README.md

A Mixed-Scale Dense Convolutional Neural Network for image analysis

This repository is a simplified implementation of the paper: A mixed-scale dense convolutional neural network for image analysis, for image segmentation. The implementation uses the PyTorch framework.

Note: Work in progress. Any contribution is appreciated.

The files are organized as follows:

.
├── config.py (contains the configuration class which handles reading config for experiments)
├── data
│   └── README.md
├── data_handler.py (data handler class for making and modifying datasets)
├── experiment
│   └── cfg.yml (template and default config file [DO NOT REMOVE])
├── main.py (main file that runs the model)
├── model.py (model architecture specifications)
├── README.md
└── utils.py (utility functions used throughout the code)

Creating an Experiment

To create a new experiment, create the following directory structure.

.
├── Annotations (Stores annotations for the generated segments from the model [TODO])
├── cfg.yml (configuration file for the experiment)
├── checkpoints (Saves model checkpoints for selective use)
└── output (stores the generated segmented images)
    └── training (images generated during the training process)

To run an experiment

The main.py script is used to run the experiment. To train the model, without using a pretrained checkpoint, to write the images in experiment directory, run the following command:

python main.py --exp_dir=<EXP_DIR> --cfg=<CONFIG_PATH> --nopretrained --write_images --train

To just run the model you have trained, update the config file with the path to the latest checkpoint and run the following command:

python main.py --exp_dir=<EXP_DIR> --cfg=<CONFIG_PATH> --pretrained --write_images --notrain --viz

Dataset

The structure used when creating the dataset is as follows:

├── README.md (contains any information about the dataset)
├── top (contains the RGB images)
└── gt (ground truth data)

This structure is to be used for all dataset creation and adaptation pruposes. Change congif file and config.py for different dataset.

Acknowledgements

Thanks to the authors of the Paper: A mixed-scale dense convolutional neural network for image analysis (Pelt, D. M. et. al.)