search

SeanNaren / warp-ctc

Pytorch Bindings for warp-ctc
Apache License 2.0
757 stars 271 forks source link

loss is converging but giving wer 100 #111

Closed vaibhav0195 closed 4 years ago

vaibhav0195 commented 5 years ago

Hi thanks for this repo, I am trying to implement wave2Letter using this ctc repo. But my loss is way too much,in the beginning it decreases steadily upto a value and then stops. but the model always gives me the blank prediction. can you help me why ?

any help is appreciated Thanks :)

farisalasmary commented 4 years ago

I had a similar problem and I discovered later that the set of characters that I used in the softmax were all uppercase characters whereas my dataset's set of characters were ALL lowercase.