PyTorch Project Template

My favorite way to organize a PyTorch research project. It also offers boilerplate for multi-GPU distributed training, mixed-precision training, and Tensorboard.

Requirements

I personally prefer using the latest version of everything.

pip install -r requirements.txt

Preprocessing

Often, we need to preprocess the downloaded dataset for our own use. generate_dataset.py should do the job of processing 'raw' datasets into 'tidy' ones. The processed result should be stored under the data/dataset directory.

python generate_dataset.py --dataset 'cifar10' --source 'downloades/raw_cifar10'

Training

To start training, run train.py.

python train.py --dataset 'cifar10' --epoch 10 --batch_size 128

Or, the .conf files can be used instead of passing lengthy command-line arguments. .conf files should be stored under the data/params directory.

python train.py @data/params/cifar10.conf

Upon completion of each epoch, the model checkpoint will be stored at data/model_checkpoints/[label]. Also, tensorboard can be used to monitor the training process.

Evaluation

Once training is done, run evaluate.py to start the evaluation.

python evaluate.py @data/params/cifar10.conf

License