Detectorch - detectron for PyTorch

(Disclaimer: this is work in progress and does not feature all the functionalities of detectron. Currently only inference and evaluation are supported -- no training)
(News: Now supporting FPN and ResNet-101!)

This code allows to use some of the Detectron models for object detection from Facebook AI Research with PyTorch.

It currently supports:

  • Fast R-CNN
  • Faster R-CNN
  • Mask R-CNN

It supports ResNet-50/101 models with or without FPN. The pre-trained models from caffe2 can be imported and used on PyTorch.


Example Mask R-CNN with ResNet-101 and FPN.


Both bounding box evaluation and instance segmentation evaluation where tested, yielding the same results as in the Detectron caffe2 models. These results below have been computed using the PyTorch code:

Model box AP mask AP model id
fast_rcnn_R-50-C4_2x 35.6 36224046
fast_rcnn_R-50-FPN_2x 36.8 36225249
e2e_faster_rcnn_R-50-C4_2x 36.5 35857281
e2e_faster_rcnn_R-50-FPN_2x 37.9 35857389
e2e_mask_rcnn_R-50-C4_2x 37.8 32.8 35858828
e2e_mask_rcnn_R-50-FPN_2x 38.6 34.5 35859007
e2e_mask_rcnn_R-101-FPN_2x 40.9 36.4 35861858


Training code is experimental. See for training Fast R-CNN. It seems to work, but slow.


First, clone the repo with git clone --recursive so that you also clone the Coco API.

The code can be used with PyTorch 0.3.1 or PyTorch 0.4 (master) under Python 3. Anaconda is recommended. Other required packages

  • torchvision (conda install torchvision -c soumith)
  • opencv (conda install -c conda-forge opencv )
  • cython (conda install cython)
  • matplotlib (conda install matplotlib)
  • scikit-image (conda install scikit-image)
  • ninja (conda install ninja) (required for Pytorch 0.4 only)

Additionally, you need to build the Coco API and RoIAlign layer. See below.

Compiling the Coco API

If you cloned this repo with git clone --recursive you should have also cloned the cocoapi in lib/cocoapi. Compile this with:

cd lib/cocoapi/PythonAPI
make install

Compiling RoIAlign

The RoIAlign layer was converted from the caffe2 version. There are two different implementations for each PyTorch version:

  • Pytorch 0.4: RoIAlign using ATen library (lib/cppcuda). Compiled JIT when loaded.
  • PyTorch 0.3.1: RoIAlign using TH/THC and cffi (lib/cppcuda_cffi). Needs to be compiled with:
cd lib/cppcuda_cffi

Quick Start

Check the demo notebook.