Zero op for hidden state is not working

domerin0 / rnn-speech

Character level speech recognizer using ctc loss with deep rnns in TensorFlow.

MIT License

77 stars 31 forks source link

Closed AMairesse closed 7 years ago

AMairesse commented 7 years ago

The rnn_state_zero_op operation in AcousticModel.py does not set the hidden_state to zeros like it should.

How to reproduce :

launch a training with size ordered set to true and a small batch size
check for the Error Rate on the Training in tensorboard : it should be decreasing at an almost steady pace but there is a bump after every time a checkpoint as been made. The state is reseted only then (because the model is rebuilt)

AMairesse commented 7 years ago

Found some interesting way to deal with variables to keep the state of the RNN : https://stackoverflow.com/questions/38441589/rnn-initial-state/41239965#41239965