Skip to content

Latest commit

 

History

History
317 lines (219 loc) · 14.6 KB

README.md

File metadata and controls

317 lines (219 loc) · 14.6 KB
Raw

Quadra Logo

Docs • Tutorials • Configurations

Quadra aims to simplify deep learning experimenting process, allowing researchers or developers to compare, monitor, and share their experiments quickly. It provides a simple and flexible way to train and deploy deep learning models using YAML configuration files and open-source tools such as Hydra, Lightning framework, and Pytorch. It lets you compose your experiment configurations from single command line interface, so you can conduct multiple experiments with different settings and hyperparameters. Every experiment can be logged using integrations provided by Lightning framework such Mlflow.

Quick Start Guide

If you use pip to manage your packages, you can install quadra from PyPi by running the following command:

pip install quadra

If instead you prefer to use poetry, you can install quadra from PyPi by running the following command:

poetry add quadra

If you don't have virtual environment ready, Let's set up our environment for using the quadra library. We have two parts in this guide: Common setup and Environment-specific setup.

Using Conda

Create and activate a new Conda environment.

conda create -n myenv python=3.10
conda activate myenv

Using Python Virtualenv

Create and activate a new virtual environment.

# replace `myenv` with the name of your virtual environment
python3 -m venv myenv
source myenv/bin/activate

Common Setup

  1. Check your git version: Make sure you have git version 2.10 or higher, to avoid any installation failures.
git --version
  1. Upgrade pip:
pip install --upgrade pip

  1. Install the package

    • Install the quadra package with pip:

      pip install quadra

    • Install the quadra package with poetry:

      curl -sSL https://install.python-poetry.org | python3 -
poetry add quadra

  2. Run from CLI: Run the following command to check if the installation was successful:

quadra experiment=default

Setup Mlflow (Optional)

To use Mlflow and leverage its functionalities such as saving models, logging metrics, saving artifacts, and visualizing results, you need to ensure that the Mlflow server is running. You can find more information about Mlflow here.

By default, the logger configuration is set to Mlflow, and experiments expect the MLFLOW_TRACKING_URI environment variable to be set to the address of the Mlflow server. There are two ways to set this variable:

Using the command line:

export MLFLOW_TRACKING_URI=http://localhost:5000

This command sets the MLFLOW_TRACKING_URI variable to http://localhost:5000. Replace this with the actual address of your Mlflow server if it's running on a different host or port.

Adding it to your environment file:

Quadra uses environment variables to store credentials and other sensitive information. Thanks to python-dotenv library, you can create a .env file in the main folder of your project and store the credentials there. During the runtime, the library will automatically load the environment variables from the .env file. You can also add the MLFLOW_TRACKING_URI variable to your environment file (e.g., .env). Open the file in a text editor and add the following line:

MLFLOW_TRACKING_URI=http://localhost:5000

Again, modify the address if your Mlflow server is running on a different host or port.

By setting the MLFLOW_TRACKING_URI variable using either method, you configure the logger to connect to the Mlflow server, enabling you to utilize its features effectively.

The export command is specific to Unix-based systems like Linux or macOS. If you are using a different operating system, refer to the appropriate method for setting environment variables.

Run Example Experiments

quadra provides a set of example experiments that can be used to test the installation and to provide some example configuration files.

By default all the experiments will run on the GPU 0, to run it on the CPU you can specify a different trainer configuration parameter:

quadra <configurations> trainer=lightning_cpu

Classification Training

To run a simple classification training on the Imagenette dataset with a Resnet18 architecture run the following command:

quadra experiment=generic/imagenette/classification/default logger=csv

This will train the model for 20 epochs and log the metrics in a csv file, at the end of the training a torchscript model will be saved for inference alongside some output images.

By default the experiment will run on the GPU 0, to run it on the CPU you can specify a different trainer configuration parameter.

Segmentation Training

To run a simple segmentation training on the Oxford pet dataset with a Unet architecture run the following command:

quadra experiment=generic/oxford_pet/segmentation/smp logger=csv

This will make use of the segmentation models pytorch library to train the model for 10 epochs, logging the metrics to a csv file. At the end of the training a torchscript model will be saved for inference alongside some output images.

(SSL) Self-supervised Learning Training

On the same dataset we can run a simple SSL training using the BYOL algorithm with the following command:

quadra experiment=generic/imagenette/ssl/byol logger=csv

BYOL is not the only SSL algorithm available, you can find a list of all the available algorithms under quadra/configs/experiment/generic/imagenette/ssl folder.

Anomaly Detection Training

To run a simple anomaly detection training on the MNIST dataset using the PADIM algorithm run the following command:

quadra experiment=generic/mnist/anomaly/padim logger=csv

This will run an anomaly detection considering on of the classes as good (default is the number 9) and the rest as anomalies.

This will make use of the anomalib library to train the model. Many different algorithms are available, you can find them under quadra/configs/experiment/generic/mnist/anomaly folder.

Running with Custom Datasets

Each task comes with a default configuration file that can be customized for your needs. Each example experiment we have seen so far uses a default configuration file that can be found under quadra/configs/experiment/base/<task>/<config_name>.yaml.

Let's see how we can customize the configuration file to run the classification experiment on a custom dataset.

Structure your dataset in the following way:

dataset/
├── class_1
│   ├── abc.xyz
│   └── ...
├── class_2
│   ├── abc.xyz
│   └── ...
├── class_3 
│   ├── abc.xyz
│   └── ...

Create a new experiment configuration file under quadra/configs/experiment/custom/<config_name>.yaml with the following content:

# @package _global_
defaults:
  - base/classification/classification # extends the default classification configuration

core:
  name: <your-custom-experiment-name> # name of the experiment

model:
  num_classes: <number-of_classes-you-have> # number of classes in your dataset

Run the experiment with the following command:

quadra experiment=custom/<config_name> logger=csv

It will run the experiment using the configuration file you have just created and it will apply the default parameters from the classification configuration file. Furthermore, it will log the metrics to a csv file. You can add or customize the parameters in the configuration file to fit your needs.

For more information about advanced usage, please check tutorials and task specific examples.

Development

First clone the repository from Github

First clone the repository from Github, then we need to install the package with optional dependencies (generally in editable mode) and enable the pre-commit hooks.

  1. git clone https://github.com/orobix/quadra.git && cd quadra
  2. Install poetry curl -sSL https://install.python-poetry.org | python3 -
  3. Install the required poetry plugins
poetry self add poetry-bumpversion
poetry self add poetry-dotenv-plugin
  1. Install quadra package in editable mode poetry install --with test,dev,docs --all-extras
  2. Install pre-commit hooks pre-commit install
  3. (Optional) Eventually build documentation by calling required commands (see below).

Now you can start developing and the pre-commit hooks will run automatically to prevent you from committing code that does not pass the linting and formatting checks.

We rely on a combination of Pylint, Mypy and Ruff to enforce code quality.

Building Documentations

  1. Activate your virtual environment.
  2. Install the quadra package with at least doc version.
  3. To run the webserver for real-time rendering and editing run mkdocs serve and visit http://localhost:8000/.
  4. If you want to export the static website to a specific folder mkdocs build -d <Destination Folder>

Acknowledgements

This project is based on many open-source libraries and frameworks, we would like to thank all the contributors for their work. Here is a list of the main libraries and frameworks we use:

  • Pytorch and Pytorch Lightning for training and deploying deep learning models. These two libraries are core part of training and testing tasks that allow us to run experiments on different devices in agile way.
  • Pretrained models are usually loaded from Pytorch Hub or Pytorch-image-models (or called as timm).
  • Each specific task may rely on different libraries. For example, segmentation task uses Segmentation_models.pytorch for loading backbones. The anomaly detection task uses a fork of Anomalib maintained by Orobix on this repository. We use light-weight ML models from scikit-learn. We have also implementation of some SOTA models inside our library.
  • Data processing and augmentation are done using Albumentations and OpenCV.
  • Hydra for composing configurations and running experiments. Hydra is a powerful framework that allows us to compose configurations from command line interface and run multiple experiments with different settings and hyperparameters. We have followed suggestions from Configuring Experiments section of Hydra documentation and lightning-hydra-template repository.
  • Documentation website is using Material for MkDocs and MkDocs. For code documentation we are using Mkdocstrings. For releasing software versions we combine Bumpver and Mike.
  • Models can be exported in different ways (torchscript or torch file). We have also added ONNX support for some models.
  • Testing framework is based on Pytest and related plug-ins.
  • Code quality is ensured by pre-commit hooks. We are using Ruff for linting, enforcing code quality and formatting, Pylint for in depth linting and Mypy for type checking.

FAQ

How can I fix errors related to GL when I install full opencv package?

If you are running on a remote server without a display and you are using opencv-python instead of opencv-python-headless you can run the following command to fix the issue:

It can be solved by correctly linking the libGL.so.1 library:

# Check where the library is located
find /usr -name libGL.so.1
# if the library is located in /usr/lib/x86_64-linux-gnu
# link it to /usr/lib
sudo ln -s /usr/lib/x86_64-linux-gnu/libGL.so.1 /usr/lib/libGL.so

or installing following packages (depending on your OS this may vary):

sudo apt-get install libgl1-mesa-glx

How can I run multiple experiments with single command?

You can run multiple experiments with a single command by passing --multirun flag:

quadra <configurations> --multirun

For example if you want to run same experiment with different seeds you can run:

quadra experiment=generic/imagenette/classification/default trainer=lightning_cpu logger=csv core.seed=1,2,3 --multirun