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

i used chinese to run the code,but there has something wrong? acc is o,avg_train_cost <0? #24

Closed dotsonliu closed 7 years ago

dotsonliu commented 7 years ago

image

image

dotsonliu commented 7 years ago

@ilovin ,i dont know whats wrong. why is this? i also use chinese index which coverd all chinese char. the dataset has 50w images. i donot know whats wrong? image

ilovin commented 7 years ago

assume you are using the beta version. the name of your image cannot match the regex expression(read the README) or you can modify the code in https://github.com/ilovin/lstm_ctc_ocr/blob/dev/lib/lstm/utils/tf_records.py#L54 if you are using the master version, you may also modify the related code. make a comment if it works.

dotsonliu commented 7 years ago

@ilovin i use chinese charset，how many images you advise？ i used 1000w imags to train the model（master version），its too slower，is beta better?

ilovin commented 7 years ago

it is much better, however you need to modify the code from generating image to reading image

dotsonliu commented 7 years ago

@ilovin I has modify the code,when i train the crnn(master),i use 300w+ chinese images, but when epoch is 2,lr is too small,near zero, and acc is 57%,not go better,how can i resolve the problem? ![Uploading image.png…]()

dotsonliu commented 7 years ago

image

ilovin commented 7 years ago

read the readme, change the super parameter https://github.com/ilovin/lstm_ctc_ocr#some-details