/pytorch_yolov1

Yolov1 pytorch implement

Primary LanguagePython

YOLO:

It train YOLO in VOC2007(train,vaild)+VOC2012(train,vaild) datasets and test on VOC2007(test).
I replace FC layer with Conv Layer to save memory. And you can also training with FC layer!

Evaluation:

Model mAp.
My model fc layer -
My model conv layer 0.66
Origin papar 0.63

Dependence:

  • Python3
  • Pytorch 1.3 or higher
  • apex

Install & Train

Install Apex

git clone https://github.com/NVIDIA/apex  
cd apex  
pip install -v --no-cache-dir ./

Train on VOCdatasets

  1. Download the training, validation, test data and VOCdevkit
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
  1. Extract all of these tars into one directory named VOCdevkit
tar xvf VOCtrainval_06-Nov-2007.tar tar xvf VOCtest_06-Nov-2007.tar tar xvf VOCdevkit_08-Jun-2007.tar
  1. It should have this basic structure
 $VOCdevkit/
 $VOCdevkit/VOC2007   
 $VOCdevkit/VOC2012
  1. Generate the train and test list
python tools/convert_voc.py --dir_path ./ 
cat VOC2007_train.txt VOC2012_train.txt VOC2007_val.txt VOC2012_val.txt >> train.txt
  1. Download the pretrain model
 wget https://pjreddie.com/media/files/darknet19_448.conv.23.
  1. Configure the training param
 bash train.sh

Test on VOCdatasets

 bash test.sh

Demo

vis_detector.ipynb

Samples:

imgs
imgs
imgs

Reference

maskrcnn_benchmark