Debug run on CPU

[VincentCoulombe]

Hi,

I have a bug in my data processing pipeline, and it is very hard to debug as all the tensors are put on the GPU automatically.

Reading through the documentation, I have not found how to force the run to be on CPU so that I can have a meaningful traceback when an error is raised.

Does anybody know how?

Thank you,
Vincent

[AwePhD]

Hi,

Here in the doc of mmdet.
Namely, use CUDA environment variable to hide your GPU to torch/mmengine interface with CUDA_VISIBLE_DEVICES=-1.