road-segmentation

This repository aims to provide a simple environment to easily train and test convolutional neural network models for semantic segmentation task.

Dependencies

The best way to set the environment up is to install Anaconda or Miniconda which should manage the installation of CUDA for you. We use python3.6.

After creating and activating conda virtual environment, install dependencies by running

$ pip install -r requirements.txt

Usage

All the script commands mentioned below are accompanied by their help description (click package).

Dataset setup

Script data/split_data.py does simple preprocessing and splits data into train, validation and test folders.

Run $ python data/split_data.py --help to find out more.

Training, Evaluating, Predicting

File run.py contains commands for training a model, evaluating it on test set and creating predictions.

Repository contains many models defined in models folder and imported in run.py.

In order to train a model, run.py contains command train.

$ python run.py train --help

Once the model is trained, we may evaluate it using evaluate command.

$ python run.py evaluate --help

If we want to predict test images, we may use predict command.

$ python run.py predict --help

The script also contains command visualize for visualizing feature maps.

$ python run.py visualize --help

Active learning

The main active learning script is located in the root of this directory and called al.py. If we want to run training procedure for specific model, it has to be specified and loaded in run.py. It is possible to specify sampling method, stopping condition and the epoch strategy, all of which are defined in alutils.py file. For more information on how to specify these parameters, run:

$ python al.py simulate --help