rm_ml_experiments

Installation

General Dependencies:

• javac 1.7+
• git
• python2.7

Camino installation

Run make install; you may need to call this with sudo depending on your operating system

Python installation

• Initialise a virtual environment (you need pip for this)
  • virtualenv -p /usr/bin/python2.7 ml_research
  • the directory may be different depending on your OS
• Activate the environment using source ml_research/bin/activate (your prompt should change)
• Install the requirements using pip install -r requirements.txt

Makefile commands

• generate: Generates voxel files; output path will be data/{dataset_name}
• train-all: Train on all models matching current src/config.py and produce visualisations
• install : Installs dependencies (including camino); see Installation
• clean_data : Deletes all model data generated by the program
• clean : Deletes all files installed by the program

Configuration

Configurations for voxel generation and training are defined in the config.py file. An example is given in config.py.example

Models

Each configuration is a list of strings in the models list, i.e:

models = [
  ["compartment1", "compartment2", "compartment3"],
]

Meta-parameters

The number of voxels to generate can also be set

dataset_size = 12500

As well as the signal to noise ratio

signal_to_noise = 20

The values supplied for a parameter can either be a single number value, or a list with two values. If two values are given, a sample will be taken in the range provided using uniform sampling.

Filtering

To filter out parameters you can define a list of parameters to strip from the training data

strip = ["theta", "phi", "alpha"]

Notes on running

If running on a mac, it is suggested to use caffeinate -i, so that the mac won't sleep until the tasks are completed:

caffeinate -i python main.py generate && caffeinate -i python main.py train-all

.
├── camino            # The Camino repository used to generate data
├── schemes
|   ├── PGSE_90_100t.scheme      # A scheme using a gradient strength of 100
|   └── ...
├── scripts
|   ├── run_generate      # Generates data for an Azuru machine
|   └── setup_ubuntu      # Sets up an Azure Ubuntu machine
├── src
|   ├── datasets          # Dataset Generation Module
|      ├── dataset_factory.py   # factory to generate and import data
|      └── gen_voxels.py        # entry point for data generation using threads
|   ├── fitting           # Fitting Module
|      └── fit_models.py        # Uses Camino to fit compartment models to generated data
|   ├── machine_learning  # Machine Learning module
|      ├── algortihms           # Contains the machine learning algorithms used to train models
|         ├── convolutional_nn.py         # Convolutional Neural Network using Keras and Tensorflow
|         ├── knn.py                      # K-Nearest Neighbours using Scikit-Learn
|         ├── multilayer_perceptron.py    # Multilayer Perceptron using Scikit-Learn
|         ├── random_forest.py            # Random Forest using Scikit-Learn
|         ├── ridge_regression.py         # Ridge Regression using Scikit-Learn
|         └── svm.py                      # Support Vector Machine using Scikit-Learn
|      ├── experiment.py        # A single experiment (algorithm, train_dataset, test_dataset)
|      └── ...
|   └── visualisation   # Visualisation Module
|      ├── graphs           # Various graphs
|         └── ...
|      ├── selectors        # Contains selector methods to select subsets of all the trained data
|         └── ...
|      └── ...
├── main.py           # The entry point for Camino-ml
├── Makefile          # A makefile which simplifies using main.py
├── Media          # A .gitignored folder containing visualisations
├── Results          # A .gitignored folder containing results
├── Data          # A .gitignored folder containing generated data
└── ...