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.
./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