experiments.py

import decentralized
import learning.nn
import learning.optimizers
import numpy as np

learning.optimizers.Variable.datatype = np.float16  # change to float64 on enough memory


def experiment(
    dataset,
    device_type=decentralized.devices.GossipDevice,
    gossip_merge=decentralized.mergers.AvgMerge,
    classifier=learning.nn.MLP,
    pretrained=False,
    gossip_pull=False,
    smoother=decentralized.mergers.NoSmooth,
    seed=0,
):
    measures = {"acc": list(), "base_acc": list()}
    network, test_labels = decentralized.simulation.create_network(
        dataset,
        device_type,
        pretrained=pretrained,
        classifier=classifier,
        gossip_merge=gossip_merge,
        gossip_pull=gossip_pull,
        seed=seed,
        smoother=smoother,
        min_communication_rate=0,
        max_communication_rate=0.1,
    )
    for epoch in range(1000):
        network.round()
        if epoch % 10 == 0:
            accuracy_base = sum(
                1.0 if network.devices[u].predict(False) == label else 0
                for u, label in test_labels.items()
            ) / len(test_labels)
            accuracy = sum(
                1.0 if network.devices[u].predict() == label else 0
                for u, label in test_labels.items()
            ) / len(test_labels)
            measures["base_acc"].append(accuracy_base)
            measures["acc"].append(accuracy)
            print(
                f"Epoch {epoch} \t Acc {accuracy:.3f} \t Base acc {accuracy_base:.3f}"
            )
    return measures


for dataset in ["citeseer", "cora", "pubmed"]:
    setting = {
        "dataset": dataset,
        "device_type": decentralized.devices.GossipDevice,
        "gossip_merge": decentralized.mergers.AvgMerge,
        "classifier": learning.nn.Tautology,
        "pretrained": False,
        "gossip_pull": False,
        "smoother": decentralized.mergers.NoSmooth,
    }
    print(setting)
    measures = {"acc": list(), "base_acc": list()}
    for repetition in range(1):
        for measure, curve in experiment(**setting, seed=repetition).items():
            if measure in measures:
                measures[measure].append(curve[-1])
    print(
        "Average "
        + str(
            {
                measure: [np.mean(values), np.std(values)]
                for measure, values in measures.items()
            }
        )
    )