Keras 3 LSTM layer's option use_cudnn=False doesn't work correctly.

[tosa-no-onchan]

Hi Keras!
About https://keras.io/api/layers/recurrent_layers/lstm/
option
use_cudnn=False

This option doesn't work correctly on some situation.

On single use is OK.
x = layers.LSTM(128, return_sequences=True,use_cudnn=False)(x)

But on complex is not.
x = layers.Bidirectional(layers.LSTM(128, return_sequences=True,use_cudnn=False))(x)

The layers.Bidirectional() calls layers.LSTM 3 times at one operation.
use_cudnn=False <-- first call is good.
use_cudnn=Auto <-- second call is no good.
use_cudnn=Auto <-- third call is no good.

Thus, how do i resolve this problem?
I think.
An one way is to change LSTM init param in direct.

class LSTM(RNN):
    ....
    def __init__(
        ....
        #use_cudnn="auto",
        use_cudnn=False,

The other way is,I want,
to add a use_cudnn option to Model.compile option like jit_compile option.

    model.compile(
        ...
        jit_compile=False,
        use_cudnn=False,
    )

My PC
Ubuntu 22.04
Python 3.10.12
virtual env
keras 3.7.0
tensorflow 2.16.2
cuDNN v8.9.6

I am looking forward for your answer.

[mehtamansi29]

Hi @tosa-no-onchan -

Thanks for reporting the issue. Can you help me with sample code to reproduce this issue ?

[tosa-no-onchan]

Thank you for your reply.

Sure!

My program is
https://github.com/tosa-no-onchan/opp_with_lstm

My use model is
https://github.com/tosa-no-onchan/opp_with_lstm/blob/main/model.py

Original program is
https://pylessons.com/speech-recognition