Closed zqalex closed 7 months ago
May be some package versions are different between docker and conda? E.g. cuda, torch, spconv, mmcv
May be some package versions are different between docker and conda? E.g. cuda, torch, spconv, mmcv
The conda envirnment is based on the env.yml you shared, and the docker file is from the project code. Actually there are some difference between them, such as MMEngine. But I am not sure, and trying to train and test in the docker container again.
The difference between conda envirnment and the docker is the reason. But I am not sure what difference caused this problem.
Have you solved?
When first testing SCANNET in a docker container:
All metrics are normal values, such as:
But after the same command in the conda environment, even though everything worked fine in the train step, all metrics are 0 at test, for example:
I also tested the model trained in the conda environment in the docker container, but all metrics turned to 0. I wonder why? I just changed num_workers=3 in oneformer3d_1xb4_scannet.py. File structure:
log when testing in conda environment: