Questions about NaN

[TwitchOnly111]

I was training on tampCOCO, and after the first layer of RGB and DCT streams when I used the pre-trained model, the data became NaN and a lot of zeros, while without the pre-trained model it was some big numbers.

before：

after：

[TwitchOnly111]

Same as issue 10

[BaoanhTrinh]

You can use python>=3.9 and suitable torch version. It works for me

[0011szy]

,同问，这种情况怎么处理