This repository contains some PyTorch example models and training code with support for distributed training on NERSC systems.

The layout of this package can also serve as a template for PyTorch projects and the provided BaseTrainer and train.py script can be used to reduce boiler plate.

The directory layout of this repo is designed to be flexible:

• Configuration files (in YAML format) go in configs/
• Dataset specifications using PyTorch's Dataset API go into datasets/
• Model implementations go into models/
• Trainer implementations go into trainers/. Trainers inherit from BaseTrainer and are responsible for constructing models as well as training and evaluating them.

All examples are run with the generic training script, train.py.

This package currently contains the following examples:

• CIFAR10 classification with ResNet50 or generic CNN model.
• HEP-CNN classification (https://arxiv.org/abs/1711.03573).
• Minimal Hello World example.

To run the examples on the Cori supercomputer, you may use the provided example SLURM batch script. Here's how to run the ResNet50 CIFAR10 example on 4 Haswell nodes:

sbatch -N 4 scripts/train_cori.sh configs/cifar10.yaml