search

emedvedev / attention-ocr

A Tensorflow model for text recognition (CNN + seq2seq with visual attention) available as a Python package and compatible with Google Cloud ML Engine.
MIT License
1.08k stars 256 forks source link

Parameters are changed when testing #127

Closed roger12337 closed 5 years ago

roger12337 commented 5 years ago

Don't know whether you have faced similar situation. After I have trained model (perplexity = 1.1), I do some testing on other images. After I continue to train the model with the same data, parameters are quite different and has perplexity of 5. Does anyone know why this occurs?

roger12337 commented 5 years ago

I explore a little more. The performance of the model is consistent when both training and testing data has same max-width and max-height.

emedvedev commented 5 years ago

That would make sense. Good accuracy with images of different sizes would require a large and extensive dataset, and if your testing data has image sizes that are completely different from training, you won't see good results.

I'll close the issue for now, but if there's any issues with the model itself, do let me know!