You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
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:
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.
@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.
Activity
Srikanth-ECE commentedon Jul 10, 2026
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 commentedon Jul 14, 2026
Regarding Framework - Adding pytorch profiler:
@Srikanth-ECE raised a point that appears most actionable:
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 commentedon Jul 16, 2026
If this is CUDA-side, I'd double-check driver/toolkit vs the build, and device placement.
— AMA