WU Chen

@kortschak Some characters (such as ÀÉÈÎÙÔ) is also cut off in title and y axis label. 

Hello, is there a workaround for this now? Due to the file size limitation, I have to use Cloudflare Workers instead of Cloudflare Pages.

The actual reason for my implementation of CTC loss becoming NaN after iterations is that the logarithm of zero is taken, not due to `slice_assign`. Now, I am not certain...

@nathanielsimard I'm trying to train a CRNN model from scratch, but after a day of training, there's still no sign of convergence in the model. Then I noticed that only...

> There are _no_ mutable operation in the tensor API, every operation returns the result that should be used! Thank you very much for pointing out the actual problem! I...

I believe the result is now the same as PyTorch. But the performance of this implementation seems to be less than ideal.

The implementation doesn't work on the NdArray backend because of #1053. It also doesn't work on the LibTorch backend because of #1055. I believe the current performance bottleneck lies in...

@louisfd Thank you! The current implementation still significantly consumes graphics memory. I believe that separately calculating the alpha values for blanks and letters can significantly reduce the graphics memory usage,...

@louisfd Thank you! Now `repeat()` is much faster.

I tried to use this implementation of ctc loss in the CRNN model, but after the first iteration loss became NaN. I don't know what went wrong. [wcshds/crnn-cjk](https://github.com/wcshds/crnn-cjk)