Deformable Convolution V2 - MASK RCNN

Implementation of Deformable Convolution V2 in Pytorch 1.0 codes

Contributions

• The only repo that tries to reproduce the total set up of Deformable Convolution V2 in full stable Pytorch codes

• The cuda codes are ported from MXNET to Pytorch, including Modulated Deformable Convolution and Modulated ROI Pooling , supporting stable pytorch1.0 version, gradient test code is provided

• Full training and test details are provided base on the framework of maskrcnn-benchmark，Feature Mimicking branch is implemented

Tips & Notice

• train.py only support single image per batch right now due to I don't have enough resource to run multi-batch and multi-cards training, but you can easily upgrade it to support batch and multi-card training if you check on the original train_net.py from the original repo, because the framework supports everything.

• The training is ongoing, so the results and pretrained models will be published later due to it's really slow to train in my single card system.

Difference with the paper

• Inspired by the idea from Rethinking ImageNet Pre-training, the model is trained from scratch, instead of finetuning from an Imagenet pretrained model

• Due to the same issue, Batchnorm is replaced by Groupnorm

• Weights for different branches are adjusted, and OHEM is used, compare to the original paper.

How to use

Preparation

Clone the this repository.

git clone git@github.com:TreB1eN/Deformable_Convolution_V2_Pytorch1.0.gitcd Deformable_Convolution_V2_Pytorch1.0/

Download the Pretrained models(Incoming)

Installation

1. Install PyTorch 1.0 and torchvision following the official instructions.

2. Install dependencies

   pip install -r requirements.txt

3. Compile cuda ops.

   ./compile.sh  # or "PYTHON=python3 ./compile.sh" if you use system python3 without virtual environments

Inference

Incoming

Evaluate

Incoming

Perform training on COCO dataset

For the following examples to work, you need to download the COCO dataset. We recommend to symlink the path to the coco dataset to datasets/ as follows

We use minival and valminusminival sets from Detectron

# symlink the coco datasetcd ~/github/maskrcnn-benchmark
mkdir -p datasets/coco
ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
ln -s /path_to_coco_dataset/train2014 datasets/coco/train2014
ln -s /path_to_coco_dataset/test2014 datasets/coco/test2014
ln -s /path_to_coco_dataset/val2014 datasets/coco/val2014# for pascal voc dataset:ln -s /path_to_VOCdevkit_dir datasets/voc

Create following folders

Deformable_Convolution_V2
├── work_space
│   ├── model
│   ├── log
│   ├── final
│   ├── save

For training, just run:

python train.py

All detailed configuration is in configs/e2e_deformconv_mask_rcnn_R_50_C5_1x.yaml

References

maskrcnn-benchmark

paper

Contact

Email : treb1en@qq.com

Questions and PRs are welcome, especially in helping me to get better trained models, due to I don't have enough resource to trained it sufficiently