search

Holmeyoung / crnn-pytorch

Pytorch implementation of CRNN (CNN + RNN + CTCLoss) for all language OCR.
MIT License
378 stars 105 forks source link

I tried to reproduce your code, but failed, cost stuck at 7.0 #40

Closed jaysimon closed 4 years ago

jaysimon commented 4 years ago

I tried to reproduce your code, but failed, cost stuck at 7.0. Code is here: https://github.com/jaysimon/crnn-houwei

By using SyntheticChineseStringDataset, your code runs well, cost decrease from 8.0 to 2.0 or less. Your model predicts well. While I tried to build the model by myself, cost stuck at 7.0. Any idea or advice? Thank you very much.

jaysimon commented 4 years ago

I tried almost everything in your repository: reshape to (32, 100), normalized to (-1.0, 1.0), same network, same ctcloss, same lmdb data, same net weight init. But failed, I dont know why.

Holmeyoung commented 4 years ago

Hi, what did you change in the net, I see no model file here https://github.com/jaysimon/crnn-houwei/tree/master/model

Maybe you can clone it first and change it little by little, see what will happen~~~

jaysimon commented 4 years ago

Hi, what did you change in the net, I see no model file here https://github.com/jaysimon/crnn-houwei/tree/master/model

Maybe you can clone it first and change it little by little, see what will happen~~~

model is in the train.py. I will check it carefully. Thanks a lot!