I'm trying to perform a training of the SODA-A dataset, only on the 'container' category, using this SODA-mmrotate repository and training on a single GPU (Tesla V100 16GB) with this command:
python ./tools/train.py ./configs/sodaa-benchmarks/rotated_retinanet_obb_r50_fpn_1x.py .
The actual train step, the iterations in each epoch, seems quite fast while the subsequent evaluating phase is really slow:
It is stuck on this evaluation step, and I see 2 cores fully working from htop, since hours and assuming it will be executed after each epoch, would require days just for evaluation...
Can you suggest something to perfom a faster evaluation step? Can I use a different evaluation procedure? How can I adjust the number of process used in this phase?
Or, given my goal of training the model only on the containers class to obtain higher AP, is there a better approach?
Sorry for using a 'Feature' tag even if it is not a proper feature proposal, but neither a bug.
Sorry for the late response.
set nproc=10 works well for evaluation in our experiments.
the evaluation spends ~3 hours since the densely packed situation, and we will update the codes for faster evaluation.
Hi all,
I'm trying to perform a training of the SODA-A dataset, only on the 'container' category, using this SODA-mmrotate repository and training on a single GPU (Tesla V100 16GB) with this command:
python ./tools/train.py ./configs/sodaa-benchmarks/rotated_retinanet_obb_r50_fpn_1x.py
.The actual train step, the iterations in each epoch, seems quite fast while the subsequent evaluating phase is really slow:
It is stuck on this evaluation step, and I see 2 cores fully working from htop, since hours and assuming it will be executed after each epoch, would require days just for evaluation...
Can you suggest something to perfom a faster evaluation step? Can I use a different evaluation procedure? How can I adjust the number of process used in this phase?
Or, given my goal of training the model only on the containers class to obtain higher AP, is there a better approach?
Sorry for using a 'Feature' tag even if it is not a proper feature proposal, but neither a bug.
Any suggestions are appreciated, Thanks