Fix keras bidirectional merge failures #1869
Merged
Conversation
This file contains hidden or 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
Signed-off-by: Kotaro Yamamoto <kota.crk@gmail.com>
| This pull request introduces 1 alert when merging 00562aa into bf4a22d - view on LGTM.com new alerts:
|
support below cases: - there are one or two Identity layers between input/output and RNN - Transpose-Reverse-backward (previously, only Reverse-Transpose-backward was supported) - return_sequences=False with no Reverse after the backward Signed-off-by: Kotaro Yamamoto <kota.crk@gmail.com>
kota-row force-pushed the fix_bidirectional branch from 00562aa to ce644d2 Compare 
fatcat-z reviewed Apr 2, 2022
| | ||
| if len(reverse_or_slice_nodes) == 1 and reverse_or_slice_nodes[0].type == "Identity": | ||
| reverse_or_slice_nodes = g.find_output_consumers(reverse_or_slice_nodes[0].output[0]) | ||
| |
Collaborator
There was a problem hiding this comment.
Do we need to do it twice here? Or update it like:
if len(reverse_or_slice_nodes) == 1 and reverse_or_slice_nodes[0].type == "Identity": reverse_or_slice_nodes = g.find_output_consumers(reverse_or_slice_nodes[0].output[0]) if len(reverse_or_slice_nodes) == 1 and reverse_or_slice_nodes[0].type == "Identity": reverse_or_slice_nodes = g.find_output_consumers(reverse_or_slice_nodes[0].output[0]) Contributor Author
There was a problem hiding this comment.
yes, tensorflow makes one or two Identity, it depends on Layer arguments.
I changed the code to nested if.
fatcat-z reviewed Apr 2, 2022
tf2onnx/rewriter/rnn_utils.py Outdated
| | ||
| if rnn_output_index == 0: | ||
| axis = 1 | ||
| # remove reverse op for rnn_bw |
Collaborator
There was a problem hiding this comment.
Please update the comments accordingly to involve tailed slice node.
- Consecutive Identity checks changed to nested if - update comment for remove Reverse or tail-slice op Signed-off-by: Kotaro Yamamoto <kota.crk@gmail.com>
fatcat-z approved these changes Apr 2, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
2 participants
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.
Fixed an issue where the Bidirectional layers generated by Keras were not correctly combined into a single LSTM/GRU.
currently biltm_rewriter/bigru_rewriter only supports the following structure
but, tf2 keras Bidirectional Layer generates different graph
I also added a test to check that all Bidirectional recurrent layers are combined into one. The first commit will fail the test.