Efficient workflow and reproducibility are extremely important components in every machine learning projects, which enable 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 upon this template. Such reasonable technology stack for deep learning prototyping offers 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 could 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 gorgeous Lightning Hydra Template, reshaped and polished it, and implemented more features which can improve overall efficiency of workflow and reproducibility.

# clone template
git clone https://github.com/gorodnitskiy/yet-another-lightning-hydra-template
cd yet-another-lightning-hydra-template

# install requirements
pip install -r requirements.txt

Or run the project in docker. See more in Docker section.

• Main technologies
• Project structure
• Workflow - how it works
  • Basic workflow
  • LightningDataModule
  • LightningModule
  • Training loop
  • Evaluation and prediction loops
  • Callbacks
  • Extensions
• Hydra configs
  • How to run pipeline with Hydra
  • Instantiating objects with Hydra
  • Command line operations
  • Additional out-of-the-box features
  • Custom config resolvers
  • Simplify complex modules configuring
• Logs
• Data
• Notebooks
• Hyperparameters search
• Docker
• Tests
• Continuous integration

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.

The structure of a machine learning project can vary depending on the specific requirements and goals of the project, as well as the tools and frameworks being used. However, here is a general outline of a common directory structure for a machine learning project:

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

In this particular case, the directory structure looks like:

├── 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                        <- Logs generated by hydra, lightning loggers, etc.
├── 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 additional attributes to pathnames
├── .gitignore                  <- List of files ignored by git
├── .pre-commit-config.yaml     <- Configuration of pre-commit hooks for code formatting
├── Dockerfile                  <- Dockerfile
├── Makefile                    <- Makefile with commands like `make train` or `make test`
├── pyproject.toml              <- Configuration options for testing and linting
├── requirements.txt            <- File for installing python dependencies
├── setup.py                    <- File for installing project as a package
└── README.md

Before starting a project, you need to think about the following things to unsure in results reproducibility:

• Docker image setting up
• Freezing python package versions
• Code Version Control
• Data Version Control. Many of which currently provide not just Data Version Control, but a lot of side very 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...

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 it is a general workflow:

1. Write your PyTorch Lightning DataModule (see examples in datamodules/datamodules.py)
2. Write your PyTorch Lightning Module (see examples in modules/single_module.py)
3. Fill up your configs, in 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.,0.00001,0.0001

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

The template contains example with MNIST classification, which uses for tests by the way. If you run python src/train.py, you will get something like this:

At the start, you need to create PyTorch Dataset for you 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 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)

from typing import Any, Dict, List, Optional, Union
from torch.utils.data import DataLoader, Dataset
from pytorch_lightning import LightningDataModule


class YourDataModule(LightningDataModule):
    def __init__(self, *args: Any, **kwargs: Any) -> None:
        super().__init__()
        self.train_set: Optional[Dataset] = None
        self.valid_set: Optional[Dataset] = None
        self.test_set: Optional[Dataset] = None
        self.predict_set: Optional[Dataset] = None
        ...

    def prepare_data(self) -> None:
        # (Optional) Perform data operations on CPU via a single process
        # - load data
        # - preprocess data
        # - etc.
        ...

    def setup(self, stage: str) -> None:
        # (Optional) Perform data operations on every GPU:
        # - count number of classes
        # - build vocabulary
        # - perform train/val/test splits
        # - create datasets
        # - apply transforms (which defined explicitly in your datamodule)
        # - etc.
        if not self.train_set and not self.valid_set and not self.test_set:
            self.train_set = ...
            self.valid_set = ...
            self.test_set = ...
        if (stage == "predict") and not self.predict_set:
            self.predict_set = ...

    def train_dataloader(self) -> Union[DataLoader, List[DataLoader], Dict[str, DataLoader]]:
        # Used to generate the training dataloader(s)
        # This is the dataloader that the Trainer `fit()` method uses
        return DataLoader(self.train_set, ...)

    def val_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
        # Used to generate the validation dataloader(s)
        # This is the dataloader that the Trainer `fit()` and `validate()` methods uses
        return DataLoader(self.valid_set, ...)

    def test_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
        # Used to generate the test dataloader(s)
        # This is the dataloader that the Trainer `test()` method uses
        return DataLoader(self.test_set, ...)

    def predict_dataloader(self) -> Union[DataLoader, List[DataLoader]]:
        # Used to generate the prediction dataloader(s)
        # This is the dataloader that the Trainer `predict()` method uses
        return DataLoader(self.predict_set, ...)

    def teardown(self, stage: str) -> None:
        # Used to clean-up when the run is finished
        ...

See examples of datamodule configs in configs/datamodule folder.

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.

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

from typing import Any
from pytorch_lightning import LightningModule


class LitModel(LightningModule):
    def __init__(self, *args: Any, **kwargs: Any):
        super().__init__()
        ...

    def forward(self, *args: Any, **kwargs: Any):
        ...

    def training_step(self, *args: Any, **kwargs: Any):
        ...

    def training_step_end(self, step_output: Any):
        ...

    def training_epoch_end(self, outputs: Any):
        ...

    def validation_step(self, *args: Any, **kwargs: Any):
        ...

    def validation_step_end(self, step_output: Any):
        ...

    def validation_epoch_end(self, outputs: Any):
        ...

    def test_step(self, *args: Any, **kwargs: Any):
        ...

    def test_step_end(self, step_output: Any):
        ...

    def test_epoch_end(self, outputs: Any):
        ...

    def configure_optimizers(self):
        ...

    def any_extra_hook(self, *args: Any, **kwargs: Any):
        ...

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.

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: use for tracking the best validation metric. Usually it can be MaxMetric or MinMetric.
• additional metrics: additional metrics.

Each metric config should contain _target_ key with metric class name and other parameters which are required by metric. The template allows to use any metrics, for example from torchmetrics or implemented by yourself (see examples in modules/metrics/components/ or torchmetrics API).

See more details about 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"

Also, the template includes few manually implemented metrics:

• Accuracy
• NDCG
• MRR
• SentiMRR
• PrecisionAtRecall
• IoU

The template offers the following Losses API:

• Loss config should contain _target_ key with loss class name and other parameters which are required by loss.
• Parameter contains 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 to use any losses, for example from PyTorch or implemented by yourself (see examples in modules/losses/components/).

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

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

• _target_: key with model class name.
• model_name: model name
• model_repo (optional): model repository
• Other parameters which are required by model.

By default, 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

By default, the template contains the following LightningModules:

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

See examples of module configs in configs/module folder.

_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 in the template contains the following stages:

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

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

Evaluation loop in the template contains 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 a 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 effectively to use csv format for a large number of predictions, it could help to avoid RAM memory overflow, because csv allows to write row by row and doesn't require to keep in RAM the whole dict like in case of json. To change output format, set predictions_saving_params.output_format variable in configs/extra/default.yaml config file.
• If you need some custom output format, for example parquet, you can easily modify src.utils.saving_utils.save_predictions() method.

See more details about Prediction API and predict_step in LightningModule.

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

By default, the template contains few of them:

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

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

Lightning provides DDP plugins which allow custom integrations to the internals of the Trainer such as custom precision, checkpointing or cluster environment implementation.

By default, the template provides GradCam package which can be extremely useful for understanding network decision and explainability in general.

See more details in package documentation.

Hydra + OmegaConf provide a flexible and efficient configuration management system that allows to dynamically create a hierarchical configurations by composition and override it through config files and the command line.

It is core component of the template which is orchestrating all other modules.

This powerful tools allow to create a simple and efficient way for managing and organizing the various configurations in one place, constructing complex configurations structure without any limits which can be essential in machine learning projects.

All of that enable to easily switch between any parameters and try different configurations without having to manually update the code.

A decorator hydra.main is supposed to be used to load Hydra config during launching of the pipeline. Here a config is being parser by Hydra grammar parser, merged, composed and passed to the pipeline main function.

import hydra
from omegaconf import DictConfig, OmegaConf
from src.train import train


@hydra.main(version_base="1.3", config_path=".", config_name="train.yaml")
def main(cfg: DictConfig) -> None:
    print(OmegaConf.to_yaml(cfg))
    train(cfg)


if __name__ == "__main__":
    main()

Also, it could be done by Hydra Compose API, use initialize, initialize_config_module or initialize_config_dir, instead of hydra.main:

from hydra import compose, initialize_config_dir
from omegaconf import OmegaConf
from src.train import train


def main() -> None:
    with initialize_config_dir(version_base="1.3", config_dir="."):
        cfg = compose(
            config_name="train.yaml", return_hydra_config=True, overrides=[]
        )
    print(OmegaConf.to_yaml(cfg))
    train(cfg)


if __name__ == "__main__":
    main()

For object instantiating from a config, it should contain _target_ key with class or function name which would be instantiated with other parameters passed to config. Hydra provides hydra.utils.instantiate() (and its alias hydra.utils.call()) for instantiating objects and calling class or function. Prefer instantiate for creating objects and call for invoking functions.

loss:
  _target_: "torch.nn.CrossEntropyLoss"

metric:
  _target_: "torchmetrics.Accuracy"
  task: "multiclass"
  num_classes: 10
  top_k: 1

Based on such config you could instantiate loss via loss = hydra.utils.instantiate(config.loss) and metric via metric = hydra.utils.instantiate(config.metric).

It supports few config parameters conversion strategies:

• none: default behavior, use OmegaConf containers
• partial: convert OmegaConf containers to dict and list, except Structured Configs, which remain as DictConfig instances.
• object: convert OmegaConf containers to dict and list, except Structured Configs, which are converted to instances of the backing dataclass / attr class using OmegaConf.to_object
• all: convert everything to primitive containers

It is managed by adding an additional _convert_ parameter to config.

Moreover, it offers a partial instantiation, which can be very useful, for example for function instantiation or recursively object instantiation. It is managed by adding an additional _partial_ bool parameter to config.

output_activation:
  _target_: "torch.softmax"
  dim: 1

output_activation = hydra.utils.instantiate(
    config.output_activation, _partial_=True
)
preds = output_activation(logits)

It supports few operations from the command line as well:

• Override existing config value by passing it as well
• Add a new config value, which doesn't exist in the config, by using +
• Override a config value if it's already in the config, or add it otherwise, by using ++

# train the model with the default config
python src/train.py

# train the model with the overridden parameter
python src/train.py model.model_name="torchvision.models/vit_l_16"

# train the model with the overridden parameter and add a new parameter
python src/train.py model.model_name="torch.hub/vit_l_16_v2" ++model.model_repo="repository"

It provides much more excited features like:

• Structured configs with extended list of available primitive types, nested structure, containers containing primitives, default values, bottom-up values overriding and much more. It gives even more possibilities to structure configs in any ways.

_target_: src.datamodules.datamodules.SingleDataModule

defaults:
  - _self_
  - loaders: default.yaml
  - transforms: default.yaml

datasets:
  train:
    _target_: src.datamodules.datasets.ClassificationDataset
    json_path: ${paths.data_dir}/train/data.json

  valid:
    _target_: src.datamodules.datasets.ClassificationDataset
    json_path: ${paths.data_dir}/valid/data.json

  test:
    _target_: src.datamodules.datasets.ClassificationDataset
    json_path: ${paths.data_dir}/test/data.json

• Colored logs for hydra/job_logging and hydra/hydra_logging
• Hyperparameters sweepers: Optuna, Nevergrad, Ax
• Custom plugins

Hydra provides a way to extend furthermore its functionality by adding custom resolvers via OmegaConf.register_new_resolver(). It allows to add custom executable expressions to the config. See more here.

By default, OmegaConf supports the following resolvers:

• oc.env: returns the value of an environment variable
• oc.create: may be used for dynamic generation of config nodes
• oc.deprecated: may be used to mark a config node as deprecated
• oc.decode: decodes a string using a given codec
• oc.select: provides a default value to use in case the primary interpolation key is not found, or select keys that are otherwise illegal interpolation keys, or works with missing values
• oc.dict.{keys,value}: analogous to the dict.keys and dict.values methods in plain Python dictionaries

But it is a powerful tool, which allows to add any other custom resolvers! For example, it could be annoying to write loss or metric names in configs in each place where it might be required, like early_stopping config, model_checkpoint config, config contained scheduler params, or somewhere else. It could be solved by adding custom resolver for replacing __loss__ and __metric__ names by the actual loss or metric name, which is passed to the config and initialized by Hydra.

Note: You need to register custom resolvers before hydra.main or Compose API calls. Otherwise, it just doesn't apply by Hydra config parser.

In my template, it is implemented as a decorator utils.register_custom_resolvers, which allows to register all custom resolvers in a single place. It supports Hydra's command line flags, which are required to override config path, name or dir. By default, it allows to replace __loss__ to loss.__class__.__name__ and __metric__ to main_metric.__class__.__name__ via such syntax: ${replace:"__metric__/valid"}. Use quotes for defining internal value in ${replace:"..."} to avoid grammar problems with hydra config parser. It can be easily expanded for any other purposes.

import hydra
from omegaconf import DictConfig, OmegaConf
from src.train import train
from src import utils

_HYDRA_PARAMS = {
    "version_base": "1.3",
    "config_path": ".",
    "config_name": "train.yaml",
}

@utils.register_custom_resolvers(**_HYDRA_PARAMS)
@hydra.main(**_HYDRA_PARAMS)
def main(cfg: DictConfig) -> None:
    print(OmegaConf.to_yaml(cfg))
    train(cfg)


if __name__ == "__main__":
    main()

This powerful tool allows to significantly simplify pipeline developing and configuring, for example:

• Instantiate modules with any custom logic under the hood, eg:
  • Instantiate the whole module with all inside submodules recursively by Hydra
  • Main module and some part of inside submodules can be initialized by Hydra and rest of them manually
  • Manually initialize main module and all submodules
• Package dynamic structures, like data augmentations to config, where you can easily set up any transforms classes, parameters or applying order. See more about implemented TransformsWrapper, which can be easily reworked for any over augmentations package. Config example:

train:
  order:
    [
      "resize",
      "random_brightness_contrast",
      "normalize",
    ]
  resize:
    _target_: albumentations.Resize
    height: 256
    width: 256
    p: 1.0
  random_brightness_contrast:
    _target_: albumentations.RandomBrightnessContrast
    brightness_limit: [-0.2, 0.2]
    contrast_limit: [-0.2, 0.2]
    p: 0.5
  normalize:
    _target_: albumentations.Normalize
    mean: [0.485, 0.456, 0.406]
    std: [0.229, 0.224, 0.225]
    p: 1.0

• And many other things... It has very low limitations, so you can use it for any custom logic, which you want to use in your project.

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 single 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/` directory
│   │   │   │   │   └── configs         <- Full copy of `configs/` directory
│   │   │   │   └── ...                 <- Any other thing saved during training
│   │   │   └── ...
│   │   │
│   │   └── multiruns                   <- Logs generated by multiruns
│   │       ├── YYYY-MM-DD_HH-MM-SS     <- Datetime of the multirun
│   │       │   ├──1                    <- Multirun job number
│   │       │   ├──2
│   │       │   └── ...
│   │       └── ...
│   │
│   └── debugs                          <- Logs generated when debugging config is attached
│       └── ...

You can change this structure by modifying paths in Hydra configuration and set another path constants in path configuration.

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

However, there are another methods and one of them calls as Hierarchical Data Format HDF5 or h5py, which has 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
    ...

Jupyter Notebook is a powerful tool for data analysis, visualization and presenting Machine Learning projects. Such sections formatting can be used to have clean and understandable structure of a jupyter notebook:

├── Summary                 <- Summary: for fastly understanding what it is about
│   ├── Objective           <- Objective of analysis
│   ├── Methods             <- Methods which are using
│   └── Results             <- Results of analysis
│
├── Config                  <- Configs and constants
│
├── Libs                    <- Imports and environment variables defining
│
└── Analysis                <- Analysis / Visualization / Presenting / etc.

Furthermore, it can be helpful to use special naming convention for jupyter notebooks: a number (for ordering), the creator's initials and a short description, all of this delimited by -, e.g. 1.0-asg-data-exploration.ipynb.

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

You can 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 is 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 more on Dockerfile and .dockerignore.
• 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 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 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 in general. 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"

Template contains few initial CI workflows via GitHub Actions platform. It makes it easy to automate and streamline development workflows, which can help to save time and efforts, 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 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.