Enabling the layer normalization decrease the performance of the model

vrenkens / nabu

Code for end-to-end ASR with neural networks, build with TensorFlow

MIT License

108 stars 43 forks source link

Enabling the layer normalization decrease the performance of the model #44

Open AzizCode92 opened 6 years ago

AzizCode92 commented 6 years ago

I have tried to add layer normalization to the layers.py by setting layer_norm to True as a default value. I used the "listener.py" as my encoder. Those modifications decreased the performance of the model ( training time is longer, both accuracy and WER decreased dramatically). I have seen other users complaining about such problems with tf.contrib.rnn.LayerNormBasicLSTMCell
in stackoverflow Am I missing something else?

vrenkens commented 6 years ago

I have experienced the same issues... Maybe it is better to use the tf.contrib.rnn.LSTMCell instead

AzizCode92 commented 5 years ago

Hi Vincent, I have used the TIMIT dataset and I used the recipe of LAS and it worked fine for my case. I have attached my validation loss (the blue one is without layer normalization and the orange one is with layer normalization). screen shot 2018-12-03 at 12 56 53

Why it didn't work for you case? P.S: I used the KALDI-ASR project to process the dataset

with best regards, Aziz

AzizCode92 commented 5 years ago

Previously, I have pointed that it didn't work with the librispeech but I do have doubts that it is due to some mistake I might have done during the processing.