PSPNet

This is an unofficial implementation of Pyramid Scene Parsing Network (PSPNet) in Chainer.

Training

Requirement

• Python 3.4.4+

• Chainer 3.0.0b1+

• ChainerMN master

• CuPy 2.0.0b1+

• ChainerCV 0.6.0+

• NumPy 1.12.0+

• tqdm 4.11.0+

pip install chainer --pre
pip install cupy --pre
pip install git+git://github.com/chainer/chainermn
pip install git+git://github.com/chainer/chainercv
pip install tqdm
````

---

# Inference using converted weights

## Requirement

- Python 3.4.4+
    - Chainer 3.0.0b1+
    - ChainerCV 0.6.0+
    - Matplotlib 2.0.0+
    - CuPy 2.0.0b1+
    - tqdm 4.11.0+

## 1. Run demo.py

### Cityscapes

$ python demo.py -g 0 -m cityscapes -f aachen_000000_000019_leftImg8bit.png

### Pascal VOC2012

$ python demo.py -g 0 -m voc2012 -f 2008_000005.jpg

### ADE20K

$ python demo.py -g 0 -m ade20k -f ADE_val_00000001.jpg

### FAQ

If you get `RuntimeError: Invalid DISPLAY variable`, how about specifying the matplotlib's backend by an environment variable?

$ MPLBACKEND=Agg python demo.py -g 0 -m cityscapes -f aachen_000000_000019_leftImg8bit.png

---

# Convert weights by yourself

**Caffe is NOT needed** to convert `.caffemodel` to Chainer model. Use `caffe_pb2.py`.

## Requirement

- Python 3.4.4+
    - protobuf 3.2.0+
    - Chainer 3.0.0b1+
    - NumPy 1.12.0+

## 1. Download the original weights

Please download the weights below from the author's repository:

- pspnet50_ADE20K.caffemodel: [GoogleDrive](https://drive.google.com/open?id=0BzaU285cX7TCN1R3QnUwQ0hoMTA)
- pspnet101_VOC2012.caffemodel: [GoogleDrive](https://drive.google.com/open?id=0BzaU285cX7TCNVhETE5vVUdMYk0)
- pspnet101_cityscapes.caffemodel: [GoogleDrive](https://drive.google.com/open?id=0BzaU285cX7TCT1M3TmNfNjlUeEU)

**and then put them into `weights` directory.**

## 2. Convert weights

$ python convert.py ```

Reference

• The original implementation by authors is: hszhao/PSPNet

• The original paper is:

• Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, Jiaya Jia, "Pyramid Scene Parsing Network", Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017