Shinji Watanabe

The youtube video is dynamic, and we cannot fully avoid the issue... Anyway, I reported this issue to the main organizers already. I recommend you to contact [email protected] They will...

If this often happens during training and training is stopped, your modification needs some more tuning. If this rarely happens (only a couple of times in each epoch), you can...

This happens when the encoder output is shorter than the label length, or you just set the wrong optimizer. Please check both cases. The second case is specific to your...

I don't think this is sudden. There is some small corruption already around 10 epochs, and the loss gradually increases (training is not regular). Due to this wrong training behavior,...

The attention loss is computed from the encoder and decoder networks. So, it would be changed if you only fix the encoder network.

> Thanks, @sw005320, though I was freezing both the encoder and decoder networks, it was still not working, but I was able to finally resolve the issue using `optimizer.zero_grad(set_to_none=True)` in...

Yes, frankly, I'm also confused, and it would be great if @rishabh135 clarifies it.

Thanks for the report! @neillu23, can you take a look at this?

> All of prediction text are same. Like prediction :一线城市的房地产市场仍然有一定的影响力的一线城市的一线城市土地市场成交量的一线城市一线城市的一线城市 I have checked that ctc is work. And I have tried to adjust decode coef ,but it didn't work. I...

Did you use your own data? I guess your training/test split may have some issues.