16864 |
6 tasks
* Use nd.copy() to initialize parameters of new operator * Add check for output states
zixuanweeei
force-pushed
the
rnn/gradients
branch
from
b34e7a5
to
40c57f3
Compare
|
CI passed. Please take a review. Thanks. @ciyongch @TaoLv @pengzhao-intel Besides, we will backport this patch into v1.7 branch as well @stu1130. |
TaoLv
approved these changes
Apr 11, 2020
zixuanweeei
added a commit
to zixuanweeei/mxnet
that referenced
this pull request
Apr 13, 2020
…che#17872) * Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 * Use nd.copy() to initialize parameters of new operator * Add check for output states * Initialize i2h/h2h_weights with zeros for rnn_relu/tanh, and reduce size * Split fused rnn layer test into tests of individual mode * Skip lstm and gru tests on CPU context without DNNL
pengzhao-intel
pushed a commit
that referenced
this pull request
Apr 15, 2020
* Support projection feature for LSTM on CPU (Only Inference) (#17702) * Support projection feature for LSTM on CPU * test solution for -Werror=maybe-uninitialized * Check device type when create state * Document the projection feature of LSTM for RNN operator * Minor fix * Re-run CI * Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 (#17872) * Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 * Use nd.copy() to initialize parameters of new operator * Add check for output states * Initialize i2h/h2h_weights with zeros for rnn_relu/tanh, and reduce size * Split fused rnn layer test into tests of individual mode * Skip lstm and gru tests on CPU context without DNNL
stu1130
pushed a commit
to stu1130/incubator-mxnet
that referenced
this pull request
Apr 15, 2020
…18038) * Support projection feature for LSTM on CPU (Only Inference) (apache#17702) * Support projection feature for LSTM on CPU * test solution for -Werror=maybe-uninitialized * Check device type when create state * Document the projection feature of LSTM for RNN operator * Minor fix * Re-run CI * Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 (apache#17872) * Fix issue of zeros gradients w.r.t. RNN bias when num_layers > 1 * Use nd.copy() to initialize parameters of new operator * Add check for output states * Initialize i2h/h2h_weights with zeros for rnn_relu/tanh, and reduce size * Split fused rnn layer test into tests of individual mode * Skip lstm and gru tests on CPU context without DNNL
7 tasks
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Patch for the issue #17818. The rnn operator produces zero gradients for bias when num_layers > 1. It is caused by a mistake in calculating the shift of bias pointer, where we used the size of fusion bias (i2h_bias + h2h_bias) but MXNet gives twice (i2h_bias, h2h_bias) as many as the fusion size.
Checklist
Changes
@ciyongch @pengzhao-intel @TaoLv