RNN state makes model output tend towards 0

[Tom2096]

Hi,

I have been following the examples in RNN-Denoise. However, when testing the quantized model, I noticed that the quantized output tends very fast towards zero:

Resetting the states of each RNN layer after each prediction makes the output more reasonable but does not give good noise reduction result:

Is there something I am missing? Note that the model used is the one provided in the repo and rest of the code is identical to the examples in RNN-Denoise. Thanks!

[Tom2096]

Note that the model works well in python keras with stateful=True. The input to the above screenshot is one of the noised audio samples from the training dataset, and is preprocessed with the same methods listed in the repo.

[majianjia]

Hi @Tom2096
Since the training isn't quantization-aware, it's possible that the weights distribution might become extreme. Keras should perform well in this scenario because it uses floating point representations. However, after quantization, there is a loss of weight resolution. Try stopping the training after a very short period, maybe just 1 or 2 epochs, to see if there's an improvement.