An obsession

[731076467]

I always have the problem of "type float can't be cast to the desired output type long" after training several epoch. Sometimes the second epoch reports an error, and sometimes it reports an error on the ninth epoch.

[isotopezzq]

I also run into this problem and after 3 epochs the traing stopped

[Ironieser]

Sry, I tried to reproduce your bug today. But I failed. Everything is OK in my experiment.
At least, in my training, the codes have no any bugs and I used the processed npz as input data rather than raw videos.
Maybe you should check your env and the input data.
The training log as the following figure shown.

[Ironieser]

If you need our help, please upload more details of your env and log. :)

[731076467]

The version of pytorch I use is 1.8.0+cu111, and my device only has a 3080-10G GPU, so the batch size for training is relatively small. Is it possible that this is the problem?

[Ironieser]

The version of pytorch I use is 1.8.0+cu111, and my device only has a 3080-10G GPU, so the batch size for training is relatively small. Is it possible that this is the problem?

No, there is no problem. Maybe others

[isotopezzq]

The version of pytorch I use is 1.8.0+cu111, and my device only has a 3080-10G GPU, so the batch size for training is relatively small. Is it possible that this is the problem?

No, there is no problem. Maybe others
Is there any chance that it would be the problem of dataset? (I noticed that in stu6_44, the label was lost)

[isotopezzq]

I delete the missing data(stu6_44.mp4),and everything goes well

[731076467]

Which file is this file?

[isotopezzq]

in train.csv line 585
delete this line and also delete the video in ./video/train

[731076467]

thank you!let me try!

[Ironieser]

The version of pytorch I use is 1.8.0+cu111, and my device only has a 3080-10G GPU, so the batch size for training is relatively small. Is it possible that this is the problem?

No, there is no problem. Maybe others
Is there any chance that it would be the problem of dataset? (I noticed that in stu6_44, the label was lost)

yeah, I think so.