Extracting Features with TSN for BMN Model on Custom Dataset: Clarifications Needed

[YancyGuo]

Firstly, I am in the process of building a dataset for BMN, but I am confused about some of the steps presented in the Doc.

I have trained a checkpoint for TSN and modified the configuration based on my dataset's specifics. My challenge is understanding how to construct a CSV similar to the standard feature. According to the Doc. eeach standard CSV should be composed of 400 features (representing 400 classes) and have 100 lines. To achieve this CSV format, I need to extract the relevant features and save them as a .pkl file.
However, when I follow Step 6 and extract features using the command below, the resulting .pkl file shows dimensions of pytorch.Size([144, 2048]):
python ../../misc/clip_feature_extraction.py /data/yancy/envs/mmaction2v3/mmaction2/tools/data/activitynet/tsn_extract_rgb_feat_config_modified.py /data/yancy/envs/mmaction2v3/mmaction2/work_dirs/113_test/best_acc_top1_epoch_76.pth ../../../data/US/rgb_val_feat --video-list /data/yancy/envs/mmaction2v2/mmaction2-main/data/data_v3/val_file_list.txt --video-root /data/yancy/envs/mmaction2v2/mmaction2-main/data/data_v3/frames_notime/val --spatial-type avg --temporal-type keep --local_rank 1

I believe the first dimension (144) is correct, but the 2048 dimension seems to be an output without the classification layer. Therefore, I adjusted the command to get the score of each frame, similar to the ActivityNet CSV, which I expect to display as (144,400). However, the modified command outputs a dimension of pytorch.Size([400])
python ../../misc/clip_feature_extraction.py /data/yancy/envs/mmaction2v3/mmaction2/tools/data/activitynet/tsn_extract_rgb_feat_config_modified.py /data/yancy/envs/mmaction2v3/mmaction2/work_dirs/113_test/best_acc_top1_epoch_76.pth ../../../data/US/rgb_val_feat --video-list /data/yancy/envs/mmaction2v2/mmaction2-main/data/data_v3/val_file_list.txt --video-root /data/yancy/envs/mmaction2v2/mmaction2-main/data/data_v3/frames_notime/val --spatial-type avg --temporal-type keep --dump-score --local_rank 1