TSM (ICCV'19) action recognition Tutorial for BYRers

---

1. 环境配置

git clone https://github.com/kravrolens/TSMaction.git
conda create -n action python=3.9
conda activate action
conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia -y
pip install -r requirements.txt

2. 下载预训练权重

bash pretrained/download.sh

3. Kinetics-400数据集准备

首先批量下载训练集和验证集压缩包：

mkdir p ~/Data/Kinetics-400 && cd ~/Data/Kinetics-400
bash ~/Code/TSM/data/download.sh ~/Code/TSM/data/train_link.list
bash ~/Code/TSM/data/download.sh ~/Code/TSM/data/val_link.list

解压数据集，路径基本格式为：

$~/Data/Kinetics-400/train/     
$~/Data/Kinetics-400/test/                      
$~/Data/Kinetics-400/train/abseiling/
...
$~/Data/Kinetics-400/test/abseiling/
...

把视频拆分成帧，由于磁盘容量限制，这里仅拆分验证集，路径是：~/Data/Kinetics-400/val_frame

mkdir -p ~/Data/Kinetics-400/val_frame && cd tools
python vid2img_kinetics.py ~/Data/Kinetics-400/val ~/Data/Kinetics-400/val_frame

生成标签，路径在：~/Code/TSM/data/val_videofolder.txt，生成格式是：<图像相对路径 jepg数量 类别>

python tools/gen_label_kinetics_val.py

若要生成训练集，只需将上述路径中的val替换为train即可，路径是：~/Data/Kinetics-400/train_frame

mkdir -p ~/Data/Kinetics-400/train_frame && cd tools
python vid2img_kinetics.py ~/Data/Kinetics-400/train ~/Data/Kinetics-400/train_frame
python tools/gen_label_kinetics_train.py

生成标签，路径在：~/Code/TSM/data/train_videofolder.txt

4. 训练

bash scripts/train_tsm_kinetics_rgb_8f.sh

5. 测试

bash scripts/test_tsm_kinetics_rgb_8f.sh

6. 参考

• TSM: Temporal Shift Module for Efficient Video Understanding
• official repo
• official Kinetics准备
• PaddlePaddle Kinetics准备
• CSDN blog Kinetics准备