[single_stage_detector] Updated run_and_time.sh for customizing number of GPUs on single node

[jamesljlster]

Dear MLCommons team,

I appreciate your work, which has helped me verify the training performance after applying hardware resource virtualization to our bare-metal server.

I want to validate the model training performance in a reasonable time on both virtualized and non-virtualized environments. However, this is very challenging with the default Open Images dataset and our relatively small GPUs in a single node. Thus, I made some changes to the performance evaluation script, which may help others with similar use cases.

This pull request modifies the run_and_time.sh of SSD training, and introduces the following changes:

• Added DATASET environment variable for dataset customization. This allows the script to run the training script with datasets other than the default Open Image dataset. (For example, I use the coco dataset.)
  (Reverted: [single_stage_detector] Updated run_and_time.sh for customizing number of GPUs on single node #808 (comment))
• Changed --nproc_per_node=1 of torchrun command line to --nproc_per_node=${DGXNGPU}. This enables the script to do training with more GPUs on a single node.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[jamesljlster]

recheck

[jamesljlster]

recheck

[ShriyaRishab]

MLPerf requires us to use the same dataset as in the reference to ensure results are comparable across submissions. So we should not make it a changeable parameter

[jamesljlster]

Understood. I've reverted the dataset customization commit.