-
Notifications
You must be signed in to change notification settings - Fork 1.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
chinese dataset with attn is worse than CTC #107
Comments
@haohaoxiao-dl Is that accuracy measured on val set or test set? I got > 87% accuracy on val but less than 10% accuracy on test :( |
just the accuracy on val set. |
@haohaoxiao-dl I use more than 1 million textboxes. That's should be enough I think |
Should there be any undetected difference between your test set and val set? |
of course these are differences between synthetic data with real data but I didn't expect the result would be too poor like that |
maybe you can fine-tune your trained model by some real data? BTW,I tried to set Transformation to TPS,but the accuracy was always zero during training.Any suggestion about that? |
sorry I'm so noob about this. IMO, the accuracy is zero means that the model does not learn anything. Usually I will check my data or try to tune the learning rate |
I used the same parameters except "Prediction", and I got 60+ accuracy with CTC while the accuracy of attn is no more than 10,than i tried an adam learning rate of 0.0003,than my training result of attn can up to 20+,which still too low.
what can I do to improve the result?
by the way, on account of the arbitrary size of my training data, I set the imgW to 512 and use "pad",and my batch_max_length was set to 64(once tried to set a bigger and a smaller one, but seems no palpable improve on accuracy)
thanks
The text was updated successfully, but these errors were encountered: