# PyTorch-IE-Hydra-Template A clean and scalable template to kickstart your deep learning based information extraction project 🚀⚡🔥
Click on [Use this template](https://github.com/ChristophAlt/pytorch-ie-hydra-template/generate) to initialize new repository.
This project is heavily inspired by [Lightning-Hydra-Template](https://github.com/ashleve/lightning-hydra-template). _Suggestions are always welcome!_

📌 Introduction

Why you should use it:

• Convenient all-in-one technology stack for deep learning based information extraction prototyping - allows you to rapidly iterate over new models, datasets and tasks on different hardware accelerators like CPUs, multi-GPUs or TPUs.
• A collection of best practices for efficient workflow and reproducibility.
• Thoroughly commented - you can use this repo as a reference and educational resource.

Why you shouldn't use it:

• Not fitted for data engineering - the template configuration setup is not designed for building data processing pipelines that depend on each other.
• Limits you as much as PyTorch-IE limits you.
• PyTorch-IE and Hydra are still evolving and integrate many libraries, which means sometimes things break - for the list of currently known problems visit this page.

Main Technologies

PyTorch-IE - a lightweight information extraction (IE) technology stack built on top of PyTorch Lightning, Huggingface Datasets and Huggingface Hub for reproducible high-performance AI research.

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 Of This Template

• Predefined Structure: clean and scalable so that work can easily be extended and replicated | #Project Structure
• Rapid Experimentation: thanks to automating pipeline with config files and hydra command line superpowers | #Your Superpowers
• Reproducibility: obtaining similar results is supported in multiple ways | #Reproducibility
• Little Boilerplate: so pipeline can be easily modified | #How It Works
• Main Configuration: main config file specifies default training configuration | #Main Project Configuration
• Experiment Configurations: can be composed out of smaller configs and override chosen hyperparameters | #Experiment Configuration
• Workflow: comes down to 4 simple steps | #Workflow
• Experiment Tracking: many logging frameworks can be easily integrated, like Tensorboard, MLFlow or W&B | #Experiment Tracking
• Logs: all logs (checkpoints, data from loggers, hparams, etc.) are stored in a convenient folder structure imposed by Hydra | #Logs
• Hyperparameter Search: made easier with Hydra built-in plugins like Optuna Sweeper | #Hyperparameter Search
• Tests: unit tests and shell/command based tests for speeding up the development | #Tests
• Best Practices: a couple of recommended tools, practices and standards for efficient workflow and reproducibility | #Best Practices

Project Structure

The directory structure of new project looks like this:

├── configs                   <- Hydra configuration files
│   ├── callbacks                <- Callbacks configs
│   ├── datamodule               <- Datamodule configs
│   ├── dataset                  <- Dataset configs
│   ├── debug                    <- Debugging configs
│   ├── experiment               <- Experiment configs
│   ├── hparams_search           <- Hyperparameter search configs
│   ├── local                    <- Local configs
│   ├── logger                   <- Logger configs
│   ├── model                    <- Model configs
│   ├── paths                    <- Project paths configs
│   ├── taskmodule               <- Taskmodule configs
│   ├── trainer                  <- Trainer configs
│   │
│   ├── predict.yaml          <- Main config for inference
│   ├── evaluate.yaml         <- Main config for testing
│   └── train.yaml            <- Main config for training
│
├── data                   <- Project data
│
├── dataset_builders       <- dataset builders
│   ├── hf                   <- Huggingface
│   └── pie                  <- PyTorch-IE
│
├── logs                   <- Logs generated by Hydra and PyTorch Lightning loggers
│   (created on demand)
│
├── models                 <- Per default (see config parameter `save_dir`), trained models and their
│   (created on demand)       respective taskmodules are copied to this location in a format to easily
│                             load them with pytorch_ie.Auto* classes.
│
├── 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
│   ├── datamodules              <- Lightning datamodules
│   ├── models                   <- Pytorch-IE models
│   ├── taskmodules              <- Pytorch-IE taskmodules
│   ├── utils                    <- Utility scripts
│   ├── vendor                   <- Third party code that cannot be installed using PIP/Conda
│   │
│   ├── eval.py                  <- Run evaluation
│   ├── predict.py               <- Run inference
│   └── train.py                 <- Run training
│
├── tests                  <- Tests of any kind
│
├── .env.example              <- Template of the file for storing private environment variables
├── .gitignore                <- List of files/folders ignored by git
├── .pre-commit-config.yaml   <- Configuration of pre-commit hooks for code formatting
├── Makefile                  <- Makefile with commands like `make train` or `make test`
├── requirements.txt          <- File for installing python dependencies
├── setup.cfg                 <- Configuration of linters and pytest
├── setup.py                  <- File for installing project as a package
└── README.md

🚀  Quickstart

# clone project
git clone https://github.com/ChristophAlt/pytorch-ie-hydra-template.git
cd pytorch-ie-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

# [OPTIONAL] symlink log directories and the default model directory to
# "$HOME/experiments/my-project" since they can grow a lot
bash setup_symlinks.sh $HOME/experiments/my-project

# [OPTIONAL] set any environment variables by creating an .env file
# 1. copy the provided example file:
cp .env.example .env
# 2. edit the .env file for your needs!

Template contains example with CONLL2003 Named Entity Recognition.
When running python train.py you should see something like this:

⚡  Your Superpowers

❤️  Contributions

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 To Get Started

• First, you should probably get familiar with PyTorch-IE
• Next, go through Hydra quick start guide and basic Hydra tutorial

How It Works

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

_target_: pytorch_ie.models.transformer_token_classification.TransformerTokenClassificationModel

model_name_or_path: bert-base-uncased
learning_rate: 0.005

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 datasets with command line arguments:

python train.py model=transformer_token_classification

The whole pipeline managing the instantiation logic is placed in src/train.py.

Main Configuration

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

Experiment Configuration

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

Local Configuration

Location: configs/local
Some configurations are user/machine/installation specific (e.g. configuration of local cluster, or harddrive paths on a specific machine). For such scenarios, a file configs/local/default.yaml can be created which is automatically loaded but not tracked by Git.

Workflow

Before creating your own setup, have a look into the Pytorch-IE documentation to make yourself familiar with the Pytorch-IE core concepts like the document, model, and taskmodule.

1. Write your PyTorch-IE dataset loader (see dataset_builders/pie/conll2003/conll2003.py for an example) or try out one of the PIE datasets hosted at huggingface.co/pie.
2. Write your PyTorch-IE model (see src/models/transformer_token_classification.py for an example) or use one of the implementations from pytorch-ie or pie-modules.
3. Write your PyTorch-IE taskmodule (see src/taskmodules/transformer_token_classification.py for example) or use one of the implementations from pytorch-ie or pie-modules.
4. Write your experiment config, containing paths to your model, taskmodule and dataset (see configs/experiment/conll2003.yaml for example). You may need to also write configs for your model, taskmodule and dataset, if you do not want to use the default ones.
5. If necessary, define additional_model_kwargs for your model class in the train.py (see line with # NOTE: MODIFY THE additional_model_kwargs IF YOUR MODEL REQUIRES ...").
6. Execute a dev run for your setup to ensure that everything works as expected (assuming that configs/experiments/experiment_name.yaml is your experiment config file): python src/train.py experiment=experiment_name +trainer.fast_dev_run=true
7. Run training with chosen experiment config on the GPU: python src/train.py experiment=experiment_name trainer=gpu

Logs

Hydra creates new working directory for every executed run. By default, logs have the following structure:

├── logs
│   ├── pipeline_type                     # Folder for the logs generated by type of pipeline, i.e. training, evaluation, or prediction
│   │   ├── runs                          # Folder for single runs
│   │   │   ├── experiment_name             # Experiment name
│   │   │   │   ├── 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                     # Folder for multiruns
│   │       ├── experiment_name             # Experiment name
│   │       │   ├── 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-IE is based on PyTorch Lightning which supports many popular logging frameworks:
**Weights&Biases · Neptune · Comet · MLFlow · Tensorboard · Aim **

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 here or take a look at TransformerTokenClassification 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.

Next, you can execute it with: python train.py -m hparams_search=conll2003_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/pipeline_type/multirun folder.

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

Continuous Integration

Template comes with CI workflows implemented in Github Actions:

• .github/workflows/tests.yaml: running tests that are not marked as "slow"
• .github/workflows/pre-commit.yaml: running code quality checks, see .pre-commit-config.yaml for configured entries

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

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.datamodule)
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.datamodule, _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: ${datamodule.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.

Inference

The following code is an example of loading model from checkpoint and running predictions.

Reproducibility

What provides reproducibility:

• Hydra manages your configs
• Hydra manages your logging paths and makes every executed run store its hyperparameters and config overrides in a separate file in logs
• Single seed for random number generators in pytorch, numpy and python.random
• LightningDataModule allows you to encapsulate data split, transformations and default parameters in a single, clean abstraction
• LightningModule separates your research code from engineering code in a clean way
• Experiment tracking frameworks take care of logging metrics and hparams, some can also store results and artifacts in cloud
• Pytorch Lightning takes care of creating training checkpoints
• Example callbacks for wandb show how you can save and upload a snapshot of codebase every time the run is executed, as well as upload ckpts and track model gradients

Best Practices

Other Repositories

License

This project 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

# Your Project Name
[](https://www.nature.com/articles/nature14539) [](https://papers.nips.cc/paper/2020)

📌 Description

What it does

🚀 Quickstart

Environment Setup

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

# [OPTIONAL] create conda environment
conda create -n your-project-name python=3.9
conda activate your-project-name

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

# install remaining requirements
pip install -r requirements.txt

# [OPTIONAL] symlink log directories and the default model directory to
# "$HOME/experiments/your-project-name" since they can grow a lot
bash setup_symlinks.sh $HOME/experiments/your-project-name

# [OPTIONAL] set any environment variables by creating an .env file
# 1. copy the provided example file:
cp .env.example .env
# 2. edit the .env file for your needs!

Model Training

Have a look into the train.yaml config to see all available options.

Train model with default configuration

# train on CPU
python src/train.py

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

Execute a fast development run (train for two steps)

python src/train.py +trainer.fast_dev_run=true

Train model with chosen experiment configuration from configs/experiment/

python src/train.py experiment=conll2003

You can override any parameter from command line like this

python train.py trainer.max_epochs=20 datamodule.batch_size=64

Start multiple runs at once (multirun):

python src/train.py seed=42,43 --multirun

Notes:

• this will execute two experiments (one after the other), one for each seed
• the results will be aggregated and stored in logs/multirun/, see the last logging output for the exact path

Model evaluation

This will evaluate the model on the test set of the chosen dataset using the metrics implemented within the model. See config/dataset/ for available datasets.

Have a look into the evaluate.yaml config to see all available options.

python src/evaluate.py dataset=conll2003 model_name_or_path=pie/example-ner-spanclf-conll03

Notes:

• add the command line parameter trainer=gpu to run on GPU

Inference

This will run inference on the given dataset and split. See config/dataset/ for available datasets. The result documents including the predicted annotations will be stored in the predictions/ directory (exact location will be printed to the console).

Have a look into the predict.yaml config to see all available options.

python src/predict.py dataset=conll2003 model_name_or_path=pie/example-ner-spanclf-conll03

Notes:

• add the command line parameter +pipeline.device=0 to run the inference on GPU 0

Evaluate Serialized Documents

This will evaluate serialized documents including predicted annotations (see Inference) using a document metric. See config/metric/ for available metrics.

Have a look into the evaluate_documents.yaml config to see all available options

python src/evaluate_documents.py metric=f1 metric.layer=entities +dataset.data_dir=PATH/TO/DIR/WITH/SPLITS

Note: By default, this utilizes the dataset provided by the from_serialized_documents configuration. This configuration is designed to facilitate the loading of serialized documents, as generated during the Inference step. It requires to set the parameter data_dir. If you want to use a different dataset, you can override the dataset parameter as usual with any existing dataset config, e.g dataset=conll2003. But calculating the F1 score on the bare conll2003 dataset does not make much sense, because it does not contain any predictions. However, it could be used with statistical metrics such as count_text_tokens or count_entity_labels.

Development

# run pre-commit: code formatting, code analysis, static type checking, and more (see .pre-commit-config.yaml)
pre-commit run -a

# run tests
pytest -k "not slow" --cov --cov-report term-missing