ptutils

Pytorch utilities for model training on GPU and TPU in a single, flexible interface that can be subclassed with your own methods. Model results are (optionally) saved to a MongoDB, for asynchronous visualization.

Installation

To install run:

git clone https://github.com/anayebi/ptutils
cd ptutils/
pip install -e .

Training

The example scripts support training ResNet-18 on ImageNet categorization, e.g.

cd ptutils/model_training/
python runner.py --config=configs/resnet18_supervised_imagenet_trainer_[gpu/tpu].json

You can substitute your own training method by importing from ptutils.model_training.runner import Runner, and subclassing Runner.train().

MongoDB

By default, this packages saves model results to MongoDB. If you would like to use it, follow these instructions to install MongoDB. Otherwise, to disable this feature, set "use_mongodb": false in your configuration json.

The function ptutils.core.utils.grab_results() is an example of how to grab the results from MongoDB for the SupervisedImageNetTrainer, and this notebook gives an example of plotting it.

Code Formatting:

Put this in .git/hooks/pre-commit, and run sudo chmod +x .git/hooks/pre-commit.

#!/usr/bin/env bash
  
echo "# Running pre-commit hook"
echo "#########################"

echo "Checking formatting"

format_occurred=false
declare -a black_dirs=("ptutils/" "setup.py")
for black_dir in "${black_dirs[@]}"; do
    echo ">>> Checking $black_dir"
    black --check "$black_dir"

    if [ $? -ne 0 ]; then
        echo ">>> Reformatting now!"
        black "$black_dir"
        format_occurred=true
    fi
done

if [ "$format_occurred" = true ]; then
    exit 1
fi

License

MIT

Contributors

• Aran Nayebi (Stanford/MIT)
• Nathan C. L. Kong (Stanford)
• Javier Sagastuy-Brena (Stanford)

Contact

If you have any questions or encounter issues, either submit a Github issue here (preferred) or email me.