Imbalanced classes about cnn_lstm_ctc_ocr HOT 3 CLOSED

I don't think what you suggest is the right approach. Two issues come to mind. CTC ignores the location of the individual elements in the input image, so I cannot see how you would practically target individual sequence items. If you could, one ex post facto way to adjust your prior in a learned conditional model to use Bayes' rule.

P(class|obs) = P(obs|class) * P(class) / P(obs)
P'(class|obs) = P(obs|class) * P'(class) / P(obs) = P(class|obs) * P'(class) / P(class)

That is, you can adjust your discriminant function (in a non-sequential model) by rescaling with the ratio of training data to test data distributions.

from cnn_lstm_ctc_ocr.

Thanks for answering!
So, if I get you right, we may have 100 ‘q’ in training set and we know that it is not enough. In the test data it may be 10,000 ‘q’. So let say we have trained our model with what we have, and on the last step, we run the model and just multiply output probability ‘q’ by P’(q)/P(q)
where
P’(q) is 10,000/total_number_of_symbols_in_test
and
P(q) is 100/total_number_of_symbols_in_train

So, in that way we a little bit increase/decrease output probability of ‘q’.
It sounds tricky to me. Could you please tell if i understand you right?
And what you can say about MJSynth dataset imbalance? The class/symbols in that dataset not balanced either. Isn’t it affect to the results when we test the model trained on MJsynth on the other datasets?

from cnn_lstm_ctc_ocr.

There are no easy solutions. Any sequence model, such as the one in this repo, will perform best on test data with the same statistics (as captured by the model) as the training data.

from cnn_lstm_ctc_ocr.