Yet Another Lightning Hydra Template for ML Experiments

Efficient workflow and reproducibility are crucially important components in every machine learning project, which enables to:

• Rapidly iterate over new models and compare different approaches faster.
• Promote confidence in the results and transparency.
• Save time and resources.

PyTorch Lightning and Hydra serve as the foundation of this template. Such reasonable technology stack for deep learning prototyping provides a comprehensive and seamless solution, allowing you to effortlessly explore different tasks across a variety of hardware accelerators such as CPUs, multi-GPUs, and TPUs. Furthermore, it includes a curated collection of best practices and extensive documentation for greater clarity and comprehension.

This template can be used as is for some basic tasks like Classification, Segmentation, or Metric Learning, or be easily extended for any other tasks due to high-level modularity and scalable structure.

As a baseline, I have used the gorgeous Lightning Hydra Template, reshaped and polished it, and implemented more features that can improve the overall efficiency of workflow and reproducibility.

Table of content

• Main technologies
• Project structure
• Workflow - how it works
  • Basic workflow
  • LightningDataModule
  • LightningModule
  • Training loop
  • Evaluation and prediction loops
  • Callbacks
• Logs
• Data
• Hyperparameters search
• Docker
• Tests
• Continuous integration

Main technologies

PyTorch Lightning - a lightweight deep learning framework / PyTorch wrapper for professional AI researchers and machine learning engineers who need maximal flexibility without sacrificing performance at scale.

Hydra - a framework that simplifies configuring complex applications. The key feature is the ability to dynamically create a hierarchical configuration by composition and override it through config files and the command line.

Project structure

The machine learning project structure may differ depending on the specific requirements and goals of the project, as well as the tools and frameworks being used. However, this is a typical directory structure of machine learning project:

• src/
• data/
• logs/
• tests/
• some additional directories, like notebooks/, docs/, etc.

In this particular case, the directory structure looks like this:

├── configs                     <- Hydra configuration files
│   ├── callbacks               <- Callbacks configs
│   ├── datamodule              <- Datamodule configs
│   ├── debug                   <- Debugging configs
│   ├── experiment              <- Experiment configs
│   ├── extras                  <- Extra utilities configs
│   ├── hparams_search          <- Hyperparameter search configs
│   ├── hydra                   <- Hydra settings configs
│   ├── local                   <- Local configs
│   ├── logger                  <- Logger configs
│   ├── module                  <- Module configs
│   ├── paths                   <- Project paths configs
│   ├── trainer                 <- Trainer configs
│   │
│   ├── eval.yaml               <- Main config for evaluation
│   └── train.yaml              <- Main config for training
│
├── data                        <- Project data
├── logs                        <- Generated logs
├── notebooks                   <- Jupyter notebooks
├── scripts                     <- Shell scripts
│
├── src                         <- Source code
│   ├── callbacks               <- Additional callbacks
│   ├── datamodules             <- Lightning datamodules
│   ├── modules                 <- Lightning modules
│   ├── utils                   <- Utility scripts
│   │
│   ├── eval.py                 <- Run evaluation
│   └── train.py                <- Run training
│
├── tests                       <- Tests of any kind
│
├── .dockerignore               <- List of files ignored by docker
├── .gitattributes              <- List of attributes to pathnames
├── .gitignore                  <- List of files ignored by git
├── .pre-commit-config.yaml     <- Configuration of pre-commit hooks
├── Dockerfile                  <- Dockerfile
├── Makefile                    <- Makefile
├── pyproject.toml              <- Config for testing and linting
├── requirements.txt            <- Python dependencies
├── setup.py                    <- Setup file
└── README.md

Workflow - how it works

Before starting a project, you should consider the following aspects to ensure the reproducibility of results:

• Docker image
• Freezing python package versions
• Git
• Data version control. Many of these currently provide not just Data Version Control, but a lot of side highly useful features like Model Registry or Experiments Tracking:

  • DVC

  • Neptune

  • Your own solution or others…

    
    

• Experiments Tracking tools:
  • Weights & Biases
  • Neptune
  • DVC
  • Comet
  • MLFlow
  • TensorBoard
  • Or just CSV files…

Basic workflow

This template could be used as is for some basic tasks like Classification, Segmentation, or Metric Learning approach, but if you need to do something more complex, here is a general workflow:

1. Write your PyTorch Lightning Module (see examples in src/modules/single_module.py)

2. Write your PyTorch Lightning DataModule (see examples in src/datamodules/datamodules.py)

3. Fill up your configs, particularly create experiment configs

4. Run experiments:

   • Run training with chosen experiment config:

   python src/train.py experiment=experiment_name.yaml
   

   • Use hyperparameter search, for example by Optuna Sweeper via Hydra:

   # using Hydra multirun mode
   python src/train.py -m hparams_search=mnist_optuna
   

   • Execute the runs with some config parameter manually:

   python src/train.py -m logger=csv module.optimizer.weight_decay=0.0,0.00001,0.0001
   

   
   

5. Run evaluation with different checkpoints or run prediction on a custom dataset for additional analysis

The template contains an example with MNIST classification, which uses for tests by the way. If you run python src/train.py, you will get something like this: Show terminal screen when running pipeline in the template documentation.

LightningDataModule

At the start, you need to create PyTorch Dataset for your task. It has to include __getitem__ and __len__ methods. Maybe you can use as is or easily modify already implemented datasets in the template. See more details in PyTorch documentation.

Also, it could be useful to see a data section about how it is possible to save data for training and evaluation.

Then, you need to create DataModule using PyTorch Lightning DataModule API. By default, API has the following methods:

• prepare_data (optional): perform data operations on CPU via a single process, like load and preprocess data, etc.
• setup (optional): perform data operations on every GPU, like train/val/test splits, create datasets, etc.
• train_dataloader: used to generate the training dataloader(s)
• val_dataloader: used to generate the validation dataloader(s)
• test_dataloader: used to generate the test dataloader(s)
• predict_dataloader (optional): used to generate the prediction dataloader(s)

See examples of datamodule configs in configs/datamodule folder.

Show LightningDataModule API in the template documentation.

By default, the template contains the following DataModules:

• SingleDataModule in which train_dataloader, val_dataloader and test_dataloader return single DataLoader, predict_dataloader returns list of DataLoaders
• MultipleDataModule in which train_dataloader return dict of DataLoaders, val_dataloader, test_dataloader and predict_dataloader return list of DataLoaders

In the template, DataModules has _get_dataset_ method to simplify datasets instantiation.

LightningModule

LightningModule API

Next, your need to create LightningModule using PyTorch Lightning LightningModule API. Minimum API has the following methods:

• forward: use for inference only (separate from training_step)
• training_step: the complete training loop
• validation_step: the complete validation loop
• test_step: the complete test loop
• predict_step: the complete prediction loop
• configure_optimizers: define optimizers and LR schedulers

Also, you can override optional methods for each step to perform additional logic:

• training_step_end: training step end operations
• training_epoch_end: training epoch end operations
• validation_step_end: validation step end operations
• validation_epoch_end: validation epoch end operations
• test_step_end: test step end operations
• test_epoch_end: test epoch end operations

Show LightningModule API methods and appropriate order in the template documentation.

In the template, LightningModule has model_step method to adjust repeated operations, like forward or loss calculation, which are required in training_step, validation_step and test_step.

Metrics

The template offers the following Metrics API:

• main metric: main metric, which also uses for all callbacks or trackers like model_checkpoint, early_stopping or scheduler.monitor.
• valid_best metric: used for tracking the best validation metric. Usually, it can be MaxMetric or MinMetric.
• additional metrics: some additional metrics.

Each metric config should contain _target_ key with the metric class name and other parameters, which are required by the metric. The template allows to use any metrics, for example from torchmetrics or implemented by yourself. See more details about  torchmetrics API, implemented Metrics API and metrics config as a part of network configs in configs/module/network folder.

Metric config example:

metrics:
  main:
    _target_: "torchmetrics.Accuracy"
    task: "binary"
  valid_best:
    _target_: "torchmetrics.MaxMetric"
  additional:
    AUROC:
      _target_: "torchmetrics.AUROC"
      task: "binary"

Loss

The template suggests the following Losses API:

• Loss config should contain _target_ key with the loss class name and other parameters required
• Parameter containing weight string in name will be wrapped by torch.tensor and cast to torch.float type before passing to loss due to requirements from most of the losses.

The template allows you to use any losses, for example from PyTorch or implemented by yourself. See more details about implemented Losses API and loss config as a part of network configs in configs/module/network folder.

Loss config examples:

loss:
  _target_: "torch.nn.CrossEntropyLoss"

loss:
  _target_: "torch.nn.BCEWithLogitsLoss"
  pos_weight: [0.25]

loss:
  _target_: "src.modules.losses.VicRegLoss"
  sim_loss_weight: 25.0
  var_loss_weight: 25.0
  cov_loss_weight: 1.0

Also, the template includes few manually implemented losses:

• VicRegLoss as example for self-supervised learning
• FocalLoss: use for extremely imbalanced tasks
• AngularPenaltySMLoss: use for Metric Learning approach

Model

The template offers the following Model API, model config should contain:

• _target_: key with the model class name
• model_name: model name
• model_repo (optional): model repository
• Other parameters required by a model

By default, a model can be loaded from:

• torchvision.models with setting up model_name as torchvision.models/<model-name>, for example torchvision.models/mobilenet_v3_large
• segmentation_models_pytorch with setting up model_name as segmentation_models_pytorch/<model-name>, for example segmentation_models_pytorch/Unet
• timm with setting up model_name as timm/<model-name>, for example timm/mobilenetv3_100
• torch.hub with setting up model_name as torch.hub/<model-name> and model_repo, for example model_name="torch.hub/resnet18" and model_repo="pytorch/vision"

See more details about implemented Model API and model config as a part of network configs in configs/module/network folder.

Model config example:

model:
  _target_: "src.modules.models.classification.Classifier"
  model_name: "torchvision.models/mobilenet_v3_large"
  model_repo: null
  weights: "IMAGENET1K_V2"
  num_classes: 1

Implemented LightningModules

By default, the template comes with the following LightningModules:

• SingleLitModule contains LightningModules for a few tasks, like common, self-supervised learning and metric learning approach, which require a single DataLoader on each step
• MultipleLitModule contains LightningModules, which require multiple DataLoaders on each step

See examples of module configs in configs/module folder. Some LightningModule config example:

_target_: src.modules.single_module.MNISTLitModule

defaults:
  - _self_
  - network: mnist.yaml

optimizer:
  _target_: torch.optim.Adam
  lr: 0.001
  weight_decay: 0.0

scheduler:
  scheduler:
    _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
    mode: "max"
    factor: 0.1
    min_lr: 1.0e-9
    patience: 10
    verbose: True
  extras:
    monitor: ${replace:"__metric__/valid"}
    interval: "epoch"
    frequency: 1

logging:
  on_step: False
  on_epoch: True
  sync_dist: False
  prog_bar: True

Training loop

Training loop in the template consists of the following stages:

• LightningDataModule instantiating
• LightningModule instantiating
• Callbacks instantiating
• Loggers instantiating
• Plugins instantiating
• Trainer instantiating
• Hyperparameters and metadata logging
• Training the model
• Testing the best model

See more details in training loop and configs/train.yaml.

Evaluation and prediction loops

Evaluation loop in the template consists of the following stages:

• LightningDataModule instantiating
• LightningModule instantiating
• Loggers instantiating
• Trainer instantiating
• Hyperparameteres and metadata logging
• Evaluating model or predicting

See more details in evaluation loop and configs/eval.yaml.

The template contains the following Prediction API:

• Set predict: True in configs/eval.yaml to turn on prediction mode.
• DataModule could contain multiple predict datasets:

datasets:
  predict:
    dataset1:
      _target_: src.datamodules.datasets.ClassificationDataset
      json_path: ${paths.data_dir}/predict/data1.json
    dataset2:
      _target_: src.datamodules.datasets.ClassificationDataset
      json_path: ${paths.data_dir}/predict/data2.json

• PyTorch Lightning returns a list of batch predictions, when LightningDataModule.predict_dataloader() returns a single dataloader, and a list of lists of batch predictions, when LightningDataModule.predict_dataloader() returns multiple dataloaders.
• Predictions log to {cfg.paths.output_dir}/predictions/ folder.
• If there are multiple predict dataloaders, predictions will be saved with _<dataloader_idx> postfix. It isn’t possible to use dataset names due to PyTorch Lightning doesn’t allow to return a dict of dataloaders from LightningDataModule.predict_dataloader() method.
• There are two possible built-in output formats: csv and json. json format is used by default, but it might be more effective to use csv format for a large number of predictions, it may help to avoid RAM memory overflow, because csv allows writing row by row and doesn’t require keeping in RAM the whole dict like in case of json. To change the output format, set predictions_saving_params.output_format variable in configs/extra/default.yaml config file.
• If you need some custom output format, for instance, parquet, you can easily modify src.utils.saving_utils.save_predictions() method.

See more details about Prediction API and predict_step in LightningModule.

Callbacks

PyTorch Lightning has a lot of built-in callbacks, which can be used just by adding them to the callbacks config, thanks to Hydra. See examples in callbacks config folder.

By default, the template contains a few of them:

• Model Checkpoint
• Early Stopping
• Model Summary
• Rich Progress Bar

However, there is an additional LightProgressBar callback, which might be more elegant and useful, instead of using RichProgressbar:

Logs

Hydra creates new output directory in logs/ for every executed run.

Furthermore, template offers to save additional metadata for better reproducibility and debugging, including:

• pip logs
• git logs
• environment logs: CPU, GPU (nvidia-smi)
• full copy of src/ and configs/ directories

Default logging structure:

├── logs
│   ├── task_name
│   │   ├── runs                        <- Logs generated by runs
│   │   │   ├── YYYY-MM-DD_HH-MM-SS     <- Datetime of the run
│   │   │   │   ├── .hydra              <- Hydra logs
│   │   │   │   ├── csv                 <- Csv logs
│   │   │   │   ├── wandb               <- Weights & Biases logs
│   │   │   │   ├── checkpoints         <- Training checkpoints
│   │   │   │   ├── metadata            <- Metadata
│   │   │   │   │   ├── pip.log         <- Pip logs
│   │   │   │   │   ├── git.log         <- Git logs
│   │   │   │   │   ├── env.log         <- Environment logs
│   │   │   │   │   ├── src             <- Full copy of `src/`
│   │   │   │   │   └── configs         <- Full copy of `configs/`
│   │   │   │   └── ...                 <- Any other saved files
│   │   │   └── ...
│   │   │
│   │   └── multiruns                   <- Logs generated by multiruns
│   │       ├── YYYY-MM-DD_HH-MM-SS     <- Datetime of the multirun
│   │       │   ├──1                    <- Multirun job number
│   │       │   ├──2
│   │       │   └── ...
│   │       └── ...
│   │
│   └── debugs                          <- Logs generated during debug
│       └── ...

Data

Usually, images or any other data files just stored on disk in folders. It is a simple and convenient way.

However, there are other methods and one of them calls as Hierarchical Data Format HDF5 or h5py, which has a few reasons why it might be more beneficial to store images in HDF5 files instead of just folders:

• Efficient storage: the data format is designed specifically for storing large amounts of data. It is particularly well-suited for storing arrays of data, like images, and can compress the data to reduce the overall size of the file. The important thing about compressing in HDF5 files is that objects are compressed independently and only the objects that you need get decompressed on output. This is clearly more efficient than compressing the entire file and having to decompress the entire file to read it.
• Fast access: HDF5 allows you to access the data stored in the file using indexing, just like you would with a NumPy array. This makes it easy and fast to retrieve the data you need, which can be especially important when you are working with large datasets.
• Easy to use: HDF5 is easy to use and integrates well with other tools commonly used in machine learning, such as NumPy and PyTorch. This means you can use HDF5 to store your data and then load it into your training code without any additional preprocessing.
• Self-describing: it is possible to add information that helps users and tools know what is in the file. What are the variables, what are their types, what tools collected and wrote them, etc. The tool you are working on can read metadata for files. Attributes in an HDF5 file can be attached to any object in the file – they are not just file level information.

This template contains a tool which might be used to easily create and read HDF5 files.

To create HDF5 file:

from src.datamodules.components.h5_file import H5PyFile

H5PyFile().create(
    filename="/path/to/dataset_train_set_v1.h5",
    content=["/path/to/image_0.png", "/path/to/image_1.png", ...],
    # each content item loads as np.fromfile(filepath, dtype=np.uint8)
)

To read HDF5 file in the wild:

import matplotlib.pyplot as plt
from src.datamodules.components.h5_file import H5PyFile

h5py_file = H5PyFile(filename="/path/to/dataset_train_set_v1.h5")
image = h5py_file[0]

plt.imshow(image)

To read HDF5 file in Dataset.__getitem__:

def __getitem__(self, index: int) -> Any:
    key = self.keys[index]  # get the image key, e.g. path
    data_file = self.data_file
    source = data_file[key]  # get the image
    image = io.BytesIO(source)  # read the image
    ...

Hyperparameters search

Hydra provides out-of-the-box hyperparameters sweepers: Optuna, Nevergrad or Ax.

You may define hyperparameters search by adding new config file to configs/hparams_search.

See example of hyperparameters search config. With this method, there is no need to add extra code, everything is specified in a single configuration file. The only requirement is to return the optimized metric value from the launch file.

Execute it with:

python src/train.py -m hparams_search=mnist_optuna

The optimization_results.yaml will be available under logs/task_name/multirun folder.

Docker

Docker is an essential part of environment reproducibility that makes it possible to easily package a machine learning pipeline and its dependencies into a single container that can be easily deployed and run on any environment. This is particularly useful due to it helps to ensure that the code will run consistently, regardless of the environment in which it is deployed.

Docker image could require some additional packages depends on which device is used for running. For example, for running on cluster with NVIDIA GPUs it requires the CUDA Toolkit from NVIDIA. The CUDA Toolkit provides everything you need to develop GPU-accelerated applications, including GPU-accelerated libraries, a compiler, development tools and the CUDA runtime.

In general, there are many way how to set up it, but to simplify this process you can use:

• Official Nvidia Docker Images Hub, where it is easy to find images with any combinations of OS, CUDA, etc. See possible structure of Dockerfile here.
• Miniconda for GPU environments.

Moreover, it can be advantageous to use:

• Additional docker container runtime options for managing resources constraints, like -cpuset-cpus, -gpus, etc.
• NVTOP - a (h)top like task monitor for AMD, Intel and NVIDIA GPUs.

Here it is some example of container running based on proposed Dockerfile and .dockerignore:

set -o errexit
export DOCKER_BUILDKIT=1
export PROGRESS_NO_TRUNC=1

docker build --tag <project-name> \
    --build-arg OS_VERSION="22.04" \
    --build-arg CUDA_VERSION="11.7.0" \
    --build-arg PYTHON_VERSION="3.10" \
    --build-arg USER_ID=$(id -u) \
    --build-arg GROUP_ID=$(id -g) \
    --build-arg NAME="<your-name>" \
    --build-arg WORKDIR_PATH=$(pwd) .

docker run \
    --name <task-name> \
    --rm \
    -u $(id -u):$(id -g) \
    -v $(pwd):$(pwd):rw \
    --gpus '"device=0,1,3,4"' \
    --cpuset-cpus "0-47" \
    -it \
    --entrypoint /bin/bash \
    <project-name>:latest

Tests

Tests are an important aspect of software development in general, and especially in Machine Learning, because here it can be much more difficult to understand if code are working correctly without testing. Consequently, template contains some generic tests implemented with pytest.

For this purpose MNIST is used. It is a small dataset, so it is possible to run all tests on CPU. However, it is easy to implement tests for your own dataset if it requires.

As a baseline the tests cover:

• Main module configs instantiation by Hydra
• DataModule
• Losses loading
• Metrics loading
• Models loading and utils
• Training on 1% of MNIST dataset, for example:
  • running 1 train, val and test steps
  • running 1 epoch, saving checkpoint and resuming for the second epoch
  • running 2 epochs with DDP simulated on CPU
• Evaluating and predicting
• Hyperparameters optimization
• Custom progress bar functionality
• Utils

All this implemented tests created for verifying that the main pipeline modules and utils are executable and working as expected However, sometimes it couldn’t be enough to ensure that the code is working correctly, especially in case of more complex pipelines and models.

For running:

# run all tests
pytest

# run tests from specific file
pytest tests/test_train.py

# run tests from specific test
pytest tests/test_train.py::test_train_ddp_sim

# run all tests except the ones marked as slow
pytest -k "not slow"

Continuous integration

The template contains a few initial CI workflows via the GitHub Actions platform. It makes it easy to automate and streamline development workflows, which can help to save time and effort, increase efficiency, and improve overall quality of the code. In particularly, it includes:

• .github/workflows/test.yaml: running all tests from tests/ with pytest on Linux, Mac and Windows platforms
• .github/workflows/code-quality-main.yaml: running pre-commits on main branch for all files
• .github/workflows/code-quality-pr.yaml: running pre-commits on pull requests for modified files only

Note: You need to enable the GitHub Actions from the settings in your repository.

See more about GitHub Actions for CI.

In the case of using GitLab, it is easy to set up GitLab CI based on GitHub Actions workflows. Here it manages by .gitlab-ci.yml file. See more here.