Framework - Adding pytorch profiler

[punit121]

When I try to run the profiler on train_net.py within my new conda environment, it fails because it can't load the COCO dataset. It seems like there's a problem with how my code is trying to find and use the dataset.

[Srikanth-ECE]

Hi @punit121,

I investigated this issue by tracing the data loading flow in the Mask R-CNN implementation:

tools/train_net.py → make_data_loader() → build_dataset() → DatasetCatalog → paths_catalog.py

From what I found, the COCO dataset path is resolved through maskrcnn_benchmark/config/paths_catalog.py, where DatasetCatalog.DATA_DIR defaults to /datasets. The dataset loader then expects paths such as:

/datasets/data/train2017
/datasets/data/annotations/instances_train2017.json

Based on the current information, I couldn't determine whether this is a code issue or an environment/configuration issue.

Could you please provide:

The full error traceback
The command you used to run train_net.py
Your dataset directory structure (or where your COCO dataset is located)
Your OS and PyTorch/CUDA versions

With those details, it will be much easier to determine whether there's a bug in the framework or a dataset configuration issue.

[2000pd3rvr]

Regarding Framework - Adding pytorch profiler:

@Srikanth-ECE raised a point that appears most actionable:

Hi @punit121, I investigated this issue by tracing the data loading flow in the Mask R-CNN implementation: tools/train_net.py → make_data_loader() → build_dataset() → DatasetCatalog → paths_catalog.py From what I found, the COCO dataset path is resolved through maskrcnn_benchm...

I recommend confirming that observation with a minimal reproduction, then documenting a workaround or preparing a focused change if the root cause is clear.

— AMA

[2000pd3rvr]

If this is CUDA-side, I'd double-check driver/toolkit vs the build, and device placement.

— AMA