Deep learning on large spatio-temporal data, including fMRI and climate data.
- Moving MNIST
- Brain fMRI
- Climate
- PredRNN++
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 fromBaseTrainer
and are responsible for constructing models as well as training and evaluating them.
All examples are run with the generic training script, train.py
.
To run the examples on the Cori supercomputer, you may use the provided example SLURM batch script. Here's how to run the Hello World example on 4 Haswell nodes:
sbatch -N 4 scripts/train_cori.sh configs/hello.yaml