Loss not decreasing for Hybrid CNN+DNN and CNN+BLSTM models.

mravanelli / pytorch-kaldi

pytorch-kaldi is a project for developing state-of-the-art DNN/RNN hybrid speech recognition systems. The DNN part is managed by pytorch, while feature extraction, label computation, and decoding are performed with the kaldi toolkit.

2.37k stars 446 forks source link

Hi,

I've been trying to build hybrid models on raw input signals. The raw input signals are sampled with a sampling length of 25ms * 16000Hz = 400 frame width.

These are the two config files I've been using (with some variations). I've tried playing with the dropout, normalizations, tried making the architecture simpler but the loss isn't decreasing.

For CNN + DNN model, the loss is pretty much constant, while for CNN + BLSTM model, the loss is sometimes increasing, sometimes decreasing and sometimes going Nan. Please have a look at the config files, CNN.txt CNN_LSTM_raw.txt

mravanelli / pytorch-kaldi

Loss not decreasing for Hybrid CNN+DNN and CNN+BLSTM models. #223