Hydronaut 2024.2

A framework for exploring the depths of hyperparameter space with Hydra and MLflow.

---

title: README author: Jan-Michael Rye

Synopsis

Hydronaut is a framework for exploring the depths of hyperparameter space with Hydra and MLflow. Its goal is to encourage and facilitate the use of these tools while handling the sometimes unexpected complexity of using them together. Users benefit from both without having to worry about the implementation and are thus able to focus on developing their models.

Hydra allows the user to organize all hyperparameters via simple YAML files with support for runtime overrides via the command-line. It also allows the user to explore the hyperparameter space with automatic sweeps that are easily parallelized. These sweeps can either explore all possible parameter combinations or they can use any of the optimizing sweepers supported by Hydra such as the Optuna Sweeper plugin. The hyperparameters used for every run are automatically saved for future reference and reproducibility.

MLflow is a platform for tracking experiments and their results, among other things. The library provides numerous logging functions to track hyperparameters, metrics, artifacts and models of every run so that nothing is ever lost or forgotten. The results can be readily perused, compared and managed via both command-line and web interfaces. It can also be used to push trained models to registries to share with others.

Links

GitLab

• Homepage
• Repository
• Documentation
• Issues
• GitLab package registry

Other Repositories

• Python Package Index
• Software Heritage
• HAL open science

Related

• Hydronaut Tutorials
• MolPred - A Hydronaut-based framework that integrates ChemFeat for building machine- and deep-learning models to predict properties of molecules.

Citations

If you use this software, please cite it using the metadata in CITATION.cff. The file can be converted to various output formats using cffconvert.

Installation

Install the Hydronaut package from the Python Package Index using any standard Python package manager, e.g.

# Uncomment the following 2 lines to create and activate a virtual environment.
# python -m venv venv
# source venv/bin/activate
pip3 install --upgrade hydronaut

Packages are also provided via the GitLab package registry.

Hydronaut can also be installed from source with any standard Python package manager that supports pyproject.toml files. For example, to install it with pip, either locally or in a virtual environment, run the following commands:

git clone --recursive https://gitlab.inria.fr/jrye/hydronaut
cd hydronaut
# Uncomment the following 2 lines to create and activate a virtual environment.
# python -m venv venv
# source venv/bin/activate
pip install --upgrade .

Note that the Hydronaut Git repository contains Git submodules. It should be recursively cloned with git clone --recursive https://gitlab.inria.fr/jrye/hydronaut to check out all requirements. Alternatively, after cloning the repository non-recursively one can run git submodule update --init --recursive to fully initialize the repository. This can also be accomplished with the script hydronaut-initialize.sh, which is provided for convenience.

The project also provides the script hydronaut-install_in_venv.sh which can be used to install the package in a virtual environment. Internally, the script uses pip-install.sh from the utility-scripts submodule which can circumvent a bug in the way thathatch-vcs handles Git submodules.

Submodules

• utility-scripts - Required for some of the scripts included in the Git repository but not required for the Python package itself.

Usage

There are only two requirements for running an experiment with Hydronaut:

• A Hydra YAML configuration file.
• A subclass of the Hydronaut Experiment class, which is defined in hydronaut.experiment.

Once the configuration file and Experiment subclass have been created, the experiment can be run with hydronaut-run or hydronaut-run_in_venv.sh (see below).

Tutorials

A series of Jupyter notebooks has been prepared to serve as an introduction and a tutorial to Hydronaut. The Git repository is available here and the exported slides are available here.

Examples

Examples of varying complexity are provided in the examples directory. The dummy examples provide the simplest albeit least interesting examples of a minimal setup. Peruse the others to get an idea of how to create more interesting experiments.

Hydra Configuration File

By default, Hydronaut expects a Hydra configuration file at the subpath conf/config.yaml relative to the current working directory. A different configuration file can be specified by setting the HYDRONAUT_CONFIG environment variable if necessary. The value of this variable will be interpreted as a subpath within the conf directory of the working directory. For example, export HYDRONAUT_CONFIG=config-test.yaml will load conf/config-test.yaml relative to the current directory.

Hydronaut uses Hydra, which in turn uses OmegaConf configuration files. The Hydra start guide provides a good and quick introduction to the functionality provided by Hydra. The basic idea is that you should set all of your experiment's parameters in the configuration file and then retrieve them from the configuration object in your code. This will grant the following advantages:

• All parameters can be modified in one place without changing the code.
• All parameters can be overridden from the command line.
• The effects of different parameters and parameter combinations can be explored automatically using Hydra's sweeper plugins, including automatic optimization.
• The exact parameters used for each run are stored systematically in structured output files along with all artifacts and metrics that your experiment creates.
• Only a single object needs to be passed around in your code instead of an ever-changing list of parameters.

In addition to the reserved Hydra fields (hydra, defaults), Hydronaut adds an experiment field with some required values:

experiment:
  name: <experiment name>                # required
  description: <experiment description>  # required
  exp_class: <module>:<class>            # required
  params: <experiment parameters>
  python:
    paths: <list of directories to add to the Python system path>
  mlflow: <MLflow configuration>
  environment: <dict of environment variable names and values>

It is strongly recommended that all experiment parameters be nested under experiment.params but this is not enforced programmatically unless experiment/hf_experiment is added to the defaults list in the configuration file.

The best way to get started is to browse the tutorial and peruse the configuration files in the provided examples. hydronaut-init can also be used to initialize a directory for a new experiment. It will create a configuration file and a subclass of the Experiment class that the user can use as a starting point. It accepts some options for initializing the configuration file with settings for sweepers and launchers. See hydronaut-init --help for details.

For further details, consult the Hydra and OmegaConf documentation, e.g.

• command-line flags
• defaults list
• extending configs
• extended override syntax
• override syntax
• tab completion
• variable interpolation

Environment Variable Configuration

Environment variables can be defined in the configuration file under experiment.environment. This should be a dictionary mapping environment variable names (strings) to values (strings). These settings will only take effect after the Hydra configuration file is loaded but before MLflow is initialized and can thus be used to configure MLflow (see below).

experiment:
    # ...
    environment:
        # e.g.
        MLFLOW_TRACKING_URI: https://example.com/mlflow/
        MLFLOW_TRACKING_PASSWORD: gToVTMvhH0C6B6yR
        # ...

MLflow Configuration

Tracking, artifact and registry servers for MLflow are configured via environment variables. These can either be set by the user prior to invoking hydronaut-run or via the configuration file as explained above. The configuration is exactly the same as when using MLflow directly.

In addition to the environment variables supported by MLflow and its eventual dependencies such as Boto3 (required for AWS support), the configuration file can also be used to pass additional arguments to mlflow.start_run via the field experiment.mlflow.start_run. This must be a dictionary mapping valid start_run keyword argument names to their values, e.g.

experiment:
    # ...
    mlflow:
        start_run:
            tags:
                tag1: one
                tag2: two

Standard Environment Variables

The user should consult the relevant documentation for supported environment variables, e.g.

• MLflow Tracking
• Boto3 environment variables for AWS credentials

For example, to set up a remote tracking server with an S3 artifact server backend, the following environment variables may be required:

• MLFLOW_TRACKING_URI
• MLFLOW_TRACKING_USERNAME
• MLFLOW_TRACKING_PASSWORD
• MLFLOW_S3_ENDPOINT_URL
• AWS_ACCESS_KEY_ID
• AWS_SECRET_ACCESS_KEY
• AWS_DEFAULT_REGION
• AWS_SESSION_TOKEN

The specific requirements will depend on the target setup.

Custom Environment Variables

• MLFLOW_REGISTRY_URI: If set, the value will be used to configure the MLflow registry URI, which otherwise defaults to the value of the tracking URI.

Resolvers

Resolvers are functions that can be used to insert values into fields of the configuration file, such as the current date (${now:%Y-%m-%d}) or the number of available CPU cores (${n_cpu:}). Hydronaut provides some custom resolvers in addition to the ones provided by OmegaConf and Hydra.

Resolver	Description
cwd	The original current working directory where Hydronaut was started, as an absolute path. The following optional arguments are recognized: resolve - resolve symlinks in the path; uri - return the path as a file URI (file:///...).
min	Returns the minimum value of its arguments: ${min: ${n_cpu:} ${n_gpu_pytorch}}.
max	Returns the maximum value of its arguments: ${max: ${n_cpu:} ${n_gpu_pytorch}}.
n_gpu_pytorch	The number of available GPUs to PyTorch: ${n_gpu_pytorch:}. If PyTorch is not installed then this will always return 0. This resolver accepts a divisor with the same interpretation as n_cpu.
n_cpu	The number of available logical CPUs: ${n_cpu:}. An optional argument may also be given to divide the number of available CPUs by an integer, which may be useful when assigning CPUs to jobs, e.g. ${n_cpu: 4} or {n_cpu: ${n_jobs}}.
url_quote	Quote the single given argument for use in URLs (e.g. sqlite:///${url_quote:${cwd:}}/database.db).

See hydronaut.hydra.resolvers for implementation details and comments.

Experiment Subclasses

Hydronaut will invoke a subclass of hydronaut.experiment.Experiment as an entry point to the user's code. The only thing that the user must define in their subclass is the __call__() method (see below). It is entirely up to the user what this function does as only the return value will be used by Hydronaut. The user can invoke other Python code, external commands, remote services, etc.

The configuration object will be available through the parent class's config attribute, from which the user should retrieve the hyperparameters defined in the configuration file.

Here are the main methods of interest to users when subclassing the Experiment class:

• __call__(self) (required) - This method accepts no arguments and returns 1 or more numerical values which should represent the target score or loss of the experiment. The return value will be tracked in MLflow under the name "Objective Value" and it will be the value optimized by any configured optimizers. When returning multiple values, the optimizer must be configured accordingly. For example, see the Optuna plugin's Multi-Objective Optimization example.

• setup(self) (optional) - This method accepts no arguments and returns None. It should be used to prepare everything needed before invoking __call__(), such as downloading and preparing datasets, initializing remote resources, etc.

• __init__(self, config) (optional) - This method accepts the Hydra/OmegaConf configuration object and returns None. If overridden, it should invoke super().__init__(config) to ensure proper setup.

The parent Experiment class also provides several MLflow logging methods for convenience. These are documented here.

Hydronaut will determine which Experiment subclass to load via the configuration file's experiment.exp_class field. The value of this field is a string with the format <module>:<class> where <module> is the Python module containing the subclass and <class> is its name. Python modules and packages that are not already on the system path can be made importable by adding their containing directories to the experiment.python.paths list in the configuration file.

See the dummy example for an example of a very simple configuration file and Experiment subclass with only one module. The other examples and the tutorial will demonstrate real-life examples.

Commands

The following commands are installed with the Python package.

hydronaut-run

Run hydronaut-run (equivalent to python -m hydronaut.run) in the working directory to load the configuration file and run the experiment. After the script has started, run mlflow ui in the same directory and then open the URL that it shows in a web browser. All of the experiments results will appear under the name given to the experiment in the configuration file.

hydronaut-run accepts all of Hydra's command-line flags. For example, to show Hydra information, run hydronaut-run --info.

# Usage:
hydronaut-run [<hydra arguments>]

# For example:
hydronaut-run --cfg job
hydronaut-run --multirun experiment.params.foo=42

hydronaut-init

Hydronaut also provides a script named hydronaut-init which will generate a commented configuration file and an Experiment subclass skeleton under the current working directory which can be used as a starting point for a new experiment.

See hydronaut-init --help for available options.

Decorators

As an alternative to subclassing hydronaut.experiment.Experiment and running an experiment with hydronaut-run, Hydronaut also provides the with_hydronaut decorator that can be used to run classes and functions with Hydronaut directly.

The decorator accepts a relative or absolute path to a configuration file as an optional argument (see the examples linked below).

• If the path is relative, it is interpreted relative to the working directory.
• The parent directory of the configuration file will be treated as the Hydra configuration directory, i.e. the directory in which Hydra will look for overrides.
• If the argument is omitted, the default path is used.

Decorated Functions

The decorator can be used with functions that accept a Hydra configuration object as their sole argument and return one or more numerical values that will serve as objective values. The decorated function is a callable object that will invoke the given function within the Hydronaut framework. The user should invoke this object as their main function.

See the dummy decorated function example for details.

Decorated Classes

The decorator can be used with classes that define __init__, __call__ and optionally setup methods with the same function signatures as hydronaut.experiment.Experiment. The decorated class is a callable object that will invoke the setup (if defined) and __call__ methods of the given class after instantiating it with the configuration object within the Hydronaut framework. The user should invoke this object as their main function.

See the dummy decorated class example for details.

Remarks

• Decorated functions and classes should be run directly by the user as shown in the examples. They are not compatible with hydronaut-run.
• Subclasses of hydronaut.experiment.Experiment are compatible with the decorator and provide wrappers around the MLflow logging functions that output log messages when they are called. It is therefore recommended to subclass the Experiment class even when using decorators.

API Documentation

The Sphinx-generated online API documentation is available here: https://jrye.gitlabpages.inria.fr/hydronaut/.

Scripts

The following convenience scripts are provided in the source repository for common operations.

hydronaut-initialize.sh

hydronaut-initialize.sh is just a convenience script for recursively checking out the submodules. It may be extended later.

# Usage:
hydronaut-initialize.sh

hydronaut-install_in_venv.sh

hydronaut-install_in_venv.sh will install Hydronaut in a virtual environment. If the virtual environment does not exist then it will be created.

See hydronaut-install_in_venv.sh -h for details.

hydronaut-lint.sh

hydronaut-lint.sh will report warnings and errors in the Hydronaut source files and examples.

# Usage:
hydronaut-lint.sh

hydronaut-rsync-results.sh

hydronaut-rsync-results.sh will transfer files that result from an experiment in the source directory to a destination using rsync. It can be used to retrieve results from a cluster, for example.

# Usage:
hydronaut-rsync-results.sh [<rsync options>] <source> <destination>

hydronaut-run_in_venv.sh

hydronaut-run_in_venv.sh is a wrapper around hydronaut-run that will first set up a virtual environment and install Hydronaut before running hydronaut-run in the current directory. If the current directory contains a pyproject.toml or requirements.txt file, they will also be installed in the virtual environment before running hydronaut-run.

See hydronaut-run_in_venv.sh -h for details.

Optuna

Usage of the Optuna Sweeper plugin is documented here. By default, Optuna studies are kept in memory and thus only last for the duration of a single program execution. This works across runs in the same execution but not between separate executions. You may resume a study on subsequent executions by using persistent storage such as an SQLite backend. This can be achieved as follows:

hydra:
  sweeper:
    # Use a common database for all runs.
    storage: sqlite:///${url_quote:${cwd:}}/optuna.db

    # Set the study to the experiment name.
    study_name: ${experiment.name}

    # ...

Please read the Optuna SQLite Storage subsection under Troubleshooting below for information regarding pending pull requests and version incompatibilities.

When using Optuna >= 3.0.0, it is possible to use Optuna Dashboard to follow optimization in real time.

MLflow

Hydronaut only uses a fraction of the full functionality of MLflow. For example, beyond tracking your experiments, it can also be used to package and distribute your code via standard platforms to facilitate collaboration with others. The interested user should consult the MLflow documentation to get an idea of what is possible.

Most of the MLflow functionality is available via the command-line interface (mlflow). For additional functionality the user may also be interested in MLflow Extra.

Useful Examples

# Find an experiment ID.
mlflow experiments search

# Export all of its parameters and metrics to a CSV file
mlflow experiments csv -x <ID>

GitLab Integration

GitLab version 15.11 introduced functionality to use GitLab as an MLflow server. The configuration only requires a GitLab access token and two environment variables and it will let you log all results obtained via Hydronaut to a GitLab project of your choice. Please refer to the official GitLab documentation for details.

GitLab does not yet fully support all MLflow client methods but these are expected to be implemented as development continues. See the "GitLab MLflow Server" subsection under "Troubleshooting" below for a temporary workaround.

Troubleshooting

CUDA/NVCC Errors With PyTorch

Check that the version of PyTorch is compatible with the one installed on the system:

# System CUDA version
nvcc -V

# PyTorch CUDA version
python -c 'import torch; print(torch.version.cuda)'

If the versions are mismatched, consult the PyTorch Start Locally guide for commands to install a compatible version in a virtual environment. If PyTorch is already installed in the current (virtual) environment, append --upgrade to the install command (e.g. pip install --upgrade ...).

Hydra Joblib Launcher Plugin And PyTorch DataLoader Threads

The Hydra Joblib Launcher plugin is currently not compatible with PyTorch DataLoader worker processes. Either disable the Joblib launcher or set the number of worker processes to 0 when instantiating DataLoaders (num_workers=0).

Hydra Configuration In Subprocesses

Due to limitations in Hydra's support for subprocesses, it is usually necessary to re-initialize the Hydra configuration in subprocesses such as PyTorch DataLoaders running as separate workers. Hydronaut provides the configure_hydra function to facilitate this.

from hydronaut.hydra.config import configure_hydra
# ...
configure_hydra(from_env=True)

The configuration uses environment variables that are set during initialization of the Experiment base class.

GitLab MLflow Server

At the time of writing, GitLab has not yet implemented full support for all MLflow client methods. In particular, artifact logging does not currently support subdirectories.

The following code can be used to override with MLflow client's artifact logging with versions that transform directory paths into flat file names. This code can be pasted in before the Hydronaut-decorated main function in user code or at the top of the hydronaut.run module.

import os
import pathlib
import tempfile
from mlflow.tracking.client import MlflowClient

def log_artifact(self, run_id, local_path, artifact_path=None):
    '''
    Override MlflowClient.log_artifact to avoid creating directories on
    endpoints that do not support them.
    '''
    separator = '__'
    if artifact_path is not None:
        artifact_path = pathlib.Path(artifact_path)
        local_path = pathlib.Path(local_path).resolve()
        name = f'{separator.join(artifact_path.parts)}{separator}{local_path.name}'
        with tempfile.TemporaryDirectory() as tmp_dir:
            tmp_path = pathlib.Path(tmp_dir) / name
            tmp_path.symlink_to(local_path)
            self._tracking_client.log_artifact(run_id, str(tmp_path), None)
            return
    self._tracking_client.log_artifact(run_id, local_path, None)


def log_artifacts(self, run_id, local_dir, artifact_path=None):
    '''
    Override MlflowClient.log_artifacts to avoid creating directories on
    endpoints that do not support them.
    '''
    if artifact_path is not None:
        artifact_path = pathlib.Path(artifact_path)

    local_dir = pathlib.Path(local_dir).resolve()
    # local_dir.walk() will be available in Python 3.12
    for root, _dirs, files in os.walk(local_dir):
        rel_root = pathlib.Path(root).relative_to(local_dir.parent)
        for fil in files:
            path = root / fil
            new_artifact_path = artifact_path / rel_root
            self.log_artifact(run_id, path, new_artifact_path)


MlflowClient.log_artifact = log_artifact
MlflowClient.log_artifacts = log_artifacts

If the need for this workaround persists, it will be integrated as an option in Hydronaut.

Optuna SQLite Storage

Hydra's current Optuna Sweeper plugin depends on Optuna 2.10.1, which is incompatible with SQLAlchemy version 2.0 and above. There is an open pull request on Hydra's GitHub to update the plugin for Optuna version >=3.0.0. Until this pull request is merged, you can work around the issue by either downgrading SQLAlchemy:

pip install --force-reinstall SQLAlchemy==1.4.44

or by installing the patched version submitted with the pull request that updates the sweeper to Optuna >= 3.0.0.

pip install 'git+https://github.com/keisuke-umezawa/hydra/@feature/fix-optuna-v3#egg=hydra-optuna-sweeper&subdirectory=plugins/hydra_optuna_sweeper'

YAML Errors When Sweeping Strings

Attempting to sweep strings from the configuration file, for example with

hydra:
  sweeper:
    params:
      ++experiment.params.foo: "a b", "c d", "e f"

will result in YAML block-parsing errors:

yaml.parser.ParserError: while parsing a block mapping
  in ...
expected <block end>, but found ','
  in ...

This occurs due to a limitation of the YAML parser used by Hydra, which expects the rest of the line to be a string value when the value starts with a quotation mark. Either remove the quotes from the value and use "\" to escape special characters and spaces (e.g. ++experiment.params.foo: a\ b, c\ d, e\ f), or use the choice function (e.g. ++experiment.params.foo: choice("a b", "c d", "e f")). The choice sweep function is documented here.

Further Reading

Optuna

• Optuna website
• Hydra Optuna Sweeper plugin documentation
• Multi-objective Optimization with Optuna

Optuna Dashboard

• Optuna website
• GitHub
• PyPI

PyTorch Lighting

• Introduction
• Callbacks