📌  Introduction

Why you might want to use it:

✅ Save on boilerplate
Easily add new models, datasets, tasks, experiments, and train on different accelerators, like multi-GPU, TPU or SLURM clusters.

✅ Education
Thoroughly commented. You can use this repo as a learning resource.

✅ Reusability
Collection of useful MLOps tools, configs, and code snippets. You can use this repo as a reference for various utilities.

Why you might not want to use it:

❌ Things break from time to time
Lightning and Hydra are still evolving and integrate many libraries, which means sometimes things break. For the list of currently known problems visit this page.

❌ Not adjusted for data engineering
Template is not really adjusted for building data pipelines that depend on each other. It's more efficient to use it for model prototyping on ready-to-use data.

❌ Overfitted to simple use case
The configuration setup is built with simple lightning training in mind. You might need to put some effort to adjust it for different use cases, e.g. lightning fabric.

❌ Might not support your workflow
For example, you can't resume hydra-based multirun or hyperparameter search.

Note: Keep in mind this is unofficial community project.

Main Technologies

PyTorch Lightning - a lightweight PyTorch wrapper for high-performance AI research. Think of it as a framework for organizing your PyTorch code.

Hydra - a framework for elegantly 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.

Main Ideas

• Rapid Experimentation: thanks to hydra command line superpowers
• Minimal Boilerplate: thanks to automating pipelines with config instantiation
• Main Configs: allow you to specify default training configuration
• Experiment Configs: allow you to override chosen hyperparameters and version control experiments
• Workflow: comes down to 4 simple steps
• Experiment Tracking: Tensorboard, W&B, Neptune, Comet, MLFlow and CSVLogger
• Logs: all logs (checkpoints, configs, etc.) are stored in a dynamically generated folder structure
• Hyperparameter Search: simple search is effortless with Hydra plugins like Optuna Sweeper
• Tests: generic, easy-to-adapt smoke tests for speeding up the development
• Continuous Integration: automatically test and lint your repo with Github Actions
• Best Practices: a couple of recommended tools, practices and standards

Project Structure

The directory structure of new project looks like this:

├── .github                   <- Github Actions workflows
│
├── configs                   <- Hydra configs
│   ├── callbacks                <- Callbacks configs
│   ├── data                     <- Data configs
│   ├── debug                    <- Debugging configs
│   ├── experiment               <- Experiment configs
│   ├── extras                   <- Extra utilities configs
│   ├── hparams_search           <- Hyperparameter search configs
│   ├── hydra                    <- Hydra configs
│   ├── local                    <- Local configs
│   ├── logger                   <- Logger configs
│   ├── model                    <- Model 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 and lightning loggers
│
├── notebooks              <- Jupyter notebooks. Naming convention is a number (for ordering),
│                             the creator's initials, and a short `-` delimited description,
│                             e.g. `1.0-jqp-initial-data-exploration.ipynb`.
│
├── scripts                <- Shell scripts
│
├── src                    <- Source code
│   ├── data                     <- Data scripts
│   ├── models                   <- Model scripts
│   ├── utils                    <- Utility scripts
│   │
│   ├── eval.py                  <- Run evaluation
│   └── train.py                 <- Run training
│
├── tests                  <- Tests of any kind
│
├── .env.example              <- Example of file for storing private environment variables
├── .gitignore                <- List of files ignored by git
├── .pre-commit-config.yaml   <- Configuration of pre-commit hooks for code formatting
├── .project-root             <- File for inferring the position of project root directory
├── environment.yaml          <- File for installing conda environment
├── 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

🚀  Quickstart

# clone project
git clone https://github.com/ashleve/lightning-hydra-template
cd lightning-hydra-template

# [OPTIONAL] create conda environment
conda create -n myenv python=3.9
conda activate myenv

# install pytorch according to instructions
# https://pytorch.org/get-started/

# install requirements
pip install -r requirements.txt

Template contains example with MNIST classification.
When running python src/train.py you should see something like this:

⚡  Your Superpowers

python train.py trainer.max_epochs=20 model.optimizer.lr=1e-4

python train.py +model.new_param="owo"

# train on CPU
python train.py trainer=cpu

# train on 1 GPU
python train.py trainer=gpu

# train on TPU
python train.py +trainer.tpu_cores=8

# train with DDP (Distributed Data Parallel) (4 GPUs)
python train.py trainer=ddp trainer.devices=4

# train with DDP (Distributed Data Parallel) (8 GPUs, 2 nodes)
python train.py trainer=ddp trainer.devices=4 trainer.num_nodes=2

# simulate DDP on CPU processes
python train.py trainer=ddp_sim trainer.devices=2

# accelerate training on mac
python train.py trainer=mps

# train with pytorch native automatic mixed precision (AMP)
python train.py trainer=gpu +trainer.precision=16

# set project and entity names in `configs/logger/wandb`
wandb:
  project: "your_project_name"
  entity: "your_wandb_team_name"

# train model with Weights&Biases (link to wandb dashboard should appear in the terminal)
python train.py logger=wandb

python train.py experiment=example

python train.py callbacks=default

# gradient clipping may be enabled to avoid exploding gradients
python train.py +trainer.gradient_clip_val=0.5

# run validation loop 4 times during a training epoch
python train.py +trainer.val_check_interval=0.25

# accumulate gradients
python train.py +trainer.accumulate_grad_batches=10

# terminate training after 12 hours
python train.py +trainer.max_time="00:12:00:00"

# runs 1 epoch in default debugging mode
# changes logging directory to `logs/debugs/...`
# sets level of all command line loggers to 'DEBUG'
# enforces debug-friendly configuration
python train.py debug=default

# run 1 train, val and test loop, using only 1 batch
python train.py debug=fdr

# print execution time profiling
python train.py debug=profiler

# try overfitting to 1 batch
python train.py debug=overfit

# raise exception if there are any numerical anomalies in tensors, like NaN or +/-inf
python train.py +trainer.detect_anomaly=true

# use only 20% of the data
python train.py +trainer.limit_train_batches=0.2 \
+trainer.limit_val_batches=0.2 +trainer.limit_test_batches=0.2

python train.py ckpt_path="/path/to/ckpt/name.ckpt"

python eval.py ckpt_path="/path/to/ckpt/name.ckpt"

# this will run 6 experiments one after the other,
# each with different combination of batch_size and learning rate
python train.py -m data.batch_size=32,64,128 model.lr=0.001,0.0005

# this will run hyperparameter search defined in `configs/hparams_search/mnist_optuna.yaml`
# over chosen experiment config
python train.py -m hparams_search=mnist_optuna experiment=example

python train.py -m 'experiment=glob(*)'

python train.py -m seed=1,2,3,4,5 trainer.deterministic=True logger=csv tags=["benchmark"]

pre-commit run -a

pre-commit autoupdate

# run all tests
pytest

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

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

python train.py tags=["mnist","experiment_X"]

>>> python train.py tags=[]
[2022-07-11 15:40:09,358][src.utils.utils][INFO] - Enforcing tags! <cfg.extras.enforce_tags=True>
[2022-07-11 15:40:09,359][src.utils.rich_utils][WARNING] - No tags provided in config. Prompting user to input tags...
Enter a list of comma separated tags (dev):

>>> python train.py -m +x=1,2,3 tags=[]
ValueError: Specify tags before launching a multirun!

❤️  Contributions

This project exists thanks to all the people who contribute.

Have a question? Found a bug? Missing a specific feature? Feel free to file a new issue, discussion or PR with respective title and description.

Before making an issue, please verify that:

• The problem still exists on the current main branch.
• Your python dependencies are updated to recent versions.

Suggestions for improvements are always welcome!

How It Works

All PyTorch Lightning modules are dynamically instantiated from module paths specified in config. Example model config:

_target_: src.models.mnist_model.MNISTLitModule
lr: 0.001
net:
  _target_: src.models.components.simple_dense_net.SimpleDenseNet
  input_size: 784
  lin1_size: 256
  lin2_size: 256
  lin3_size: 256
  output_size: 10

Using this config we can instantiate the object with the following line:

model = hydra.utils.instantiate(config.model)

This allows you to easily iterate over new models! Every time you create a new one, just specify its module path and parameters in appropriate config file.

Switch between models and datamodules with command line arguments:

python train.py model=mnist

Example pipeline managing the instantiation logic: src/train.py.

Main Config

Location: configs/train.yaml
Main project config contains default training configuration.
It determines how config is composed when simply executing command python train.py.

# order of defaults determines the order in which configs override each other
defaults:
  - _self_
  - data: mnist.yaml
  - model: mnist.yaml
  - callbacks: default.yaml
  - logger: null # set logger here or use command line (e.g. `python train.py logger=csv`)
  - trainer: default.yaml
  - paths: default.yaml
  - extras: default.yaml
  - hydra: default.yaml

  # experiment configs allow for version control of specific hyperparameters
  # e.g. best hyperparameters for given model and datamodule
  - experiment: null

  # config for hyperparameter optimization
  - hparams_search: null

  # optional local config for machine/user specific settings
  # it's optional since it doesn't need to exist and is excluded from version control
  - optional local: default.yaml

  # debugging config (enable through command line, e.g. `python train.py debug=default)
  - debug: null

# task name, determines output directory path
task_name: "train"

# tags to help you identify your experiments
# you can overwrite this in experiment configs
# overwrite from command line with `python train.py tags="[first_tag, second_tag]"`
# appending lists from command line is currently not supported :(
# https://github.com/facebookresearch/hydra/issues/1547
tags: ["dev"]

# set False to skip model training
train: True

# evaluate on test set, using best model weights achieved during training
# lightning chooses best weights based on the metric specified in checkpoint callback
test: True

# simply provide checkpoint path to resume training
ckpt_path: null

# seed for random number generators in pytorch, numpy and python.random
seed: null

Experiment Config

Location: configs/experiment
Experiment configs allow you to overwrite parameters from main config.
For example, you can use them to version control best hyperparameters for each combination of model and dataset.

# @package _global_

# to execute this experiment run:
# python train.py experiment=example

defaults:
  - override /data: mnist.yaml
  - override /model: mnist.yaml
  - override /callbacks: default.yaml
  - override /trainer: default.yaml

# all parameters below will be merged with parameters from default configurations set above
# this allows you to overwrite only specified parameters

tags: ["mnist", "simple_dense_net"]

seed: 12345

trainer:
  min_epochs: 10
  max_epochs: 10
  gradient_clip_val: 0.5

model:
  optimizer:
    lr: 0.002
  net:
    lin1_size: 128
    lin2_size: 256
    lin3_size: 64

data:
  batch_size: 64

logger:
  wandb:
    tags: ${tags}
    group: "mnist"

Workflow

Basic workflow

1. Write your PyTorch Lightning module (see models/mnist_module.py for example)
2. Write your PyTorch Lightning datamodule (see data/mnist_datamodule.py for example)
3. Write your experiment config, containing paths to model and datamodule
4. Run training with chosen experiment config:

   python src/train.py experiment=experiment_name.yaml

Experiment design

Say you want to execute many runs to plot how accuracy changes in respect to batch size.

1. Execute the runs with some config parameter that allows you to identify them easily, like tags:

   python train.py -m logger=csv data.batch_size=16,32,64,128 tags=["batch_size_exp"]

2. Write a script or notebook that searches over the logs/ folder and retrieves csv logs from runs containing given tags in config. Plot the results.

Logs

Hydra creates new output directory for every executed run.

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
│   │   │   │   └── ...                     # 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.

Experiment Tracking

PyTorch Lightning supports many popular logging frameworks: Weights&Biases, Neptune, Comet, MLFlow, Tensorboard.

These tools help you keep track of hyperparameters and output metrics and allow you to compare and visualize results. To use one of them simply complete its configuration in configs/logger and run:

python train.py logger=logger_name

You can use many of them at once (see configs/logger/many_loggers.yaml for example).

You can also write your own logger.

Lightning provides convenient method for logging custom metrics from inside LightningModule. Read the docs or take a look at MNIST example.

Tests

Template comes with generic tests implemented with pytest.

# run all tests
pytest

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

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

Most of the implemented tests don't check for any specific output - they exist to simply verify that executing some commands doesn't end up in throwing exceptions. You can execute them once in a while to speed up the development.

Currently, the tests cover cases like:

• running 1 train, val and test step
• running 1 epoch on 1% of data, saving ckpt and resuming for the second epoch
• running 2 epochs on 1% of data, with DDP simulated on CPU

And many others. You should be able to modify them easily for your use case.

There is also @RunIf decorator implemented, that allows you to run tests only if certain conditions are met, e.g. GPU is available or system is not windows. See the examples.

Hyperparameter Search

You can define hyperparameter search by adding new config file to configs/hparams_search.

# @package _global_

defaults:
  - override /hydra/sweeper: optuna

# choose metric which will be optimized by Optuna
# make sure this is the correct name of some metric logged in lightning module!
optimized_metric: "val/acc_best"

# here we define Optuna hyperparameter search
# it optimizes for value returned from function with @hydra.main decorator
hydra:
  sweeper:
    _target_: hydra_plugins.hydra_optuna_sweeper.optuna_sweeper.OptunaSweeper

    # 'minimize' or 'maximize' the objective
    direction: maximize

    # total number of runs that will be executed
    n_trials: 20

    # choose Optuna hyperparameter sampler
    # docs: https://optuna.readthedocs.io/en/stable/reference/samplers.html
    sampler:
      _target_: optuna.samplers.TPESampler
      seed: 1234
      n_startup_trials: 10 # number of random sampling runs before optimization starts

    # define hyperparameter search space
    params:
      model.optimizer.lr: interval(0.0001, 0.1)
      data.batch_size: choice(32, 64, 128, 256)
      model.net.lin1_size: choice(64, 128, 256)
      model.net.lin2_size: choice(64, 128, 256)
      model.net.lin3_size: choice(32, 64, 128, 256)

Next, execute it with: python train.py -m hparams_search=mnist_optuna

Using this approach doesn't require adding any boilerplate to code, everything is defined in a single config file. The only necessary thing is to return the optimized metric value from the launch file.

You can use different optimization frameworks integrated with Hydra, like Optuna, Ax or Nevergrad.

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

This approach doesn't support resuming interrupted search and advanced techniques like prunning - for more sophisticated search and workflows, you should probably write a dedicated optimization task (without multirun feature).

Continuous Integration

Template comes with CI workflows implemented in Github Actions:

• .github/workflows/test.yaml: running all tests with pytest
• .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

Distributed Training

Lightning supports multiple ways of doing distributed training. The most common one is DDP, which spawns separate process for each GPU and averages gradients between them. To learn about other approaches read the lightning docs.

You can run DDP on mnist example with 4 GPUs like this:

python train.py trainer=ddp

Note: When using DDP you have to be careful how you write your models - read the docs.

Accessing Datamodule Attributes In Model

The simplest way is to pass datamodule attribute directly to model on initialization:

# ./src/train.py
datamodule = hydra.utils.instantiate(config.data)
model = hydra.utils.instantiate(config.model, some_param=datamodule.some_param)

Note: Not a very robust solution, since it assumes all your datamodules have some_param attribute available.

Similarly, you can pass a whole datamodule config as an init parameter:

# ./src/train.py
model = hydra.utils.instantiate(config.model, dm_conf=config.data, _recursive_=False)

You can also pass a datamodule config parameter to your model through variable interpolation:

# ./configs/model/my_model.yaml
_target_: src.models.my_module.MyLitModule
lr: 0.01
some_param: ${data.some_param}

Another approach is to access datamodule in LightningModule directly through Trainer:

# ./src/models/mnist_module.py
def on_train_start(self):
  self.some_param = self.trainer.datamodule.some_param

Note: This only works after the training starts since otherwise trainer won't be yet available in LightningModule.

Best Practices

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

conda update -n base -c defaults conda

conda create -n myenv python=3.10
conda activate myenv

pip install pre-commit

pre-commit install

pre-commit run -a

pre-commit autoupdate

MY_VAR=/home/user/my_system_path

path_to_data: ${oc.env:MY_VAR}

self.log("train/loss", loss)

from torchmetrics.classification.accuracy import Accuracy


class LitModel(LightningModule):
    def __init__(self)
        self.train_acc = Accuracy()
        self.val_acc = Accuracy()

    def training_step(self, batch, batch_idx):
        ...
        acc = self.train_acc(predictions, targets)
        self.log("train/acc", acc)
        ...

    def validation_step(self, batch, batch_idx):
        ...
        acc = self.val_acc(predictions, targets)
        self.log("val/acc", acc)
        ...

dvc init

dvc add data/MNIST

git add data/MNIST.dvc data/.gitignore
git commit -m "Add raw data"

pip install -e .

pip install git+git://github.com/YourGithubName/your-repo-name.git --upgrade

from project_name.models.mnist_module import MNISTLitModule
from project_name.data.mnist_datamodule import MNISTDataModule

# @package _global_

defaults:
  - override /hydra/launcher@_here_: submitit_slurm

data_dir: /mnt/scratch/data/

hydra:
  launcher:
    timeout_min: 1440
    gpus_per_task: 1
    gres: gpu:1
  job:
    env_set:
      MY_VAR: /home/user/my/system/path
      MY_KEY: asdgjhawi8y23ihsghsueity23ihwd

Resources

This template was inspired by:

• PyTorchLightning/deep-learning-project-template
• drivendata/cookiecutter-data-science
• lucmos/nn-template

Other useful repositories:

• jxpress/lightning-hydra-template-vertex-ai - lightning-hydra-template integration with Vertex AI hyperparameter tuning and custom training job

License

Lightning-Hydra-Template is licensed under the MIT License.

MIT License

Copyright (c) 2021 ashleve

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

DELETE EVERYTHING ABOVE FOR YOUR PROJECT

Description

What it does

Installation

Pip

# clone project
git clone https://github.com/YourGithubName/your-repo-name
cd your-repo-name

# [OPTIONAL] create conda environment
conda create -n myenv python=3.9
conda activate myenv

# install pytorch according to instructions
# https://pytorch.org/get-started/

# install requirements
pip install -r requirements.txt

Conda

# clone project
git clone https://github.com/YourGithubName/your-repo-name
cd your-repo-name

# create conda environment and install dependencies
conda env create -f environment.yaml -n myenv

# activate conda environment
conda activate myenv

How to run

Train model with default configuration

# train on CPU
python src/train.py trainer=cpu

# train on GPU
python src/train.py trainer=gpu

Train model with chosen experiment configuration from configs/experiment/

python src/train.py experiment=experiment_name.yaml

You can override any parameter from command line like this

python src/train.py trainer.max_epochs=20 data.batch_size=64