Implementation of paper Temporal Sentence Grounding in Streaming Videos (Accepted by ACM Multimedia 2023).
Please refer to install.md for installation.
Please refer to data_preparation.md for a general knowledge of data preparation.
- ActivityNet Captions
# If you have 8 gpus
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
./tools/dist_train.sh configs/tpn/benchmark/anet_mtl_16_16_tpn_dec2_rnn_dot_s8_l64_b8*64_kd04.py 8 --validate --test-best
# If you have fewer gpus (like 2), you have three options:
# a) Increase batch size
export CUDA_VISIBLE_DEVICES="0,1"
./tools/dist_train.sh configs/tpn/benchmark/anet_mtl_16_16_tpn_dec2_rnn_dot_s8_l64_b8*64_kd04.py 2 --validate --test-best --cfg-options data.videos_per_gpu=256
# b) Decrease learning rate if GPU memory not enough for a)
export CUDA_VISIBLE_DEVICES="0,1"
./tools/dist_train.sh configs/tpn/benchmark/anet_mtl_16_16_tpn_dec2_rnn_dot_s8_l64_b8*64_kd04.py 2 --validate --test-best --cfg-options optimizer.lr=2.5e-6
# c) combine the a) and b)- TACoS
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
./tools/dist_train.sh configs/tpn/benchmark/tacos_mtl_16_tpn_dec1_rnn_dot_s8_l64_b8*64_kd02.py 8 --validate --test-best- MAD
# Train
export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
./tools/dist_train.sh configs/tpn/benchmark/mad_mtl_16_16_tpn_dec2_rnn_dot_s8_l32_b8*128_kd02.py 8 --validate
# Test
export CUDA_VISIBLE_DEVICES="0,1"
for i in {0..19} # split the dataset into 20 chunks to avoid OOM
do
./tools/dist_test.sh work_dirs/mad_mtl_16_16_tpn_dec2_rnn_dot_s8_l32_b8*128_kd02/mad_mtl_16_16_tpn_dec2_rnn_dot_s8_l32_b8*128_kd02.py \
work_dirs/mad_mtl_16_16_tpn_dec2_rnn_dot_s8_l32_b8*128_kd02/best_R@1,IoU=0.3_epoch_13.pth 2 --eval R@N,IoU=M \
--out work_dirs/mad_mtl_16_16_tpn_dec2_rnn_dot_s8_l32_b8*128_kd02/best_pred_${i}_20.pkl \
--cfg-options data.workers_per_gpu=10 data.test.portion=[$i,20]
done@inproceedings{tian2017tsgvs,
title={{Temporal} {Sentence} in {Streaming} {Videos}},
author={Tian Gan, Xiao Wang, Yan Sun, Jianlong Wu, Qingpei Guo, Liqiang Nie},
booktitle={{International} {Conference} on {Multimedia}},
publisher= {{ACM}},
year={2023}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent