emedvedev/attention-ocr

Space prediction

Closed this issue · 5 comments

Hello,

Say I want to predict an image that contains 2 words with a space between them, something like: 'ABC 1234'. Is attention ocr able to do such prediction? I have tried to generate such sequences and train but I haven't been able to obtain decent results... I am not entirely sure how to store the lables in the txt files considering it has a space, should I use quotation marks? But then is it able to disregard that character and not consider it as part of the label?

Regards,
Vlad

It is possible to do this by using the the --full-ascii argument in the training command.

It is possible to do this by using the the --full-ascii argument in the training command.

is it working by using --full-ascii are you getting spaces in your prediction

Yes, what that does is adding all the ascii characters, including ' ' to your charmap.

Yes, what that does is adding all the ascii characters, including ' ' to your charmap.

How did you created the dataset of multiple words
Can you share training hyperparameters