Yolov2 Pytorch Implementation

This repository aims to learn and understand the YOLO algorithm. I am a beginner of deep learning, and I found the best way to learn a deep learning algorithm is to implement it from scratch. So if you also feel this way, just follow this repo! The code in this projects is clear and easier to understand, and I also documented it as much as possible.

Purpose

• train pascal voc
• multi-GPUs support
• test
• pascal voc validation
• data augmentation
• pretrained network
• reorg layer
• multi-scale training
• reproduce original paper's mAP

Main Results

	training set	test set	mAP@416	mAP@544
this repo	VOC2007+2012	VOC2007	72.7	74.6
original paper	VOC2007+2012	VOC2007	76.8	78.6

Running time: ~19ms (52FPS) on GTX 1080

Prerequisites

• python 3.5.x
• pytorch 0.4.1
• tensorboardX
• opencv3
• pillow

Preparation

First clone the code

git clone https://github.com/tztztztztz/yolov2.pytorch.git

Install dependencies

pip install -r requirements.txt

Then create some folder

mkdir output 
mkdir data

Demo

Download the pretrained weights

wget http://pjreddie.com/media/files/yolo-voc.weights

You can run the demo with cpu mode

python demo.py

Or with gpu mode

python demo.py --cuda true

Training on PASCAL VOC

Prepare the data

1. Download the training data.

   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar
   
   # download 2012 data
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar

2. Extract the training data, all the data will be in one directory named VOCdevit. We use $VOCdevit to represent the data root path

   tar xvf VOCtrainval_06-Nov-2007.tar
   tar xvf VOCtest_06-Nov-2007.tar
   tar xvf VOCdevkit_08-Jun-2007.tar
   
   # 2012 data
   tar xvf VOCtrainval_11-May-2012.tar

3. It should have this basic structure

   $VOCdevkit/                           # development kit
   $VOCdevkit/VOCcode/                   # VOC utility code
   $VOCdevkit/VOC2007                    # image sets, annotations, etc.
   

4. Create symlinks for the PASCAL VOC dataset

   cd yolov2.pytorch
   mkdir data
   cd data
   mkdir VOCdevkit2007
   cd VOCdevkit2007
   ln -s $VOCdevit/VOC2007 VOC2007
   
   # mkdir VOCdevkit2012
   # cd VOCdevkit2012
   # ln -s $VOCdevit/VOC2012 VOC2012
   

Download pretrained network

cd yolov2.pytorch
cd data
mkdir pretrained
cd pretrained
wget https://pjreddie.com/media/files/darknet19_448.weights

Train the model

python train.py --cuda true

If you want use multiple GPUs to accelerate the training. you can use the command below.

python train.py --cuda true --mGPUs true

NOTE: Multi-scale training uses more GPU memory. If you have only one GPU with 8G memory, it's better to set multi-scale=False in config/config.py. See link.

Testing

python test.py --cuda true