We published a new paper on arXiv.
We also added the following new models and their Kinetics pretrained models in this repository.
- ResNet-50, 101, 152, 200
- Pre-activation ResNet-200
- Wide ResNet-50
- ResNeXt-101
- DenseNet-121, 201 In addition, we supported new datasets (UCF-101 and HDMB-51) and fine-tuning functions.
Some minor changes are included.
- Outputs are normalized by softmax in test.
- If you do not want to perform the normalization, please use
--no_softmax_in_test
option.
- If you do not want to perform the normalization, please use
This is the PyTorch code for the following papers:
This code includes training, fine-tuning and testing on Kinetics, ActivityNet, UCF-101, and HMDB-51.
If you want to classify your videos or extract video features of them using our pretrained models,
use this code.
The Torch (Lua) version of this code is available here.
Note that the Torch version only includes ResNet-18, 34, 50, 101, and 152.
If you use this code or pre-trained models, please cite the following:
@article{hara3dcnns,
author={Kensho Hara and Hirokatsu Kataoka and Yutaka Satoh},
title={Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?},
journal={arXiv preprint},
volume={arXiv:1711.09577},
year={2017},
}
Pre-trained models are available here.
All models are trained on Kinetics.
ResNeXt-101 achieved the best performance in our experiments. (See paper in details.)
resnet-18-kinetics.pth: --model resnet --model_depth 18 --resnet_shortcut A
resnet-34-kinetics.pth: --model resnet --model_depth 34 --resnet_shortcut A
resnet-34-kinetics-cpu.pth: CPU ver. of resnet-34-kinetics.pth
resnet-50-kinetics.pth: --model resnet --model_depth 50 --resnet_shortcut B
resnet-101-kinetics.pth: --model resnet --model_depth 101 --resnet_shortcut B
resnet-152-kinetics.pth: --model resnet --model_depth 152 --resnet_shortcut B
resnet-200-kinetics.pth: --model resnet --model_depth 200 --resnet_shortcut B
preresnet-200-kinetics.pth: --model preresnet --model_depth 200 --resnet_shortcut B
wideresnet-50-kinetics.pth: --model wideresnet --model_depth 50 --resnet_shortcut B --wide_resnet_k 2
resnext-101-kinetics.pth: --model resnext --model_depth 101 --resnet_shortcut B --resnext_cardinality 32
densenet-121-kinetics.pth: --model densenet --model_depth 121
densenet-201-kinetics.pth: --model densenet --model_depth 201
Fine-tuned models on UCF-101 or HMDB-51 will be also available.
conda install pytorch torchvision cuda80 -c soumith
- FFmpeg, FFprobe
wget http://johnvansickle.com/ffmpeg/releases/ffmpeg-release-64bit-static.tar.xz
tar xvf ffmpeg-release-64bit-static.tar.xz
cd ./ffmpeg-3.3.3-64bit-static/; sudo cp ffmpeg ffprobe /usr/local/bin;
- Python 3
- Download datasets using official crawler codes
- Convert from avi to jpg files using
utils/video_jpg.py
python utils/video_jpg.py avi_video_directory jpg_video_directory
- Generate fps files using
utils/fps.py
python utils/fps.py avi_video_directory jpg_video_directory
- Download the Kinetics dataset using official crawler codes
- Locate test set in
video_directory/test
.
- Locate test set in
- Convert from avi to jpg files using
utils/video_jpg_kinetics.py
python utils/video_jpg_kinetics.py avi_video_directory jpg_video_directory
- Generate n_frames files using
utils/n_frames_kinetics.py
python utils/n_frames_kinetics.py jpg_video_directory
- Generate annotation file in json format similar to ActivityNet using
utils/kinetics_json.py
python utils/kinetics_json.py train_csv_path val_csv_path test_csv_path json_path
Assume the structure of data directories is the following:
~/
data/
kinetics_videos/
jpg/
.../ (directories of class names)
.../ (directories of video names)
... (jpg files)
results/
save_100.pth
kinetics.json
Confirm all options.
python main.lua -h
Train ResNets-34 on the Kinetics dataset (400 classes) with 4 CPU threads (for data loading).
Batch size is 128.
Save models at every 5 epochs.
All GPUs is used for the training.
If you want a part of GPUs, use CUDA_VISIBLE_DEVICES=...
.
python main.py --root_path ~/data --video_path kinetics_videos/jpg --annotation_path kinetics.json \
--result_path results --dataset kinetics --model resnet \
--model_depth 34 --n_classes 400 --batch_size 128 --n_threads 4 --checkpoint 5
Continue Training from epoch 101. (~/data/results/save_100.pth is loaded.)
python main.py --root_path ~/data --video_path kinetics_videos/jpg --annotation_path kinetics.json \
--result_path results --dataset kinetics --resume_path results/save_100.pth \
--model_depth 34 --n_classes 400 --batch_size 128 --n_threads 4 --checkpoint 5
Fine-tuning conv5_x and fc layers of a pretrained model (~/data/models/resnet-34-kinetics.pth) on UCF-101.
python main.py --root_path ~/data --video_path ucf101_videos/jpg --annotation_path ucf101_01.json \
--result_path results --dataset ucf101 --n_classes 400 --n_finetune_classes 101 \
--pretrain_path models/resnet-34-kinetics.pth --ft_begin_index 4 \
--model resnet --model_depth 34 --resnet_shortcut A --batch_size 128 --n_threads 4 --checkpoint 5