1.Requirements
Environment
- Python 3.6.5
- PyTorch 0.4.1
- TorchVison 0.2.1
- FFmpeg, FFprobe
bash conda create -n pt0.4.1 python=3.6.5 #创建一个新环境
pip install --upgrade pip #更新pip
conda install numpy=1.15.4 pillow=5.3.0 scipy=1.1.0 scikit-learn=0.20.1 PyYAML=3.13 matplotlib=3.0.1 ffmpeg=4.0 #安装所需要的包
pip install torch-0.4.1-cp36-cp36m-linux_x86_64.whl #安装pytorch
pip install torchvision==0.2.1 #安装torchvision
pip install tensorboardX #安装tensorboardX
2.Data Preparation
Basically, the processing of video data can be summarized into 3 steps:
- Get the videos and Decompress them into a folder
- Convert from videos to frames
- Construct file lists for training and validation
Get the videos and Decompress them into a folder
###Kinetics Because kinetics datasets are not original, so we first make label file in the same form as UCF-101:
python video_txt.py <videodir_path> <txtfile_path>
###UCF-101
the ucf101 videos are archived in the downloaded file. Please use unrar x UCF101.rar to extract the videos.
###HMDB-51
the HMDB51 video archive has two-level of packaging. The following commands illustrate how to extract the videos.
mkdir rars && mkdir videos
unrar x hmdb51_org.rar rars/
for a in $(ls rars); do unrar x "rars/${a}" videos/; done;
###Something-Somethingv1
the something-something-v1 frames are archived in the downloaded file. Please use cat 20bn-something-something-v1-?? | tar zx to decompress them into a folder.
###Something-Somethingv2
the something-something-v2 videos are archived in the downloaded file. Please use cat 20bn-something-something-v2-?? | tar zx to decompress them into a folder.
and you should use python video_frame.py <videodir_path> <framedir_path> to convert it from videos to frames.
###Jester
the jester-v1 frames are archived in the downloaded file. Please use cat 20bn-jester-v1-?? | tar zx to decompress them into a folder.
Convert from videos to frames
sh scripts/extract_frames.sh <videodir_path> <framedir_path>
or
python utils/video_frame.py <videodir_path> <framedir_path>
Construct file lists for training and validation
sh scripts/gen_dataset_list.sh <dataset_name>
3.Running the code
Training
sh scripts/train_somethingv1.sh
Testing
sh scripts/test_somethingv1.sh