Skip to content

Latest commit

 

History

History
52 lines (41 loc) · 3 KB

README.md

File metadata and controls

52 lines (41 loc) · 3 KB
Raw

A template for creating a deep learning project.

Your really detailed description goes here.

Using this template

  1. Create a new environment with pipenv shell
  2. Install flit with pip install flit
  3. Install the template via flit: flit install -s
  4. Train a mnist model: train
  5. Evaluate the trained model: evaluate

Simple, huh? Well, kind of.

The guts of the project are in src/__init__.py, src/train_config.py and src/eval_config.py. src/__init__.py dynamically loads the classes required for training and evaluation (a model class, defined in src/models/, a dataset, in src/datasets/, and an evaluator class src/evaluators/). It then smushes the models/datasets/evaluators together to either train or evaluate a model. Pytorch Lightning takes care of all of the training specifics through the trainer_args class in src/train_config.py.

All arguments to the model/dataset/evaluator classes are defined in the config files: src/train_config.py, and src/eval_config.py.

This structure tries to disentangle model, dataset, and evaluation code. Good rules of thumb: Don't import anything from src.models, src.datasets, or src.evaluators into other scripts. Rely on util/util.py to store helper functions and other operations that you use frequently.

It's ok to hardcode variables for certain experiments since you can always copy/paste code when you want to start a new experiment.

Another cool feature: This project relies on sacred (https://github.com/IDSIA/sacred) to manage its CLI. This means that you can change arguments for training or evaluating via the command line. For example, if I wanted to train the MNISTModel for 10 epochs, I'd run:

train with trainer_args.max_epochs=10

You can go on and configure your whole experiment via the command line if you want - training hyperparameters will be saved in the model log directory (configured in train_config.py).

Adding new models/datasets/evaluators

Easy! Just create a new file in src/models: touch src/models/my_cool_idea.py and create a class that inherits from pl.LightningModule (just as the MNISTModel does in src/models/mnist_classifier.py). Be sure to fill out the same methods as the MNISTModel. Creating a new dataset/evaluator is the exact same, just replace models/ with datasets/evaluators. The dataset you create has to inherit from the DataModule class (in src/datasets/__init__.py), and the evaluator class has to inherit from the Evaluator class (in src/evaluators/__init__.py).

What's this .pre-commit-config.yaml thing?

Pre-commit is a tool that runs checks when you enter git commit. Install in your virtual environment with pip install pre-commit, set up with pre-commit install, and run an initial check with pre-commit run --all-files. If a test fails, you need to go in and fix it, either by running the command (like black or isort) or by fixing the file manually. You don't have to set this up but having consistent formatting makes collaboration easier.