Fix Edge Case: Shared Layer Output Between Model Output and Internal Layers #2407
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
…nd also the output of the model.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Fix Edge Case: Shared Layer Output Between Model Output and Internal Layers
🎯 Summary
This PR addresses a specific edge case that occurs when converting TensorFlow models to ONNX. It improves the handling of cases where a layer's output is used simultaneously as both the input to another layer and the model's final output.
🔧 Changes
Edge case handling: Resolves the case where a layer's output is used as both the model output and the input to an internal layer.
Improved Transpose output logic:
Uses identity nodes to handle output branching.
Differentiates between model output consumers and internal consumers.
Introduces a two-step process: edge case handling → regular case handling.
Optimizer order optimization: Moves remove_identity to the front of the optimizer chain.
Improved error messages: Provides clearer error messages during graph output validation.
Technical Details
Before:
When a layer's output is connected to both the model output and internal layers, the transpose operation fails, causing the graph structure to break.
After:
Output is branched through identity nodes.
Transpose is applied only to the model output consumer.
Internal consumers maintain the original output.
🔍 Files Changed
tf2onnx/tfonnx.py - Major improvements to the transpose_outputs function.
tf2onnx/graph.py - Improved error messages and code cleanup.
tf2onnx/optimizer/init.py - Changed optimizer order.
Test model file
This file is a test model created by extracting part of the edge case I encountered.
https://drive.google.com/file/d/10ChR5OS4k6az1yG13vdbiassrsxYUkuJ/view?usp=sharing