Several methods was integrated into a single framework.
- SeqPAN:
- SeqPANBackBone:
- remvoe the match loss and dual attention
- SeqPANBert:
- replace the BERT word embedding to Glove.
- VSLNet
- BAN
- https://github.com/DJX1995/BAN-APR.git
- (still exsit bug, can run, but can't train )
- CCA
- CPL
- https://github.com/minghangz/cpl.git
- (uncompleted)
NOTE
There are many draft codes, we don't test the robustness. Especially the part of generating various labels in BaseDataset.py That easily produce bugs to decrease performance. Please test the consistency of label and vfeat length before traning.
Download the video feature from https://huggingface.co/datasets/k-nick/NLVL. We collect several version feature for Charades, ActivityNet, and Tacos. Then modify "paths" value within the ./config/anet/SeqPAN.yaml.
# train
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/charades/SeqPAN.yaml
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/anet/VSLNet.yaml
# debug
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/charades/SeqPAN.yaml --debug
# test accuracy
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/charades/SeqPAN.yaml --mode test --checkpoint ./ckpt/charades/best_SeqPAN.pkl
# test efficiency
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/charades/SeqPAN.yaml --mode summary