Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
120 lines (99 loc) · 3.98 KB

ARCHITECTURE.md

File metadata and controls

120 lines (99 loc) · 3.98 KB
Raw

Track-to-Learn

The overall structure of the project is

- TrackToLearn
  - algorithms
  - datasets
  - environments
  - experiment
  - searchers
  - trainers
  - utils
- example_model
- scripts
- cc_scripts
- ttl_track.py
- ttl_validation.py

In TrackToLearn, you will find the codebase for the project. In scripts, you will find all the scripts used to train agents in What matters in ...1 and Incorporating anatomical priors into...2. In cc_scripts, you will find slurm scripts that have been used to do the architecture search of 1. example_model contains the weights and hyperparameters of an agent trained on the ISMRM2015 dataset.

The entry points for launching TrackToLearn are in the root folder, trainers or searchers.

- TrackToLearn
  - ttl_track.py
  - ttl_validation.py
  - searchers
    - a2c_searcher.py
    - acktr_searcher.py
    - ddpg_searcher.py
    - ppo_searcher.py
    - sac_auto_searcher.py
    - sac_searcher.py
    - td3_searcher.py
    - trpo_searcher.py
    - vpg_searcher.py
  - trainers
    - a2c_train.py
    - acktr_train.py
    - ddpg_train.py
    - ppo_train.py
    - sac_auto_train.py
    - sac_train.py
    - td3_train.py
    - trpo_train.py
    - vpg_train.py

The runners module contains scripts for tracking either on a "dataset" (validation.py) or on arbitrary files (track.py, similarly to launching tracking in scilpy3). The searchers module contains scripts for launching an hyperparameter search for the relevant algorithm. The trainers module contains scripts for launching training for the relevant algorithm.

The algorithms module contains several implementations of RL algorithms.

- TrackToLearn
  - algorithms
    - rl.py
    - utils.py
    - a2c.py
    - acktr.py
    - ddpg.py
    - ppo.py
    - sac_auto.py
    - sac.py
    - td3.py
    - trpo.py
    - vpg.py
    - shared
      - onpolicy.py
      - offpolicy.py
      - replay.py

The rl submodule contains the core of all RL algorithms implementations and most things that are relevant to all (such as the RL loop at inference, for example). The algorithms/utils submodule contains functions relevant to most RL algorithms. The shared submodule mostly contains classes relevant to polices and critics. Other files are implementations of RL algorithms.

The experiment submodule contains core classes and functions for launching, monitoring and reproducing experiments.

- TrackToLearn
  - experiment 
    - experiment.py
    - train.py
    - ttl.py

experiment.py contains the base class for experiments as well as most of the arguments used in entry-point scripts. ttl.py contains the base class for TrackToLearn experiments, which may be training or tracking or other. train.py contains the base class for training runs, from which "trainers" inherit.

The enviroments submodule contains everything related to RL environments.

- TrackToLearn
  - environments
    - env.py
    - interface_tracker.py
    - noisy_tracker.py
    - reward.py
    - tracker.py
    - utils.py

env.py contains the base abstract class for environments, BaseEnv, in Track-to-Learn. tracker.py contains several concrete classes that inherit from BaseEnv. interface_tracker.py and noisy_tracker.py contain classes that inherit from classes in tracker. reward.py contains the class handling the reward function.

Finally, the datasets submodule contains everything related to the creation and processing of datasets.

- TrackToLearn
  - datasets
    - create_dataset.py
    - processing.py
    - utils.py

The create_dataset.py script can be called to create a HDF5 containing training, validation and test subjects. processing.py contains util functions related to dataset creation.

Footnotes

  1. "What matters in reinforcement learning for tractography" ↩ ↩2

  2. Incorporating anatomical priors into Track-to-Learn, ISMRM Workshop on Diffusion MRI: From Research to Clinic, poster #34. ↩

  3. scilpy: https://github.com/scilus/scilpy ↩