A Faster Pytorch Implementation of Faster R-CNN

Introduction

💥 Good news! This repo supports pytorch-1.0 now!!! We borrowed some code and techniques from maskrcnn-benchmark. Just go to pytorch-1.0 branch!

This project is a faster pytorch implementation of faster R-CNN, aimed to accelerating the training of faster R-CNN object detection models. Recently, there are a number of good implementations:

rbgirshick/py-faster-rcnn, developed based on Pycaffe + Numpy
longcw/faster_rcnn_pytorch, developed based on Pytorch + Numpy
endernewton/tf-faster-rcnn, developed based on TensorFlow + Numpy
ruotianluo/pytorch-faster-rcnn, developed based on Pytorch + TensorFlow + Numpy

During our implementing, we referred the above implementations, especially longcw/faster_rcnn_pytorch. However, our implementation has several unique and new features compared with the above implementations:

It is pure Pytorch code. We convert all the numpy implementations to pytorch!
It supports multi-image batch training. We revise all the layers, including dataloader, rpn, roi-pooling, etc., to support multiple images in each minibatch.
It supports multiple GPUs training. We use a multiple GPU wrapper (nn.DataParallel here) to make it flexible to use one or more GPUs, as a merit of the above two features.
It supports three pooling methods. We integrate three pooling methods: roi pooing, roi align and roi crop. More importantly, we modify all of them to support multi-image batch training.
It is memory efficient. We limit the image aspect ratio, and group images with similar aspect ratios into a minibatch. As such, we can train resnet101 and VGG16 with batchsize = 4 (4 images) on a single Titan X (12 GB). When training with 8 GPU, the maximum batchsize for each GPU is 3 (Res101), totaling 24.
It is faster. Based on the above modifications, the training is much faster. We report the training speed on NVIDIA TITAN Xp in the tables below.

What we are doing and going to do

Support both python2 and python3 (great thanks to cclauss).
Add deformable pooling layer (mainly supported by Xander).
Support pytorch-0.4.0 (this branch).
Support tensorboardX.
Support pytorch-1.0 (go to pytorch-1.0 branch).

Other Implementations

Feature Pyramid Network (FPN)
Mask R-CNN (~~ongoing~~ already implemented by roytseng-tw)
Graph R-CNN (extension to scene graph generation)

Tutorial

Blog by ankur6ue

Benchmarking

We benchmark our code thoroughly on three datasets: pascal voc, coco and visual genome, using two different network architectures: vgg16 and resnet101. Below are the results:

1). PASCAL VOC 2007 (Train/Test: 07trainval/07test, scale=600, ROI Align)

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
VGG-16	1	1	1e-3	5	6	0.76 hr	3265MB	70.1
VGG-16	1	4	4e-3	8	9	0.50 hr	9083MB	69.6
VGG-16	8	16	1e-2	8	10	0.19 hr	5291MB	69.4
VGG-16	8	24	1e-2	10	11	0.16 hr	11303MB	69.2
Res-101	1	1	1e-3	5	7	0.88 hr	3200 MB	75.2
Res-101	1	4	4e-3	8	10	0.60 hr	9700 MB	74.9
Res-101	8	16	1e-2	8	10	0.23 hr	8400 MB	75.2
Res-101	8	24	1e-2	10	12	0.17 hr	10327MB	75.1

2). COCO (Train/Test: coco_train+coco_val-minival/minival, scale=800, max_size=1200, ROI Align)

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
VGG-16	8	16	1e-2	4	6	4.9 hr	7192 MB	29.2
Res-101	8	16	1e-2	4	6	6.0 hr	10956 MB	36.2
Res-101	8	16	1e-2	4	10	6.0 hr	10956 MB	37.0

NOTE. Since the above models use scale=800, you need add "--ls" at the end of test command.

3). COCO (Train/Test: coco_train+coco_val-minival/minival, scale=600, max_size=1000, ROI Align)

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
Res-101	8	24	1e-2	4	6	5.4 hr	10659 MB	33.9
Res-101	8	24	1e-2	4	10	5.4 hr	10659 MB	34.5

4). Visual Genome (Train/Test: vg_train/vg_test, scale=600, max_size=1000, ROI Align, category=2500)

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
VGG-16	1 P100	4	1e-3	5	20	3.7 hr	12707 MB	4.4

Thanks to Remi for providing the pretrained detection model on visual genome!

Click the links in the above tables to download our pre-trained faster r-cnn models.
If not mentioned, the GPU we used is NVIDIA Titan X Pascal (12GB).

Preparation

First of all, clone the code

git clone https://github.com/jwyang/faster-rcnn.pytorch.git

Then, create a folder:

cd faster-rcnn.pytorch && mkdir data

prerequisites

Python 2.7 or 3.6
Pytorch 1.0 (for Pytorch 0.4.0 go to master branch)
CUDA 8.0 or higher

Data Preparation

PASCAL_VOC 07+12: Please follow the instructions in py-faster-rcnn to prepare VOC datasets. Actually, you can refer to any others. After downloading the data, create softlinks in the folder data/.
COCO: Please also follow the instructions in py-faster-rcnn to prepare the data.
Visual Genome: Please follow the instructions in bottom-up-attention to prepare Visual Genome dataset. You need to download the images and object annotation files first, and then perform proprecessing to obtain the vocabulary and cleansed annotations based on the scripts provided in this repository.

Pretrained Model

We used two pretrained models in our experiments, VGG and ResNet101. You can download these two models from:

VGG16: Dropbox, VT Server
ResNet101: Dropbox, VT Server

Download them and put them into the data/pretrained_model/.

NOTE. We compare the pretrained models from Pytorch and Caffe, and surprisingly find Caffe pretrained models have slightly better performance than Pytorch pretrained. We would suggest to use Caffe pretrained models from the above link to reproduce our results.

If you want to use pytorch pre-trained models, please remember to transpose images from BGR to RGB, and also use the same data transformer (minus mean and normalize) as used in pretrained model.

Compilation

Install all the python dependencies using pip:

pip install -r requirements.txt

Compile the cuda dependencies using following simple commands:

cd lib
make

Compile python cocoapi (repository from https://github.com/cocodataset/cocoapi):

cd lib\python_api_coco
make

It will compile all the modules you need, including NMS, ROI_Pooing, ROI_Align and ROI_Crop. The default version is compiled with Python 2.7, please compile by yourself if you are using a different python version.

As pointed out in this issue, if you encounter some error during the compilation, you might miss to export the CUDA paths to your environment.

Train

Before training, set the right directory to save and load the trained models. Change the arguments "save_dir" and "load_dir" in trainval_net.py and test_net.py to adapt to your environment.

To train a faster R-CNN model with vgg16 on pascal_voc, simply run:

CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net.py \
                   --dataset pascal_voc --net vgg16 \
                   --bs $BATCH_SIZE --nw $WORKER_NUMBER \
                   --lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
                   --cuda

where 'bs' is the batch size with default 1. Alternatively, to train with resnet101 on pascal_voc, simple run:

 CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net.py \
                    --dataset pascal_voc --net res101 \
                    --bs $BATCH_SIZE --nw $WORKER_NUMBER \
                    --lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
                    --cuda

Above, BATCH_SIZE and WORKER_NUMBER can be set adaptively according to your GPU memory size. On Titan Xp with 12G memory, it can be up to 4.

If you have multiple (say 8) Titan Xp GPUs, then just use them all! Try:

python trainval_net.py --dataset pascal_voc --net vgg16 \
                       --bs 24 --nw 8 \
                       --lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
                       --cuda --mGPUs

Change dataset to "coco" or 'vg' if you want to train on COCO or Visual Genome.

Test

If you want to evaluate the detection performance of a pre-trained vgg16 model on pascal_voc test set, simply run

python test_net.py --dataset pascal_voc --net vgg16 \
                   --checksession $SESSION --checkepoch $EPOCH --checkpoint $CHECKPOINT \
                   --cuda

Specify the specific model session, checkepoch and checkpoint, e.g., SESSION=1, EPOCH=6, CHECKPOINT=416.

Demo

If you want to run detection on your own images with a pre-trained model, download the pretrained model listed in above tables or train your own models at first, then add images to folder $ROOT/images, and then run

python demo.py --net vgg16 \
               --checksession $SESSION --checkepoch $EPOCH --checkpoint $CHECKPOINT \
               --cuda --load_dir path/to/model/directoy

Then you will find the detection results in folder $ROOT/images.

Note the default demo.py merely support pascal_voc categories. You need to change the line to adapt your own model.

Below are some detection results:

Webcam Demo

You can use a webcam in a real-time demo by running

python demo.py --net vgg16 \
               --checksession $SESSION --checkepoch $EPOCH --checkpoint $CHECKPOINT \
               --cuda --load_dir path/to/model/directoy \
               --webcam $WEBCAM_ID

The demo is stopped by clicking the image window and then pressing the 'q' key.

Authorship

This project is equally contributed by Jianwei Yang and Jiasen Lu, and many others (thanks to them!).

Citation

@article{jjfaster2rcnn,
    Author = {Jianwei Yang and Jiasen Lu and Dhruv Batra and Devi Parikh},
    Title = {A Faster Pytorch Implementation of Faster R-CNN},
    Journal = {https://github.com/jwyang/faster-rcnn.pytorch},
    Year = {2017}
}

@inproceedings{renNIPS15fasterrcnn,
    Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
    Title = {Faster {R-CNN}: Towards Real-Time Object Detection
             with Region Proposal Networks},
    Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
    Year = {2015}
}

nico-zck/faster_rcnn_pytorch