Fully-Convolutional-Network-PyTorch

For pixel-level binary classification (semantic segmentation):
One object, two classes(background, object)

Environments

OS: Ubuntu 16.04 LTS
Framework: PyTorch 1.2.0 with cuda
Packages: NumPy, PIL, OpenCV-Python(3.4.2), etc...
GPU: Nvidia Quadro-p4000 8gb

IMAGE SIZE

My image (and label) size: (3, 500, 500)

DATA FORM

Raw images should be in the directory of "/root/data/img"
Label images should be in the directory of "/root/data/label"
change root in fcn_dataset.py to your root path

TRAIN

1. python gen_data_txt.py
2. python fcn_train.py

INFERENCE

1. change "PATH_TO_PTH" in predict.py to your path_to_model_parameters
2. change root to your dir_to_inference_images
3. python predict.py

Reference

https://zhuanlan.zhihu.com/p/32506912