Hugin

Hugin helps scientists run Machine Learning experiments on geospatial raster data.

Overall Hugin aims to facilitate experimentation with multiple machine learning problems, like:

Classification
Segmentation
Super-Resolution

Currently Hugin builds on top of the Keras machine learning library but it also aims to support, in the future, additional backends like scikit-learn.

Installation

Prerequisits

As previously mentioned Hugin builds on top of Keras but, depending on your usecase it might also depend on TensorFlow for providing additional IO functionality (particularly cloud storage), or for supporting models specific for various Keras backends

Using pip

From PyPi

ToDo

From source repository

You can install Hugin using the following command:

pip install git+http://github.com/mneagul/hugin#egg=hugin

From source code

When installing from source code we recommend installation inside a specially created virtual environment.

Installing from source code involves running the setup.py inside you python environment.

python setup.py install

Using

Using Hugin involves two steps:

training
prediction

Both steps are driven using dedicated configuration files.

Training

Training can be started as follows:

hugin train --config training_config.yaml

An example training configuration can be found in docs/examples/train.yaml.

Prediction

Prediction can be started as follows:

hugin predict \
    --ensemble-config prediction.yaml \
    --input-dir /path/to/input/dir \
    --output-dir /tmp/output

An example prediction configuration can be found in docs/examples/predict.yaml

The predict command requires at least three arguments:

--ensemble-config: representing the prediction configuration file
--input-dir: representing the directory holding data that should server as input for prediction
--output-dir: directory for storing the outputs (predictions)

Developing your own models

Developing your own is really simple. The only thing needed is to create a python file that creates the corresponding Keras model.

The code building the model needs to be resolvable by Hugin: it needs to be available in the PYTHONPATH.

Let's consider you are preparing a new segmentation related experiment.

The most simple approach would be to create a new directory containing both the source code and model configuration, like in the following example:

mysegmentation/
├── model.py
├── predict.yaml
└── train.yaml

The files involved in this example are:

model.py: it contains the source code for the model. An example can be found in the source distribution, in src/hugin/models/unet/unetv14.py
train.yaml: the configuration used for training. An example can be found in docs/examples/train.yaml
predict.yaml: the configuration used for prediction. An example can be found in docs/examples/predict.yaml

After creating your model and preparing your experiment configuration you can start training, by running:

hugin train --config train.yaml

As specified in the model configuration the model will train for 10 epochs and produce the final model.

alexm98/hugin