PyTorch Implementation of Multiple Instance Detection Network with Online Instance Classifier Refinement (OICR)
git clone http://www.github.com/jd730/OICR-pytorch
- Python 3.5 or higher
- Pytorch 0.4.0 (not 0.4.1)
- CUDA 8.0 or higher
- 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, creat softlinks in the folder data/.
wget https://dl.dropboxusercontent.com/s/orrt7o6bp6ae0tc/selective_search_data.tgz
tar -xvf selective_search_data.tgz
rm -rf selective_search_data.tgz
move selective_search_data folder into data folder.
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.
As pointed out by ruotianluo/pytorch-faster-rcnn, choose the right -arch in make.sh file, to compile the cuda code:
| GPU model | Architecture |
|---|---|
| TitanX (Maxwell) | sm_52 |
| TitanX (Pascal) | sm_61 |
| TitanV or V100 | sm_70 |
| GTX 960M | sm_50 |
| GTX 1080 (Ti) | sm_61 |
| Grid K520 (AWS g2.2xlarge) | sm_30 |
| Tesla K80 (AWS p2.xlarge) | sm_37 |
More details about setting the architecture can be found here or here
-
Install pip dependency
pip install -r requirement.txt -
Compile th ecuda dependencies
cd lib & sh make.sh
Green rectangulars are the results of OICR and red rectangulars are ground truth.
trainval_net.py : main training code.
test_oicr.py : test code modified from oicr test code.
datasets/ : loading pascal_voc
roi_data_layer/ : loading batch, making roidb and minibatch
model/ includes network and roi_align,crop,pooling.
In model/oicr*/ there are two files. One is vgg network which assign each layers and the other is oicr class which decides how to make a forward and how network is composed of.
- Use tensorboard
--use_tbflag, but sometime tensorflow session is dead abruptly. - Use logger. In this code, the program automatically generate
log.txtandprogress.csvin your directory. You can check this using note.ipynb. You can easily understanding through reading the example code in the notebook.
file_path = 'jdhwang/1006_seq_tr3/log/progress.csv'
plot_reward_curve_seborn(file_path, mavg=True, mavg_v=1, n=N, target_field='midn_loss', print_header=False,txt_offset=1.0, newfig=False, conv=20)
CUDA_VISIBLE_DEVICES=1,2 python3 trainval_net.py --dataset pascal_voc --net vgg16 \ --bs 4 --nw 4 --save_dir='output' --model='oicr' \ --lr 0.001 --cuda --disp_interval 50 --mGPUs --vis \ --checkpoint_interval=500
CUDA_VISIBLE_DEVICES=2 python3 test_oicr.py --dataset pascal_voc --net vgg16 --checkpoint 70000 --load_dir='output' --cuda --model='oicr'--vis
bs (batch_size) should be divisable by 2. On Caffe, batch is defined as how many forward operations before bacward, and the it does not divide the accumulated loss. To follow this definition and improve the performance, my code automatically forward twice and backward once without divison. See here