flame

Flame is a lightweight and efficient library for developing neural networks, built using PyTorch Ignite. It simplifies the training, evaluation, and experimentation processes by providing reusable templates and a highly configurable setup. With Flame, you can easily manage everything from logging and checkpoints to custom training loops, making it ideal for both beginners and experienced researchers looking to streamline their workflow.

Contents

• Why Use Flame?
• Concepts
• Installation
• Getting Started

Why Use Flame?

Flame addresses two common needs in neural network development:

• Experiment Templates: Provides templates for training and evaluation with utilities like checkpoint saving, training resumption, logging, and more.
• Flexible Functionality: Easily customize training and testing by editing configuration files. Integrate with Ignite's metrics and handlers or use your own.

Concepts

Details coming soon.

Installation

Set up a Python 3 environment and install Flame using pip:

pip install git+https://github.com/tronglh241/flame.git

Getting Started

Usage

Flame includes two commands:

• Initialize a new project:

  flame init [-h] [-f] [directory]
  

  • directory: Location to initialize the project. Defaults to the current directory.
  • -f, --full: Creates a full project template with an additional run.py file.

• Run training or testing:

  flame run [-h] file
  

  • file: Path to the config file.

Run Your First Experiment

Follow these steps to run a simple classification experiment using the MNIST dataset:

1. Initialize a project:

   mkdir mnist-classification && cd mnist-classification
   flame init
   

   Or, run directly:

   flame init mnist-classification
   

   The folder structure will look like this:

   mnist-classification/
   └── configs
       ├── test_ddp.yml
       ├── test.yml
       ├── train_ddp.yml
       └── train.yml
   

   Use -f for an extra run.py file:

   mnist-classification/
   ├── configs
   │   ├── test_ddp.yml
   │   ├── test.yml
   │   ├── train_ddp.yml
   │   └── train.yml
   └── run.py
   

2. Install additional dependencies:

   pip install torchvision
   

3. Run the training:

   flame run configs/train.yml
   

   To monitor training, use Tensorboard:

   tensorboard --logdir logs/
   

4. Evaluate the model: Update checkpoint.loader.kwargs.path in configs/test.yml to point to the trained model, e.g., checkpoints/best_model.pt:

   checkpoint:
     loader:
       module: flame.handlers
       name: CheckpointLoader
       kwargs:
         path: "'checkpoints/best_model.pt'"
   

   Run the evaluation:

   flame run configs/test.yml
   

That's it! You've successfully trained and evaluated a model using Flame.