Wuziyi616
commented
6 months ago Hi, this is what I got when running your command:
python vis_pred.py model=rnndet dataset=gen1 dataset.path=./datasets/gen1/ checkpoint="pretrained/Sec.4.2-WSOD_SSOD/gen1-WSOD/rvt-s-gen1x0.01_ss-final.ckpt" +experiment/gen1="small.yaml" model.postprocess.confidence_threshold=0.1 num_video=5 reverse=False
Using python-based detection evaluation
Set MaxViTRNN backbone (height, width) to (256, 320)
Set partition sizes: (8, 10)
Set num_classes=2 for detection head
Found 470 sequences in /datasets/automotive-megapixel/gen1/test
Using 1.25% of data --> 6 sequences
creating streaming test datasets: 100%|█████████████████████████████████████| 6/6 [00:00<00:00, 65.62it/s]
num_full_sequences=6
num_splits=0
num_split_sequences=0
Event sequence: 0it [00:00, ?it/s]
[2024-04-02 17:29:44,920][py.warnings][WARNING] - /fs01/home/ziyiwu/conda/event/lib/python3.9/site-packages/torch/functional.py:504: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /opt/conda/conda-bld/pytorch_1682343997789/work/aten/src/ATen/native/TensorShape.cpp:3483.)
return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
Predict: 100%|█████████████████████████████████████████████████████████| 640/640 [00:06<00:00, 104.51it/s]
Plot: 640it [00:02, 315.64it/s]████████████████████████████████████████| 640/640 [00:06<00:00, 121.03it/s]I notice that the number of sequence detected in the test set is different -- your logs say 46 sequences detected in datasets/gen1/test, while mine say 470 sequences (and it's the correct number in the paper). As a result, the dataset only loads one sequence for testing (set here to reduce RAM consumption), and that's actually skipped (see these lines) because I noticed that test sequence 1 on Gen1 has very few motion, and thus there is almost no events which makes it impossible to detect anything.
To fix this, you can comment out this line, and the dataset should load all 46 sequences, test on the first 5 of them, and save them.
Tushar-Isamaliya
Problem: The video .mp4 is not getting generated in given folder (/vis/gen1_rnndet_small/pred/) but code runs successfully.
after successful run of the visualisation code :
python vis_pred.py model=rnndet dataset=gen1 dataset.path=./datasets/gen1/ checkpoint="pretrained/Sec.4.2-WSOD_SSOD/gen1-WSOD/rvt-s-gen1x0.01_ss-final.ckpt" +experiment/gen1="small.yaml" model.postprocess.confidence_threshold=0.1 num_video=5 reverse=False
Using python-based detection evaluation Set MaxViTRNN backbone (height, width) to (256, 320) Set partition sizes: (8, 10) Set num_classes=2 for detection head Found 46 sequences in datasets/gen1/test Using 1.25% of data --> 1 sequences creating streaming test datasets: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 418.26it/s] num_full_sequences=1 num_splits=0 num_split_sequences=0 Event sequence: 0it [00:00, ?it/s] [2024-04-02 20:47:06,734][py.warnings][WARNING] - /home/tushar/miniforge3/envs/leod/lib/python3.9/site-packages/torch/functional.py:504: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at /opt/conda/conda-bld/pytorch_1678402412426/work/aten/src/ATen/native/TensorShape.cpp:3483.) return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
Predict: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 6.55it/s] Plot: 2it [00:00, 258.37it/s]████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 6.38it/s] Event sequence: 1it [00:01, 1.73s/it] == Timing statistics ==