Decoupled Dynamic Filter Networks

This repo is the official implementation of CVPR2021 paper: "Decoupled Dynamic Filter Networks".

Introduction

DDF is an alternative of convolution which decouples dynamic filters into spatial and channel filters.

We illustrate the DDF operation and the DDF module. The orange color denotes spatial dynamic filters / branch, and the green color denotes channel dynamic filters / branch. The filter application means applying the convolution operation at a single position. ‘GAP’ means the global average pooling and ‘FC’ denotes the fully connected layer.

Please refer to our project page and paper for more details.

Model zoo

Model	#Params	Pretrain	Resolution	Top1 Acc	Download
ddf_mul_resnet50	16.8M	ImageNet 1K	224	79.1	google / baidu
ddf_mul_resnet101	28.1M	ImageNet 1K	224	80.5	google / baidu
ddf_add_resnet50	16.8M	ImageNet 1K	224	78.8	google / baidu
ddf_add_resnet101	28.1M	ImageNet 1K	224	79.9	baidu

The code for baidu is vub3

Use ddf as a basic building layer

Please directly copy the ddf folder to your repo and build the ddf operation. Then, you can easily import the ddf operation, the DDFPack, and the DDFUpPack.

You can design your own module with the ddf operation.

For example, you can get a carafe/involution-like module by fixing all values in the channel filter to 1 for 'mul' combination or 0 for 'add' combination.

channel_filter = torch.ones(filter_size)
output = ddf(input, channel_filter, spatial_filter,
             kernel_size, dilation, stride, 'mul')

channel_filter = torch.zeros(filter_size)
output = ddf(input, channel_filter, spatial_filter,
             kernel_size, dilation, stride, 'add')

Similarly, you can get a WeightNet-like depthwise filter by fixing all values in the spatial filter to 1 for 'mul' combination or 0 for 'add' combination.

spatial_filter = torch.ones(filter_size)
output = ddf(input, channel_filter, spatial_filter,
             kernel_size, dilation, stride, 'mul')

spatial_filter = torch.zeros(filter_size)
output = ddf(input, channel_filter, spatial_filter,
             kernel_size, dilation, stride, 'add')

Almost ALL exisitng weight-dynamic depthwise operation (not grid-dynamic like deformable convolution) can be implemented by our ddf operation. Have fun exploring.

Install

Clone this repo:

git clone https://github.com/theFoxofSky/ddfnet.git
cd ddfnet

Create a conda virtual environment and activate it:

conda create -n ddfnet python=3.7 -y
conda activate ddfnet

Install CUDA==10.1 with cudnn7 following the official installation instructions
Install PyTorch==1.7.1 and torchvision==0.8.2 with CUDA==10.1:

conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.1 -c pytorch

Install timm==0.4.5:

pip install timm==0.4.5

Install Apex:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Install other requirements:

pip install pyyaml ipdb

Build the ddf operation:

cd ddf
python setup.py install
mv build/lib*/* .

Verify the ddf operation:

cd <path_to_ddfnet>
python grad_check.py

Data preparation

We use standard ImageNet dataset, you can download it from http://image-net.org/. Please prepare it under the following file structure:

$ tree data
imagenet
├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   ├── img2.jpeg
│   │   └── ...
│   ├── class2
│   │   ├── img3.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class1
    │   ├── img4.jpeg
    │   ├── img5.jpeg
    │   └── ...
    ├── class2
    │   ├── img6.jpeg
    │   └── ...
    └── ...

Training from scratch

To train a model, for example ddf_mul_resnet50, on ImageNet from scratch with 8 RTX 2080Ti, run:

./distributed_train.sh 8 <path_to_imagenet> --model ddf_mul_resnet50 --lr 0.4 \
--warmup-epochs 5 --epochs 120 --sched cosine -b 128 -j 6 --amp --dist-bn reduce

Evaluation

To evaluate a pre-trained model, for example ddf_mul_resnet50, on ImageNet val, run:

python validate.py <path_to_imagenet> --model ddf_mul_resnet50 --checkpoint <path_to_checkpoint>

Inference time

To measure the inference time, run:

python test_time.py

Acknowledgement

Codebase from pytorch-image-models.

Citation

If you find this code useful for your research, please cite our paper.

@inproceedings{zhou_ddf_cvpr_2021,
               title = {Decoupled Dynamic Filter Networks},
               author = {Zhou, Jingkai and Jampani, Varun and Pi, Zhixiong and Liu, Qiong and Yang, Ming-Hsuan},
               booktitle = {IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR)},
               month = jun,
               year = {2021}
               }

theFoxofSky / ddfnet