Closed mysxs closed 2 months ago
Hi @mysxs, thanks for considering Amphion as your codebase!
I also met the RuntimeError: cuDNN error: CUDNN_STATUS_INTERNAL_ERROR
error. However, usually the error is raised by a wrong input value for Embedding layer or CrossEntroyLoss (see here). I recommend you check the range of x
of your EmotionEncoder, to see whether there is some vocab ids which not ranges from (0, 1582).
By the way, for nn.Embedding
, the first parameter should be the number of vocabulary instead of the input dim (now it is 1583
in your code)?
The best way to debug for cuDNN error
is to use CPU to run your code again. This is because on GPU, the traceback report is not reliable and the bug position can not be accurately found.
Hi @mysxs, thanks for considering Amphion as your codebase!
I also met the
RuntimeError: cuDNN error: CUDNN_STATUS_INTERNAL_ERROR
error. However, usually the error is raised by a wrong input value for Embedding layer or CrossEntroyLoss (see here). I recommend you check the range ofx
of your EmotionEncoder, to see whether there is some vocab ids which not ranges from (0, 1582).By the way, for
nn.Embedding
, the first parameter should be the number of vocabulary instead of the input dim (now it is1583
in your code)?The best way to debug for
cuDNN error
is to use CPU to run your code again. This is because on GPU, the traceback report is not reliable and the bug position can not be accurately found.
谢谢同学,问题解决了,你们真肝,好及时回复问题,辛苦了!
Problem Overview
Hello, I changed the Amohion code, added features, and added the corresponding encoder. The input feature dimension is 1583, and the output feature dimension is 384. The encoder is as follows:
Then the error is as follows. I checked and it seems that there is a problem with embedded id. How can I solve it?
Screenshots
Amphion/modules/diffusion/bidilconv/bidilated_conv.py: