search

ilovin / lstm_ctc_ocr

Use CTC + tensorflow to OCR
https://ilovin.github.io/2017-04-06/tensorflow-lstm-ctc-ocr/
354 stars 140 forks source link

why printed “decode” same？ #13

Closed BrianGong closed 7 years ago

BrianGong commented 7 years ago

hi， thanks for your code， but I run the code in TF 1.2 get the print information like this： image

after training for 8 hours ， the accuracy is 0 ,and sevral “decoded” in seq are same. is that correct?

ilovin commented 7 years ago

No, if you cannot see a significant change in the decoded code for about epoch, you may consider adjusting parameters. I notice that the "origin" code are ranging in 0-10, are you using your own data? If so, how many pics are you using?

BrianGong commented 7 years ago

thanks for your help!!! use more training data（128k is ok）and wait longer time ，now the accuracy close to 1. but I run the test.sh ,cant restore model from check point folder. the error like this image ,I read the code and test train.sh modify "restore = 0 " to "resotre = 1",the code works well, do you hava and idea? I use tensorflow 1.2.

ilovin commented 7 years ago

I forget to change the LSTM_test, the network in LSTM_test shall be the same as LSTM_train

ilovin commented 7 years ago

If you do not have further question please close the issue