hiyouga / image-segmentation-pytorch

U-Net for image segmentation, PyTorch implementation.

Image-Segmentation-PyTorch

PyTorch implementation of the U-Net for image segmentation.

Requirement

• Python 3
• numpy
• matplotlib
• Pillow
• torch
• torchvision
• pydensecrf

Dataset

Based on the shoe dataset [Google Drive] provided by our teacher.

Usage

Training

python train.py

Inference

python train.py --inference True --checkpoint [*.pt]

The checkpoint files can be found in the state_dict folder.

Show help message

python train.py -h

Implemented model

Ronneberger, O., Fischer, P., and Brox, T. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI). [pdf]

Notes

The model was trained from scratch on Tesla V100 32GB*4. Training the model takes 3.6GB of memory and predicting images takes 1.3GB. If you don't have enough GPU memory, consider using bilinear up-sampling rather than transposed convolution in the model.

Acknowledgements

• Some of the code is borrowed from milesial.
• This is a personal homework for "Machine Learning Theory and Application" in BUAA Summer School.

License

MIT