| === Profiling parquet === |
| File: /lustre/fsw/portfolios/dir/projects/dir_cosmos_base_lustre/cosmos3_action_datasets/bridge_raw/data/chunk-000/file-000.parquet |
|
|
| Rows: 1,496,285 Row groups: 2 Columns: 7 |
| Chunk size (15 fps, 76 frames): 76 rows per sample |
| File size on disk: 84.22 MiB (59 bytes/row) |
|
|
| Columns hand_pose uses (present in parquet): |
| task_index |
| episode_index |
| frame_index |
| timestamp |
|
|
| Columns in parquet NOT in hand_pose list: |
| observation.state |
| action |
| index |
|
|
| Arrow table size (in-memory): 146.09 MiB (102 bytes/row) |
| observation.state: 48037.5 KiB (list<element: float>) |
| action: 48037.5 KiB (list<element: float>) |
| timestamp: 6027.5 KiB (float) |
| frame_index: 11872.4 KiB (int64) |
| episode_index: 11872.4 KiB (int64) |
| index: 11872.4 KiB (int64) |
| task_index: 11872.4 KiB (int64) |
|
|
| Full-table read time (1,496,285 rows): |
| All columns (7): 0.0874s (mean of 5) |
| Subset (4 cols): 0.0080s (mean of 5) |
| Subset is 0.09x of full |
| => Subset reads less data; loader column selection could help. |
|
|
| Chunk read (76 rows) at random offsets (in-memory slice, mean of 5): 0.0000s |
|
|
| Note: LeRobot loads the full HF dataset (all columns). Column selection |
| only helps when the omitted columns are a large share of the table. |
|
|
| === Profiling video === |
| Root: /lustre/fsw/portfolios/dir/projects/dir_cosmos_base_lustre/cosmos3_action_datasets/bridge_raw |
| Video key: observation.images.main |
| Profiling 5 video(s), 76 frames each, warmup=1 repeat=10 access=random |
|
|
| /usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you. |
| import pynvml # type: ignore[import] |
| /usr/local/lib/python3.12/dist-packages/torch_automated_profiler/utils/misc.py:13: UserWarning: pkg_resources is deprecated as an API. See https: |
| from pkg_resources import parse_version |
| videos/observation.images.main/chunk-000/file-000.mp4 |
| open+create=0.0018s decode=0.2420s total=0.2438s frames=76 |
| videos/observation.images.main/chunk-000/file-001.mp4 |
| open+create=0.0024s decode=0.2358s total=0.2382s frames=76 |
| videos/observation.images.main/chunk-000/file-002.mp4 |
| open+create=0.0025s decode=0.2687s total=0.2712s frames=76 |
| videos/observation.images.main/chunk-000/file-003.mp4 |
| open+create=0.0023s decode=0.2687s total=0.2710s frames=76 |
| videos/observation.images.main/chunk-000/file-004.mp4 |
| open+create=0.0020s decode=0.2589s total=0.2609s frames=76 |
|
|
| Summary (mean over videos): |
| open+create decoder: 0.0022s |
| decode (get_frames_at): 0.2548s |
| total per video load: 0.2570s |
| DONE_PROFILE |
| RESULT: bridge_raw_profile PASSED |
|
|
