Open JonTriebenbach opened 2 years ago
@JonTriebenbach I had the same issue with tensorflow==2.7.0 and tf2onnx==1.9.2.
A simple upgrade from to tf2onnx==1.9.3 solved the issue, that lstms are shown as loops. At least for me.
This PR Add Keras LSTM support involved in tf2onnx verison == 1.9.3
.
So This issue should be fixed after tf2onnx == 1.9.3 theoretically.
I also tried to test it in tensorflow==2.7.1 and tfonnx == 1.9.3/1.10.0, and the lstms are both LSTM op instand of loop..
Could you check it again? Thanks.
@hwangdeyu Yes, we have other LSTM models that now convert as expected with tf2onnx 1.9.3. Thank you for those fixes that corrected those other models. The problem is that this new model is not converting as expected. The problem shows on both 1.9.3, and on the latest 1.10.0.
I do not know what is different about this new model that is not converting as expected. We have other models that have multiple LSTM layers, which convert as expected, this model is only a single layer LSTM model, which indicates it is not a multi-layer model issue. There is some other complexity in this new model that is causing a conversion issue.
Please continue to investigate. Thank you.
any updates on this one?
Is there any progress with this issue?
Is there any ETA on this issue?
Is there any ETA on this issue?
Not yet.
The attached example shows a Keras model with a single layer LSTM that is converted to a loop instead of the expected ONNX LSTM op.
Test system configuration: Using
tensorflow/tensorflow:2.7.1
docker image:This issue appears to be similar to tf.keras.layers.LSTM not converted to ONNX LSTM layer #1546 with pull request Add Keras LSTM support .
onnx_issue.zip
Please review this issue. Thanks.