/detr

End-to-End Object Detection with Transformers

Primary LanguagePythonApache License 2.0Apache-2.0

DE⫶TR: End-to-End Object Detection with Transformers

PyTorch training code and pretrained models for DETR (DEtection TRansformer). We replace the full complex hand-crafted object detection pipeline with a Transformer, and match Faster R-CNN with a ResNet-50, obtaining 42 AP on COCO using half the computation power (FLOPs) and the same number of parameters. Inference in 50 lines of PyTorch.

DETR

What it is. Unlike traditional computer vision techniques, DETR approaches object detection as a direct set prediction problem. It consists of a set-based global loss, which forces unique predictions via bipartite matching, and a Transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. Due to this parallel nature, DETR is very fast and efficient.

About the code. We believe that object detection should not be more difficult than classification, and should not require complex libraries for training and inference. DETR is very simple to implement and experiment with, and we provide a standalone Colab Notebook showing how to do inference with DETR in only a few lines of PyTorch code. Training code follows this idea - it is not a library, but simply a main.py importing model and criterion definitions with standard training loops.

For details see End-to-End Object Detection with Transformers by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko.

Model Zoo

We provide baseline DETR and DETR-DC5 models, and plan to include more in future. AP is computed on COCO 2017 val5k, and inference time is over the first 100 val5k COCO images, with torchscript transformer.

name backbone schedule inf_time box AP url size
0 DETR R50 500 0.036 42.0 download 159Mb
1 DETR-DC5 R50 500 0.083 43.3 download 159Mb
2 DETR R101 500 0.050 43.5 download 232Mb
3 DETR-DC5 R101 500 0.097 44.9 download 232Mb

COCO val5k evaluation results can be found in this gist.

COCO panoptic val5k models:

name backbone box AP segm AP PQ url size
0 DETR R50 38.7 31.4 44.0 download 165Mb
1 DETR-DC5 R50 40.2 31.9 44.6 download 165Mb
2 DETR R101 40.1 33 45.1 download 237Mb

The models are also available via torch hub, to load DETR R50 with pretrained weights simply do:

model = torch.hub.load('facebookresearch/detr', 'detr_resnet50', pretrained=True)

Usage

There are no extra compiled components in DETR and package dependencies are minimal, so the code is very simple to use. We provide instructions how to install dependencies via conda. First, clone the repository locally:

git clone https://github.com/facebookresearch/detr.git

Then, install PyTorch 1.5+ and torchvision 0.6+:

conda install -c pytorch pytorch torchvision

Install pycocotools (for evaluation on COCO) and scipy (for training):

conda install cython scipy
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

That's it, should be good to train and evaluate detection models.

(optional) to work with panoptic install panopticapi:

pip install git+https://github.com/cocodataset/panopticapi.git

Data preparation

Download and extract COCO 2017 train and val images with annotations from http://cocodataset.org. We expect the directory structure to be the following:

path/to/coco/
  annotations/  # annotation json files
  train2017/    # train images
  val2017/      # val images

Training

To train baseline DETR on a single node with 8 gpus for 300 epochs run:

python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --coco_path /path/to/coco 

A single epoch takes 28 minutes, so 300 epoch training takes around 6 days on a single machine with 8 V100 cards. To ease reproduction of our results we provide results and training logs for 150 epoch schedule (3 days on a single machine), achieving 39.5/60.3 AP/AP50.

We train DETR with AdamW setting learning rate in the transformer to 1e-4 and 1e-5 in the backbone. Horizontal flips, scales an crops are used for augmentation. Images are rescaled to have min size 800 and max size 1333. The transformer is trained with dropout of 0.1, and the whole model is trained with grad clip of 0.1.

Evaluation

To evaluate DETR R50 on COCO val5k with a single GPU run:

python main.py --batch_size 2 --no_aux_loss --eval --resume https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth --coco_path /path/to/coco

We provide results for all DETR detection models in this gist. Note that numbers vary depending on batch size (number of images) per GPU. Non-DC5 models were trained with batch size 2, and DC5 with 1, so DC5 models show a significant drop in AP if evaluated with more than 1 image per GPU.

Multinode training

Distributed training is available via Slurm and submitit:

pip install submitit

Train baseline DETR-6-6 model on 4 nodes for 300 epochs:

python run_with_submitit.py --timeout 3000 --coco_path /path/to/coco

License

DETR is released under the Apache 2.0 license. Please see the LICENSE file for more information.

Contributing

We actively welcome your pull requests! Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.