Closed tumbleintoyourheart closed 5 years ago
Yes, you are right!
In main.py
there is an evaluation right before the training loop: evaluator.evaluate()
.
So from your answer, the predicted sequence output at this step is supposed to be garbage, right?
(In my case it outputs all random Latin characters in a Japanese/Chinese text recognition problem)
I have done some Japanese/Chinese ocr with CRNN+CTC and the outputs at the very first steps were all blank characters (which is reasonable when using CTC loss). When we are using the Attention decoder, is the above-mentioned phenomenon normal?
For attention-based methods, it is normal to output random characters at the very first steps.
I really appreciate the support. Thank you @ayumiymk.
Dear @ayumiymk, Thank you for this repo, In inference time, the script prints out the accuracy, so I want it to print out also the actual predicted sequence. Is it stored in
lib/evaluators.py
pred_list
variable?