High Loss while training for Kanji Character about keras-ocr HOT 3 CLOSED

Looking at your screenshot, I can see the training loss is dropping steadily, so I don't think there is a misconfiguration issue with training here. I would try the following (in order).

• It could be that training just needs to go on a bit longer. Let training continue for at least 100 epochs and see if validation loss starts to drop. This means changing the EarlyStopping parameters to make them more permissive (e.g., increase patience to something much higher). Wouldn't it be nice for things to be that simple?
• It could be the training set is not diverse enough and so the model is quickly overfitting. Are you able to apply augmentation somehow to make the training set more diverse?
• It could be the validation set is too different from the training set. I wouldn't be able to help identify this without seeing the data.

I'm sorry for the trouble you're having and look forward to your feedback to try and make the package easier to work with. Thanks!

from keras-ocr.

Closing due to lack of feedback. If the issue persists, please comment back and we can re-open. Thanks!

from keras-ocr.

@Zepharchit is your model working fine now. I am using coco_text_detection dataset which has 40k images but surprisingly the loss is very high just like yours. Please respond if you have found the solution.

from keras-ocr.