loss value stays constant during the training

mcf06 / theano_ctc

Theano bindings for Baidu's CTC library.

BSD 3-Clause "New" or "Revised" License

20 stars 5 forks source link

Open HariKrishna-Vydana opened 7 years ago

githubnemo commented 7 years ago

I have not experienced this and the mini batch size is passed to warp-ctc. Can you give a minimal example demonstrating this effect?

HariKrishna-Vydana commented 7 years ago

@githubnemo i am using a minibatch size of 10,running on keras the error is drastcically high around 1000,which is not the case for batch_size of 1

githubnemo commented 7 years ago

OK but can you share your code that reproduces this behavior?