README.md

A block-random algorithm for learning on distributed, heterogeneous data

Python environments

The quickest way to setup the Python environment is to use pipenv install and then pipenv shell.

Benchmarks on classification EMNIST data sets

Reproducing the benchmark results from the paper entails the following steps:

1. Download and unzip the EMNIST data sets into ./data/emnist-dataset folder

2. Train CNN on different data sets with shuffled, sorted and block-random orderings, at different batch sizes:

   $ python classification-tests.py -d fashion
   $ python classification-tests.py -d digits
   $ python classification-tests.py -d letters
   $ python classification-tests.py -d byclass
   $ python classification-tests.py -d bymerge
   $ python classification-tests.py -d balanced
   $ python classification-tests.py -d mnist"

   The outputs are stored in the batch_size_study directory.

3. Plot the results: $ python plot.py

Predicting $\tau_{ij}$ for LES of channel flow

Reproducing the channel flow results from the paper entails the following steps:

1. Generate the filtered data from the DNS: $ python scaling.py. This creates scaled.npy in the data directory which has the filtered velocities, gradients and $\tau_ij$ terms.

2. Generate the training and test data for the various runs:

   $ python gen_data.py -o shuffled-1m -p shuffled -b 16 -n 1000000
   $ python gen_data.py -o shuffled-16 -p shuffled -b 16
   $ python gen_data.py -o block-16 -p block -b 16
   $ python gen_data.py -o sorted-16 -p sorted -b 16
   

3. Perform hyperparameter sweeps using the shell script:

   $ sh parameter_sweeps.sh 1
   $ sh parameter_sweeps.sh 2
   $ sh parameter_sweeps.sh 3
   $ sh parameter_sweeps.sh 4

4. Plot comparisons of results: $ python compare_runs.py -r runs

5. Train the models using different types of algorithms: $ sh model_runs.sh

6. Plot a given model result: $ python plot_run.py -r runs/${directory_name}

Citation for this work

@article{Mohan19,
    author    = {P. Mohan, M. T. Henry de Frahan, R. King, and R. W. Grout},
    title     = {A block-random algorithm for learning on distributed, heterogeneous data},
    journal   = {arXiv:1903.00091},
    year      = {2019}
}