This is the official implementation of the paper Revisiting Classifier: Transferring Vision-Language Models for Video Recognition.
- [Dec 22, 2022] Models: The trained models & logs.
- [Nov 30, 2022] Config: All the configs (general/few-shot/zero-shot video recognition) on Kinetics-400 & 600, ActivityNet, UCF, and HMDB.
- [Nov 30, 2022] Code: Zero-shot Evaluation: Half-classes evaluation and Full-classes evaluation.
- [Nov 29, 2022] Code: Distributed training for InfoNCE and Compatible with CE.
- [Nov 28, 2022] Code: Multi-Machine Multi-GPU Distributed Training
- [Nov 28, 2022] Code: Single-Machine Multi-GPU Distributed Training, Distributed testing.
- [Nov 19, 2022] 🎉Our paper has been accepted by AAAI-2023.
- [Jul 1, 2022] 💡Our initial Arxiv paper is released.
The code is built with following libraries:
- PyTorch >= 1.8
- RandAugment
- pprint
- tqdm
- dotmap
- yaml
- csv
- Optional: decord (for on-the-fly video training)
- Optional: torchnet (for mAP evaluation on ActivityNet)
(Recommend) To train all of our models, we extract videos into frames for fast reading. Please refer to MVFNet repo for the detaied guide of data processing.
The annotation file is a text file with multiple lines, and each line indicates the directory to frames of a video, total frames of the video and the label of a video, which are split with a whitespace. Here is the format:
abseiling/-7kbO0v4hag_000107_000117 300 0
abseiling/-bwYZwnwb8E_000013_000023 300 0(Optional) We can also decode the videos in an online fashion using decord. This manner should work but are not tested. All of the models offered have been trained using offline frames. Example of annotation:
abseiling/-7kbO0v4hag_000107_000117.mp4 0
abseiling/-bwYZwnwb8E_000013_000023.mp4 0Annotation information consists of two parts: video label, and category description.
- Video Label: As mentioned above, this part is same as the traditional video recognition. Please refer to
lists/k400/kinetics_rgb_train_se320.txtfor the format. - Category Description: We also need a textual description for each video category. Please refer to
lists/kinetics_400_labels.csvfor the format.
This implementation supports Multi-GPU DistributedDataParallel training, which is faster and simpler than DataParallel used in ActionCLIP.
- Single Machine: To train our model on Kinetics-400 with 8 GPUs in Single Machine, you can run:
# For example, train the 8 Frames ViT-B/32.
sh scripts/run_train.sh configs/k400/k400_train_rgb_vitb-32-f8.yaml- Mulitple Machines: We also provide the script to train larger model with Mulitple Machines (e.g., 2 machines and 16 GPUs), you can run:
# For example, we train the 8 Frames ViT-L/14 with 2 machines as follows:
# For first machine, you need to set the ip of your first machine as the --master_addr, --nnodes is 2.
# Compared with the Single-Machine training script, only one node_id needs to be added.
sh scripts/run_train_multinodes.sh configs/k400/k400_train_rgb_vitl-14-f8.yaml 0
# For second machine, --master_addr is still the ip of your first machine
sh scripts/run_train_multinodes.sh configs/k400/k400_train_rgb_vitl-14-f8.yaml 1- Few-shot Recognition: To train our model under Few-shot scenario, you just need to add one line in the general config file:
# You can refer to config/k400/k400_few_shot.yaml
data:
... # general configurations
shot: 2 # i.e., 2-shot settingWe support single view validation and multi-view (4x3 views) validation.
# Single view evaluation. e.g., ViT-B/32 8 Frames on Kinetics-400
sh scripts/run_test.sh configs/k400/k400_train_rgb_vitb-32-f8.yaml exp/k400/ViT-B/32/f8/last_model.pt
# Multi-view evalition (4clipsx3crops). e.g., ViT-B/32 8 Frames on Kinetics-400
sh scripts/run_test.sh configs/k400/k400_train_rgb_vitb-32-f8.yaml exp/k400/ViT-B/32/f8/last_model.pt --test_crops 3 --test_clips 4We use the Kinetics-400 pre-trained model (e.g., ViT-L/14 with 8 frames) to perform cross-dataset zero-shot evaluation, i.e., UCF101, HMDB51, ActivityNet, Kinetics-600.
-
Half-classes Evaluation: A traditional evaluation protocol involves selecting half of the test dataset's classes, repeating the process ten times, and reporting the mean accuracy with a standard deviation of ten times.
-
Full-classes Evaluation: Perform evaluation on the entire dataset.
# On ActivityNet: reporting the half-classes and full-classes results
sh scripts/run_test_zeroshot.sh configs/anet/anet_zero_shot.yaml exp/k400/ViT-L/14/f8/last_model.pt
# On UCF101: reporting the half-classes and full-classes results
sh scripts/run_test_zeroshot.sh configs/ucf101/ucf_zero_shot.yaml exp/k400/ViT-L/14/f8/last_model.pt
# On HMDB51: reporting the half-classes and full-classes results
sh scripts/run_test_zeroshot.sh configs/hmdb51/hmdb_zero_shot.yaml exp/k400/ViT-L/14/f8/last_model.pt
# On Kinetics-600: manually calculating the mean accuracy with standard deviation of three splits.
sh scripts/run_test.sh configs/k600/k600_zero_shot_split1.yaml exp/k400/ViT-L/14/f8/last_model.pt
sh scripts/run_test.sh configs/k600/k600_zero_shot_split2.yaml exp/k400/ViT-L/14/f8/last_model.pt
sh scripts/run_test.sh configs/k600/k600_zero_shot_split3.yaml exp/k400/ViT-L/14/f8/last_model.ptHere we provide some off-the-shelf pre-trained checkpoints of our models in the followin tables.
#Frame = #input_frame x #spatial crops x #temporal clips
| Architecture | #Frame | Top-1 Acc.(%) | checkpoint | Train log | config |
|---|---|---|---|---|---|
| ViT-B/32 | 8x3x4 | 80.0 | Github | log | config |
| ViT-B/32 | 16x3x4 | 80.5 | Github | log | config |
| ViT-B/16 | 8x3x4 | 82.9 | Github | log | config |
| ViT-B/16 | 16x3x4 | 83.6 | Github | log | config |
| ViT-L/14* | 8x3x4 | 86.4 | GoogleDrive | log | config |
| ViT-L/14-336 | 8x3x4 | 87.1 | BaiduCloud pwd:uwl6 | log | config |
| ViT-L/14-336 | 32x3x1 | 87.8 | GoogleDrive | log | config |
Note: * indicates that this ViT-L model is used for the zero-shot evaluation on UCF, HMDB, ActivityNet and Kinetics-600.
| Architecture | #Frame | mAP (%) | checkpoint | Train log | config |
|---|---|---|---|---|---|
| ViT-L/14 | 16x1x1 | 96.5 | BaiduCloud pwd:e9m7 | log | config |
| ViT-L/14-336 | 16x1x1 | 96.9 | BaiduCloud pwd:duq7 | log | config |
| Architecture | #Frame | Top-1 Acc. (%) | checkpoint | Train log | config |
|---|---|---|---|---|---|
| ViT-L/14 | 16x1x1 | 98.1 | BaiduCloud pwd:i93p | log | config |
| Architecture | #Frame | Top-1 Acc. (%) | checkpoint | Train log | config |
|---|---|---|---|---|---|
| ViT-L/14 | 16x1x1 | 81.3 | Baidu Cloud pwd:023f | log | config |
If you find this repository useful, please star🌟 this repo and cite📑 our paper:
@article{wu2022transferring,
title={Revisiting Classifier: Transferring Vision-Language Models for Video Recognition},
author={Wu, Wenhao and Sun, Zhun and Ouyang, Wanli},
booktitle={AAAI Conference on Artificial Intelligence (AAAI)},
year={2023}
}
This repository is built based on ActionCLIP and CLIP. Sincere thanks to their wonderful works.
For any question, please file an issue or contact Wenhao Wu