troubleshoot.md

environment troubleshooting

pytorch

import torch
print(torch.cuda.is_available())
print(torch.zeros(1).cuda())

if you see error information, it's probably that torch 2 is incompatible with your cuda driver or device. (For example, On 2080Ti and CUDA11.7). You can try torch 1.13.1:

# Warning: you will PROBABLY get different results from ours.
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117

and you will likely need mmcv 1.6.0 to work with torch 1.13.1.

mmcv

If you encounter any of these problems below while training/evaluating:

• mmcv.ext_ does not exist
• DDP multi-card evaluation does not work, saying 'TypeError: object of type 'DataContainer' has no xxx' (reference)

it's because the mmcv is not correctly setup. mmcv heavily relies on your torch version, just install another version.

As we have tested,** mmcv 1.5.2 works with torch 2.0. And for torch 1.13.1, mmcv 1.6.0 works.** (we have no idea why 1.5.2&1.6.2 not compatible with torch 1.13.1, and 1.5.3 does not support ddp eval)

mim install mmcv-full==1.6.0

If you have unfortunately encountered this problem and can only use this workaround, you need to re-install ALL mm-series packages, not just this one.

conda exported yml

we also provide an exported conda environment yml to help you determine if minor package version discrepancies exist.