CenterCLIP achieves state-of-the-art text-video retrieval performance and decent computation cost reduction on MSVD, MSRVTT, LSMDC, and ActivityNet through performing multi-segment token clustering on video tokens in the vision transformer of CLIP.
- [02/05/2022] create repo.
This is the code for the paper
CenterCLIP: Token Clustering for Efficient Text-Video Retrieval.
In this work, to reduce the number of redundant video tokens,
we design a multi-segment token clustering algorithm
to find the most representative tokens and drop the non-essential ones.
As the frame redundancy occurs mostly in consecutive frames, we divide videos into multiple segments and conduct segment-level clustering.
Center tokens from each segment are later concatenated into a new sequence, while their original spatial-temporal relations are well maintained.
We instantiate two clustering algorithms to efficiently find deterministic medoids and iteratively partition groups in high dimensional space.
Through this token clustering and center selection procedure,
we successfully reduce computation costs by removing redundant visual tokens.
This method further enhances segment-level semantic alignment between video and text representations, enforcing the spatio-temporal interactions of tokens from within-segment frames.
Our method, coined as CenterCLIP, surpasses existing state-of-the-art
by a large margin on typical text-video benchmarks, while reducing the training memory cost by 35%
and accelerating the inference speed by 14% at the best case.
- Different datasets, i.e., MSR-VTT, MSVD, DiDeMo, ActivityNet, LSMDC
- Automated mixed precision training + Distributed training (tested with multi-GPUs on multi-nodes)
- Fast PyAv video decoding + sparse frame sampling
- Fast clustering algorithms supporting batch operations
- LMDB database to accelerate IO
We are open to pull requests.
Experiments on MSVD need at least 2 RTX 3090 GPUs.
Experiments on ActivityNet need at least 8 Tesla V100 32GB GPUs.
- Install dependencies via docker
Please install PyTorch-1.9.0 and Python3.6+. PyTorch-1.6.0+ should work.
We recommend you to use our established PyTorch docker image: zhaosssss/torch_lab:1.9.3.
docker pull zhaosssss/torch_lab:1.9.3
If you have not installed docker, see https://docs.docker.com/.
After you install docker and pull our image, you can cd to script directory and run
./run_docker.sh
to create a running docker container.
NOTE: We map some directories in run_docker.sh, if you do not have these directories,
you need to modify the script.
By default, run_docker.sh runs container in background
and you need run docker exec -it ${DOCKER-ID} bash
to do some interactive operations.
- Install dependencies via
pip
If you do not want to use docker, try
pip install -r requirements.txt
However, this is not suggested.
Generally, directories are organized as following:
${HOME}
├── dataset (save the dataset)
│ │
│ ├── activitynet
│ ├── lsmdc
│ └── msrvtt
│
├── models
│ │
│ ├── eclip (save the output checkpoints)
│ └── pretrained (save the CLIP pre-trained weights)
│
├── github (save the code)
│ │
│ └── centerclip
│ │
│ ├── dataloaders
│ ├── modules
│ ├── scripts
│ └── preprocess
...
-
Some dataset splits can be found in misc/splits.
-
Video preprocessing can be done by preprocess/compress_video.py. By default we use 3 fps and 224 shorter side of frames.
-
Download CLIP pre-trained weights and place them in
${HOME}/models/pretrained.
- You can make a LMBD video database via preprocess/folder2lmdb.py if you think your IO is a bottleneck.
Download the splits and captions from CLIP4clip:
wget https://github.com/ArrowLuo/CLIP4Clip/releases/download/v0.0/msrvtt_data.zip
Download the videos from Frozen️-in-Time:
wget https://www.robots.ox.ac.uk/~maxbain/frozen-in-time/data/MSRVTT.zip
Download videos from https://www.cs.utexas.edu/users/ml/clamp/videoDescription/.
Splits can be found in https://github.com/albanie/collaborative-experts/tree/master/misc/datasets/msvd.
Or you can download them from CLIP4clip
wget https://github.com/ArrowLuo/CLIP4Clip/releases/download/v0.0/msvd_data.zip
You must obtain permission from MPII to download and use the data https://sites.google.com/site/describingmovies/download.
The videos are large than 2T, you can use preprocess/download_lsmdc.py to achieve online downloading and resizing.
It is also a multi-processes LSMDC downloader.
Set only_down=True for only downloading without resizing.
Download from http://activity-net.org/download.html. Splits can be found in https://github.com/albanie/collaborative-experts/tree/master/misc/datasets/activity-net or in misc/splits/activitynet.
For the meaning of hyper-parameters, run
python params.py --help
Or see the comments in modules/cluster/cluster.py.
See
scripts/lsmdc.sh
I add some experiments in the file, you can choose and run them.
Be careful about the batch_size and your gpu numbers.
Generally, batch_size x #GPUs = 128 as I use 128 as the total batch size.
batch_size in the scripts means single gpu batch size.
scripts/msvd.sh
scripts/msrvtt.sh
scripts/activitynet.sh
tensorboard --logdir=your_logdir --port=your_port
# or run scripts/tensorboard.sh
Checkpoints trained on Tesla V100 GPUs are not available now. We provide some checkpoints trained on 2 RTX 3090 GPUs for you to play around with. Results of checkpoints on LSMDC are the same as the paper's data. Checkpoints on MSR-VTT and MSVD come from middle stages of our work. They have comparable performance with the paper's results (CenterCLIP, ViT-B/32).
Third-party reproduction and checkpoints are warmly welcomed.
Each zip file contains 4 types of files
- a checkpoint of the model, typically, named as
ckpt.best.pth.tar - log file, named as
log.txt - a hyper-parameter json file, typically, named as
hparams_train.json tensorboardlog file, you can usetensorboardto visualize the log. It is in thetensorboarddirectory within the zip file.
| Checkpoint ID | Dataset | T2V R@1 | V2T R@1 | URL |
|---|---|---|---|---|
| eclip_new_abla_lsmdc_04 | lsmdc | 21.9 | 21.1 | zip file |
| eclip_new_abla_lsmdc_09 | lsmdc | 21.7 | 21.4 | zip file |
| eclip_new_abla_lsmdc_22 | lsmdc | 21.6 | 20.6 | zip file |
| eclip_new_abla_lsmdc_23 | lsmdc | 21.4 | 19.5 | zip file |
| eclip_msrvtt_62 | msrvtt (7k) / 1k-A | 44.1 | 41.9 | zip file |
| eclip_msrvtt_63 | msrvtt (7k) / 1k-A | 44.2 | 43.2 | zip file |
| eclip_msrvtt_80 | msrvtt (7k) / 1k-A | 43.9 | 42.6 | zip file |
| eclip_msvd_22 | msvd | 47.5 | 61.4 | zip file |
Set
# train or eval
do_train=0
do_eval=1
in the training scripts to get the evaluation results of these checkpoints.
Corresponding settings are ready in the bash scripts.
@inproceedings{2022_centerclip,
author = {Shuai Zhao and Linchao Zhu and Xiaohan Wang and Yi Yang},
title = {CenterCLIP: Token Clustering for Efficient Text-Video Retrieval},
booktitle = {{SIGIR} '22: The 45th International {ACM} {SIGIR} Conference on Research
and Development in Information Retrieval, July 11–15, 2022, Madrid, Spain},
year = {2022},
}
This project is under the CC-BY-NC 4.0 license. See LICENSE for details..
- pytorch/vision
- ArrowLuo/CLIP4Clip
- m-bain/frozen-in-time
- albanie/collaborative-experts
- openai/CLIP
- mlfoundations/open_clip
- huggingface/transformers
- facebookresearch/pytorchvideo
- DeMoriarty/fast_pytorch_kmeans
- subhadarship/kmeans_pytorch
- PyAV-Org/PyAV
- sallymmx/ActionCLIP
- VideoNetworks/TokShift-Transformer
- yjxiong/tsn-pytorch
- mit-han-lab/temporal-shift-module
- mzhaoshuai/Divide-and-Co-training
- ZJULearning/RMI