CTCLoss predicts all blank characters

vision
1

I am using crnn.pytorch (https://github.com/meijieru/crnn.pytorch), Pytorch 0.4.0 and SeanNaren/warp-ctc to recognize handwritten documents. The training loss does not decrease and the model predicts all BLANK characters.

2

It turns out that the learning rate should be relatively small (e.g. 0.001 instead of 0.01).