q-ycong-p

Results 15 comments of q-ycong-p

Memory reporting capabilities are unclear and undocumented.

There is a section in [User Guide](https://github.com/google/benchmark/blob/main/docs/user_guide.md#memory-usage) that briefly mentions the supported functionality of benchmarking memory usage. But the documentation gives no concrete instruction of how to use it. Please...

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

@hwangdeyu The reason why it's converted to loop in this case is because the `SelectV2` in frozen graph pattern is not handled in `lstm_tf2_rewriter` - itself can be skipped to...

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

Hi @hwangdeyu and tf2onnx contributors, This feature keras2onnx has implemented. I tested with the minimal example keras model shared above - [keras2onnx implementation](https://github.com/onnx/keras-onnx/pull/386) gives correct inference results on post-padded inputs....

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

> Thank you so much for so detailed issue, I am not very familiar with this part. I guess the keras2onnx is a good example to fix it. Need to...

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

Hi @hwangdeyu and tf2onnx community, is there any plan on picking up this work soon?

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

Hi @hwangdeyu, could you give us an update? We'll need to work around the urgent request (07/01) we got depending on your plan for this fix. Pls let us know.

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

@hwangdeyu Thanks for the update. Could you provide us with an ETA?

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

@hwangdeyu Given the request on our end, we need to work towards a fix too. Could you give us some pointer/direction so that we can contribute/collaborate?

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

Thanks for the pointer. Regarding the reproduction, I think the minimal example shared at top of post is sufficiently simple to verify wip fix? We can assert both inference result...

Keras LSTM not converted correctly when precedent Embedding layer specifies mask_zero=True

Hi @hwangdeyu, thanks for letting me know. Even inference result is correct, LSTM being converted to loop op means important optimizations cannot be taken advantaged of at inference time. I'm...