search

ku21fan / STR-Fewer-Labels

Scene Text Recognition (STR) methods trained with fewer real labels (CVPR 2021)
MIT License
173 stars 27 forks source link

fix bilstm bug #7

Closed simplify23 closed 1 year ago

simplify23 commented 1 year ago

bilstm output 256 for hidden_size, however the feature_extraction output 512 for output channel,this is wrong. SequenceModeling_output need to be out_channel

ku21fan commented 1 year ago

Hi, simplify.

It is not a bug. It is intended, similar to the original CRNN code: the BILSTM makes 512 to 256. https://github.com/bgshih/crnn/blob/master/model/crnn_demo/config.lua#L73

Best