MNIST Experiment: IID and Balanced Dataset

MNIST Experiment: IID and Balanced DatasetΒΆ

import torch
import torchvision

import numpy as np
import math

Loading in the data

mnist_trainset = torchvision.datasets.MNIST(root='./data', train=True, download=True, transform=torchvision.transforms.Compose([
                               torchvision.transforms.ToTensor(),
                               torchvision.transforms.Normalize(
                                 (0.1307,), (0.3081,))
                             ]))
mnist_testset = torchvision.datasets.MNIST(root='./data', train=False, download=True, transform=torchvision.transforms.Compose([
                               torchvision.transforms.ToTensor(),
                               torchvision.transforms.Normalize(
                                 (0.1307,), (0.3081,))
                             ]))

Setting the hyperparameters

## Parameters:
n_epochs = 3
batch_size_train = 10000
batch_size_test = 500
log_interval = 500

train_loader = torch.utils.data.DataLoader(mnist_trainset,batch_size=batch_size_train, shuffle=True)
test_loader = torch.utils.data.DataLoader(mnist_testset,batch_size=batch_size_test, shuffle=False)

import papayaclient

Writing the model class. Here we use a simple two layer FC Net.

class TheModel(torch.nn.Module):

    def __init__(self):
        super(TheModel, self).__init__()

        self.linear1 = torch.nn.Linear(784, 10)
        #self.linear2 = torch.nn.Linear(400, 10)
        #self.relu = torch.nn.ReLU()

    def forward(self, x):
        x1 = x.flatten(start_dim = 1)
        return self.linear1(x1)
        #return self.linear2(self.relu(self.linear1(x1)))

Creating the clients using papaya

clients = []
for batchno, (ex_data, ex_labels) in enumerate(train_loader):
    clients.append(papayaclient.PapayaClient(dat = ex_data,
                                            labs = ex_labels,
                                            batch_sz = 500,
                                            num_partners = 5,
                                            model_class = TheModel,
                                            loss_fn = torch.nn.CrossEntropyLoss))

import random
random.shuffle(clients)

## Train the Nodes
num_epochs_total = 100
num_epochs_per_swap = 5
num_times = (num_epochs_total // num_epochs_per_swap)
for i in range(0, num_times):
    for n in clients:
        for j in range(0, num_epochs_per_swap):
            n.model_train_epoch()
    if i > 1 and i < num_times - 1 :
        for n in clients:
            n.select_partners(3)
        for n in clients:
            for i in range(0, 4) :
                n.update_partner_weights()
            n.average_partners()

for c in clients :
    print(c.logs['stringy'][99])

node3010epoch 99 loss 0.2327485978603363
node3802epoch 99 loss 0.27158960700035095
node3790epoch 99 loss 0.3395092487335205
node432epoch 99 loss 0.2874259650707245
node2642epoch 99 loss 0.2938651144504547
node2787epoch 99 loss 0.3081212043762207

accuracies = {}
with torch.no_grad():
    for i in clients :
        accuracies_node = []
        for batchno, (ex_data, ex_labels) in enumerate(test_loader) :
            accuracies_node.append(((i.model.forward(ex_data).argmax(dim = 1) == ex_labels).float().mean()).item())
        accuracies[i.node_id] = np.array(accuracies_node).mean()

accuracies

{3010: 0.9139999985694885,
 3802: 0.9152999937534332,
 3790: 0.9150999933481216,
 432: 0.914999994635582,
 2642: 0.9155000001192093,
 2787: 0.9150000065565109}

Above we see the accuracy that the model at each node achieves on the held out test set.

previous

Papaya

next

MNIST Non IID