We release the PyTorch code of the TDN(Temporal Difference Networks). This code is based on the TSN and TSM codebase. The core code to implement the Temporal Difference Module are ops/base_module.py
and ops/tdn_net.py
.
TL; DR. We generalize the idea of RGB difference to devise an efficient temporal difference module (TDM) for motion modeling in videos, and provide an alternative to 3D convolutions by systematically presenting principled and detailed module design.
[Dec 1, 2021] We update the TDN-ResNet101 on SSV2 in model zoo.
[Mar 5, 2021] TDN has been accepted by CVPR 2021.
[Dec 26, 2020] We have released the PyTorch code of TDN.
The code is built with following libraries:
- Python 3.6 or higher
- PyTorch 1.4 or higher
- Torchvision
- TensorboardX
- tqdm
- scikit-learn
- ffmpeg
- decord
We have successfully trained TDN on Kinetics400, UCF101, HMDB51, Something-Something-V1 and V2 with this codebase.
-
The processing of Something-Something-V1 & V2 can be summarized into 3 steps:
- Extract frames from videos(you can use ffmpeg to get frames from video)
- Generate annotations needed for dataloader ("<path_to_frames> <frames_num> <video_class>" in annotations) The annotation usually includes train.txt and val.txt. The format of *.txt file is like:
dataset_root/frames/video_1 num_frames label_1 dataset_root/frames/video_2 num_frames label_2 dataset_root/frames/video_3 num_frames label_3 ... dataset_root/frames/video_N num_frames label_N
- Add the information to
ops/dataset_configs.py
.
-
The processing of Kinetics400 can be summarized into 3 steps:
- We preprocess our data by resizing the short edge of video to 320px. You can refer to MMAction2 Data Benchmark for TSN and SlowOnly.
- Generate annotations needed for dataloader ("<path_to_video> <video_class>" in annotations) The annotation usually includes train.txt and val.txt. The format of *.txt file is like:
dataset_root/video_1.mp4 label_1 dataset_root/video_2.mp4 label_2 dataset_root/video_3.mp4 label_3 ... dataset_root/video_N.mp4 label_N
- Add the information to
ops/dataset_configs.py
.
Note: We use decord to decode the Kinetics videos on the fly.
Here we provide some off-the-shelf pretrained models. The accuracy might vary a little bit compared to the paper, since the raw video of Kinetics downloaded by users may have some differences.
Model | Frames x Crops x Clips | Top-1 | Top-5 | checkpoint |
---|---|---|---|---|
TDN-ResNet50 | 8x1x1 | 52.3% | 80.6% | link |
TDN-ResNet50 | 16x1x1 | 53.9% | 82.1% | link |
Model | Frames x Crops x Clips | Top-1 | Top-5 | checkpoint |
---|---|---|---|---|
TDN-ResNet50 | 8x1x1 | 64.0% | 88.8% | link |
TDN-ResNet50 | 16x1x1 | 65.3% | 89.7% | link |
TDN-ResNet101 | 8x1x1 | 65.8% | 90.2% | link |
8x3x1 | 67.1% | 90.5% | - | |
TDN-ResNet101 | 16x1x1 | 66.9% | 90.9% | link |
8x3x1 | 68.2% | 91.6% | - | |
TDN-ResNet101 | (8+16)x1x1 | 68.2% | 91.6% | - |
(8+16)x3x1 | 69.6% | 92.2% | - |
Model | Frames x Crops x Clips | Top-1 (30 view) | Top-5 (30 view) | checkpoint |
---|---|---|---|---|
TDN-ResNet50 | 8x3x10 | 76.6% | 92.8% | link |
TDN-ResNet50 | 16x3x10 | 77.5% | 93.2% | link |
TDN-ResNet101 | 8x3x10 | 77.5% | 93.6% | link |
TDN-ResNet101 | 16x3x10 | 78.5% | 93.9% | link |
- For center crop single clip, the processing of testing can be summarized into 2 steps:
- Run the following testing scripts:
CUDA_VISIBLE_DEVICES=0 python3 test_models_center_crop.py something \ --archs='resnet50' --weights <your_checkpoint_path> --test_segments=8 \ --test_crops=1 --batch_size=16 --gpus 0 --output_dir <your_pkl_path> -j 4 --clip_index=0
- Run the following scripts to get result from the raw score:
python3 pkl_to_results.py --num_clips 1 --test_crops 1 --output_dir <your_pkl_path>
- Run the following testing scripts:
- For 3 crops, 10 clips, the processing of testing can be summarized into 2 steps:
- Run the following testing scripts for 10 times(clip_index from 0 to 9):
CUDA_VISIBLE_DEVICES=0 python3 test_models_three_crops.py kinetics \ --archs='resnet50' --weights <your_checkpoint_path> --test_segments=8 \ --test_crops=3 --batch_size=16 --full_res --gpus 0 --output_dir <your_pkl_path> \ -j 4 --clip_index <your_clip_index>
- Run the following scripts to ensemble the raw score of the 30 views:
python pkl_to_results.py --num_clips 10 --test_crops 3 --output_dir <your_pkl_path>
- Run the following testing scripts for 10 times(clip_index from 0 to 9):
This implementation supports multi-gpu, DistributedDataParallel
training, which is faster and simpler.
- For example, to train TDN-ResNet50 on Something-Something-V1 with 8 gpus, you can run:
python -m torch.distributed.launch --master_port 12347 --nproc_per_node=8 \ main.py something RGB --arch resnet50 --num_segments 8 --gd 20 --lr 0.01 \ --lr_scheduler step --lr_steps 30 45 55 --epochs 60 --batch-size 8 \ --wd 5e-4 --dropout 0.5 --consensus_type=avg --eval-freq=1 -j 4 --npb
- For example, to train TDN-ResNet50 on Kinetics400 with 8 gpus, you can run:
python -m torch.distributed.launch --master_port 12347 --nproc_per_node=8 \ main.py kinetics RGB --arch resnet50 --num_segments 8 --gd 20 --lr 0.02 \ --lr_scheduler step --lr_steps 50 75 90 --epochs 100 --batch-size 16 \ --wd 1e-4 --dropout 0.5 --consensus_type=avg --eval-freq=1 -j 4 --npb
We especially thank the contributors of the TSN and TSM codebase for providing helpful code.
This repository is released under the Apache-2.0. license as found in the LICENSE file.
If you think our work is useful, please feel free to cite our paper ๐ :
@InProceedings{Wang_2021_CVPR,
author = {Wang, Limin and Tong, Zhan and Ji, Bin and Wu, Gangshan},
title = {TDN: Temporal Difference Networks for Efficient Action Recognition},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021},
pages = {1895-1904}
}