Deep ResUnet and ResUnet ++

Unofficial Pytorch implementation of following papers :

• Deep ResUnet
• ResUnet ++

Note

• This repo written for experimentation (fun) purpose and heavily hard coded, so avoid to use this as it is in production environement.
• I only wrote ResUnet and ResUnet++ model, Unet is pre-implemented and borrows from this repo.
• Use your own pre-processing and dataloader, dataloader and pre-processing of this repo written for specific use case.
• This repo only tested on Massachusetts Roads Dataset.

Pre-processing

• This pre-processing is for specific use case and follows strict directory structure.

python preprocess.py --config "config/default.yaml" --train training_files_dir --valid validation_files_dir

• Training and validation directories passed in args above should contain two folders input for input images and output for target images. And all images are of fixed square size (in this case 1500 * 1500 pixels).
• Pre-processing crop each input and target image into several fixed size (in this case 224 * 224) small cropped images and saved into input_crop and mask_crop respectively on training and validation dump directories as in config file.
• You can change training and validation dump directories from config file i.e. configs/default.yaml.

Training

python train.py --name "default" --config "config/default.yaml"

For Tensorboard: tensorboard --logdir logs/

References

• DenseASPP for Semantic Segmentation in Street Scenes
• ResUNet++ with Conditional Random Field
• SENet
• Road Extraction Using PyTorch
• ASPP Module
• Deep Residual-Unet
• Squeeze-and-Excitation Networks
• ResUNet++
• Unet
• Brain tumor segmentation