cannot import name 'container_abcs' from 'torch._six'

[SSSilvester]

Hello, I followed the steps in the README and encountered an ImportError: cannot import name 'container_abcs' from 'torch._six' while running. How should I handle this error? Please translate this for me into English.

[woldier]

This is an import error, could you please provide more details about the error?
@SSSilvester

[SSSilvester]

Thank you for your reply. Here is my specific error message.

[woldier]

Unfortunately, you are providing the error logs from torch.dist, and there is no valuable information here
@SSSilvester

[SSSilvester]

when I change (/usr/local/lib/python3.8/dist-packages/torch/_six.py)，add "import collections.abc as container_abcs" can solve this problem. However,when I use the train command"CUDA_VISIBLE_DEVICES=0,1,2,7 PORT=10985 ./tools/dist_train.sh configs/SiamSeg/siamseg_daformer_sepaspp_mitb5_512x512_40k_PotsdamIRRG_2_VaihingenIRRG.py 4" ,the Train stops in

In addition to that, I checked my GPU memory usage, and it only occupies a small part. Moreover, my training logs are also strange, with a lot of repeated initializations. Is this reasonable? You're welcome once again for your reply.

[SSSilvester]

20241225_203044.log

[woldier]

Is ok with A lot of repeated initializations. This is just an exp log.

[SSSilvester]

It still stops in “Use DDP Wrapper” and it only occupies a small part of GPU memory.

[woldier]

I suggest u reduce batch size and try again.

[SSSilvester]

I use samples_per_gpu=2, workers_per_gpu=2, But my process is still stuck there without proceeding.

[woldier]

I'm very curious if the version of torch&&torchvision you're using is the same as the one mentioned in the README.

[SSSilvester]

It can be seemed in my train log.

[woldier]

This is very strange, I have an idea, you use a single GPU and do a single step debugging to see where the stuck position is.

[SSSilvester]

This is very strange, I have an idea, you use a single GPU and do a single step debugging to see where the stuck position is.

I will immediately try training with a single GPU. Thank you very much for your help and suggestions. Wishing you academic longevity.

[woldier]

Thanks🙂, it's been a pleasure talking to you.

[SSSilvester]

Hi!🙂 I have successfully run the code using a single GPU, but I've encountered an issue during the class_mix step. It appears that the process is treating the boundaries of the source domain ground truth as an additional category for class mixing. Could you please help me understand why this might be happening?