chenjun2hao / bert_ocr.pytorch Goto Github PK

I wonder what data set you're training for. Using the ST data set and MJ data set, I only got 0.68 accuracy on the CUTE80.

Hi, thanks for your awesome repo. I wonder there is any progress? I try to implement this paper too and achieve similar performance with yours, but far away from the paper.

hello, thank you for sharing the codes. I cannot understand the following line in dataset.py:
out_of_char = f'[^{self.opt.character}]'
self.opt.character is 0123456789abcdefghijklmnopqrstuvwxyz
out_of_char is [^**0123456789abcdefghijklmnopqrstuvwxyz**]
can you tell me what the three new characters [ ,^, and ] mean?

我看代码，cnn出来后map H好像变为了1，最后的attention是1D attention?不是2D

when I train it on Chinese dataset, it gives an error: RuntimeError: tabulate: failed to synchronize: cudaErrorAssert: device-side assert triggered,How can I fix it?

对应位置没有这个.txt文件，我应该如何下载？

Dear author, have you visualized the attention masks like the pics in the original paper? Can that really find the location of each character?