search

PatrickLib / captcha_recognize

Image Recognition captcha without image segmentation 无需图片分割的验证码识别
Apache License 2.0
556 stars 173 forks source link

training.py Running Time #23

Closed tomorrow1pan closed 5 years ago

tomorrow1pan commented 5 years ago

I used Tensorflow-CPU. The captcha_train.py had run over 3 hours and the step was over 2700, but the loss was between 2.9 and 3. It decreased very slowly. Can you tell me how long did your training.py run? And, doed it stop until the loss reach zero ? THX

He-haitao commented 5 years ago

i face the same problem

tomorrow1pan commented 5 years ago

After trying several times, I had found that you needn't to wait the train.py to converge and stop. The trained models was saved under folder _captchatrain. And you can run eval.py and test.py at any time, no matter the trainig.py is runnnig or stop. But the training steps will affect the prediction accuracy. You can choose the best trained model. Hope this can help you. @He-haitao

He-haitao commented 5 years ago

Yeah, i know this. but when i run captcha_eval.py, the accuracy is 0.00. my step is about 7000-8000steps and from 2000steps ,the loss is around 3.9, i just saw that they run 20000steps and get a pretty good result, i just don't know why.

He-haitao commented 5 years ago

do you get any result when you run the eval.py?

tomorrow1pan commented 5 years ago

yeah, The accuracy rate of prediction and evaluation was around 45% after 16000 steps training. But at the beginning of training, the result of eval.py and recognize.py was 0%. I didn't know why. Maybe the train.py had to run at least certain steps??? @He-haitao