Training on a single GPU

[YEKAI-2022]

你好，我在本机debug代码时，出现并行的问题local_rank和world_size,是参数设置有错误吗？

[YEKAI-2022]

我本机是单机单卡的，像这种debug出现并行问题该怎么解决，

[tian-qing001]

Hello @YEKAI-2022. For training our module on a single GPU, the recommended command is:

python -m torch.distributed.launch --nproc_per_node=1 main.py --cfg <path-to-config-file> --data-path <imagenet-path> --output <output-path>

Upon reviewing the screenshot you shared, it appears that the command you are currently using is functionally equivalent to:

python main.py --cfg <path-to-config-file> --data-path <imagenet-path> --output <output-path>

[tian-qing001]

If your objective is solely to debug the model, consider creating a new script where you can load the model without executing the main.py.

[YEKAI-2022]

torch.distributed.launch --nproc_per_node=1如果debug的话，上述的参数也要和--cfg --data-path --output 加到参数列表吗

[tian-qing001]

@YEKAI-2022 You can refer to sections 3.2 and 3.3 of this blog. It's important to note that these settings are not exclusive to our project and can be found online.

[YEKAI-2022]

Thank you very much for your answer