shiangwang/fcn

PyTorch Implementation of Fully Convolutional Networks, for VGG and ResNet backbones.

FCN

PyTorch implementation of Fully Convolutional Networks, for VGG and RESNET backbones.

Requirements

• pytorch >= 1.1.0
• torchvision >= 0.3.0
• fcn
• Jupyter Notebook

pip install -r requirements.txt

Data

sh download_dataset.sh will download the PascalVOC dataset inside data folder.

Training

python main.py --gpu_id=0 --backbone=vgg --fcn=32s --root_dataset=./data/Pascal_VOC --mode=train

• backbone: vgg or resnet [vgg default].
• fcn: 32s, 16s, 8s for vgg, and 101, 50 for resnet [32s default].
• mode: train, val, demo [train default].
• gpu_id: [-1 default (cpu)].
• resume: For val and demo, if no resume path is given, code will use original FCN pre-trained weights.

Acknowledgement

This repo is built upon wkentaro's code and some snippets can be just a mirror.