Tutorial job

submit.sh

#! /bin/bash
#SBATCH -c 1
#SBATCH --mem 1GB
#SBATCH --time 10:00
#SBATCH -A stats
#SBATCH --array 0-8

module load anaconda
conda load pytorch
python tutorial.py --shard $SLURM_ARRAY_TASK_ID

summarise.py

import numpy as np
import pandas as pd

def read_all(files):
    results = pd.concat([pd.read_csv(f) for f in files])
    return results.groupby(['n', 'distribution'])['pvalue'].agg(level_50=lambda x: np.mean(x < 0.5), level_95=lambda x: np.mean(x < 0.95))


def main():
        import argparse
        parser = argparse.ArgumentParser()
        parser.add_argument('files', type=str, nargs='+')

        args = parser.parse_args()
        result = read_all(args.files)
        print(result)

if __name__ == '__main__':
    main()

tutorial.py

import numpy as np
import scipy.stats
import pandas as pd
import itertools


def simulate_t_test_pvalue(n, dist='norm', rng=None):
    if rng is None:
        rng = np.random

    if dist == 'normal':
        x = rng.randn(n)
    elif dist == 't':
        x = rng.standard_t(3, n)
    elif dist == 'cauchy':
        x = rng.standard_cauchy(n)
    else:
        raise ValueError('invalid distribution')

    test_result = scipy.stats.ttest_1samp(x, 0)
    return test_result.pvalue


def run_simulation(n, dist, num_replications, rng=None):
    pvalues = [simulate_t_test_pvalue(n, dist, rng) for _ in range(num_replications)]
    return pd.DataFrame({
        'pvalue': pvalues,
        'n': n,
        'distribution': dist
    })


def main():
    import argparse

    parser = argparse.ArgumentParser()
    parser.add_argument('--shard', type=int)

    args = parser.parse_args()

    ns = [10, 20, 40]
    dists = ['normal', 't', 'cauchy']

    shard = args.shard

    configs = list(itertools.product(ns, dists))

    result = run_simulation(*configs[shard], num_replications=2000, rng=np.random.RandomState(shard))
    result.to_csv('output_{0}.csv'.format(shard), index=False)


if __name__ == '__main__':
    main()