Description

The cmd-wheat-detection Python package includes code to train and evaluate a basic model for the Kaggle Global Wheat Detection problem. This deep learning model uses pytorch with pytorch lightning to run torchvision object detection models.

You can see how I use this code in a Kaggle notebook here.

Usage

Set up virtual environment

python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -e .

Activate virtual environment

source venv/bin/activate

Command-line scripts

Run training as a script using the default configuration

train  # using python entry point

or

python wheat/scripts/train.py

As documented in the PyTorch Lightning docs, PyTorch Lightning CLI flags can be used to configure the training run. For example, to set the number of training epochs for to 10, add --max_epochs=10 to the train command.

Run evaluation on the validation set with weights from a saved checkpoint

evaluate <path-to-checkpoint> # using python entry point

or

python wheat/scripts/evaluate.py <path-to-checkpoint>

Run inference on the test set and save predictions in the Kaggle submission format

predict <path-to-checkpoint> # using python entry point

or

python wheat/scripts/predict.py <path-to-checkpoint>

Jupyter notebook

To install Jupyter notebook support, with the virtual environment active, run pip install -e .[notebook], then start Jupyter.

Provided notebooks

• Visualize dataset examples: notebooks/visualize_dataset.ipynb
• Run training: notebooks/train.ipynb
• Run evaluation and analyze results: notebooks/evaluate.ipynb

Configuration

Configuration is managed using configobj. The full set of configurable parameters with type information is available in wheat/config/configspec.ini. Overrides to default values can be specified in a custom config.ini file.

Example custom config file

Indentation is not required but is included because I find it helps me see which parameters are in which section or sub-section.

['train']
    batch_size = 2
    [['optimizer']]
        initial_lr = 0.0001

Run training with a custom configuration

train --config-path my_config.ini

TODO

Honestly, I'm not sure when I'll have time to do any of these, but it's nice to have goals, right?

• Get rid of "numpy array is not writeable" error
• Replace config with PyTorch Lightning hyperparameters
• Maybe replace numpy seed with PyTorch Lightning seed feature
• Add stochastic weight averaging
• Add tests?
• Add test-time augmentation?