TengdaHan/DPC

Video Representation Learning by Dense Predictive Coding

This repository contains the implementation of Dense Predictive Coding (DPC).

arch

DPC Results

Original result from our paper:

Pretrain Dataset	Resolution	Backbone	Finetune Acc@1 (UCF101)	Finetune Acc@1 (HMDB51)
UCF101	128x128	2d3d-R18	60.6	-
Kinetics400	128x128	2d3d-R18	68.2	34.5
Kinetics400	224x224	2d3d-R34	75.7	35.7

Also re-implemented by other researchers:	Pretrain Dataset	Resolution	Backbone	Finetune Acc@1 (UCF101)	Finetune Acc@1 (HMDB51)
UCF101	128x128	2d3d-R18	61.35 @kayush95	45.31 @kayush95

News

2020/06/05: Update the link for 3D-ResNet34-Kinetics400-224x224-runningStats, the old one didn't save BN running statistics, thus couldn't be used to evaluate linear probe. Now it saves (without changing any weights).

Installation

The implementation should work with python >= 3.6, pytorch >= 0.4, torchvision >= 0.2.2.

The repo also requires cv2 (conda install -c menpo opencv), tensorboardX >= 1.7 (pip install tensorboardX), joblib, tqdm, ipdb.

Prepare data

Follow the instructions here.

Self-supervised training (DPC)

Change directory cd DPC/dpc/

example: train DPC-RNN using 2 GPUs, with 3D-ResNet18 backbone, on Kinetics400 dataset with 128x128 resolution, for 300 epochs
```
python main.py --gpu 0,1 --net resnet18 --dataset k400 --batch_size 128 --img_dim 128 --epochs 300
```
example: train DPC-RNN using 4 GPUs, with 3D-ResNet34 backbone, on Kinetics400 dataset with 224x224 resolution, for 150 epochs
```
python main.py --gpu 0,1,2,3 --net resnet34 --dataset k400 --batch_size 44 --img_dim 224 --epochs 150
```

Evaluation: supervised action classification

Change directory cd DPC/eval/

example: finetune pretrained DPC weights (replace {model.pth.tar} with pretrained DPC model)

python test.py --gpu 0,1 --net resnet18 --dataset ucf101 --batch_size 128 --img_dim 128 --pretrain {model.pth.tar} --train_what ft --epochs 300

example (continued): test the finetuned model (replace {finetune_model.pth.tar} with finetuned classifier model)

python test.py --gpu 0,1 --net resnet18 --dataset ucf101 --batch_size 128 --img_dim 128 --test {finetune_model.pth.tar}

DPC-pretrained weights

It took us more than 1 week to train the 3D-ResNet18 DPC model on Kinetics-400 with 128x128 resolution, and it tooks about 6 weeks to train the 3D-ResNet34 DPC model on Kinetics-400 with 224x224 resolution (with 4 Nvidia P40 GPUs).

Download link:

3D-ResNet18-Kinetics400-128x128,
3D-ResNet34-Kinetics400-224x224,
3D-ResNet34-Kinetics400-224x224-runningStats

example: finetune 3D-ResNet34-Kinetics400-224x224

python test.py --gpu 0,1 --net resnet34 --dataset ucf101 --batch_size 44 --img_dim 224 --pretrain {model.pth.tar} --train_what ft --epochs 300

Citation

If you find the repo useful for your research, please consider citing our paper:

@InProceedings{Han19dpc,
  author       = "Tengda Han and Weidi Xie and Andrew Zisserman",
  title        = "Video Representation Learning by Dense Predictive Coding",
  booktitle    = "Workshop on Large Scale Holistic Video Understanding, ICCV",
  year         = "2019",
}

For any questions, welcome to create an issue or contact Tengda Han (htd@robots.ox.ac.uk).