svrg mnist

main.py

from __future__ import print_function
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
from torch.autograd import Variable
import svrg_solver

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
                    help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
                    help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=10, metavar='N',
                    help='number of epochs to train (default: 10)')
parser.add_argument('--inners', type=int, default=200, metavar='N',
                    help='number of inner updates to train (default: 200)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
                    help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
                    help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='enables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
                    help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=100, metavar='N',
                    help='how many batches to wait before logging training status')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()

torch.manual_seed(args.seed)
if args.cuda:
    torch.cuda.manual_seed(args.seed)


kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
train_loader = torch.utils.data.DataLoader(
    datasets.MNIST('../data', train=True, download=True,
                   transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.1307,), (0.3081,))
                   ])),
    batch_size=args.batch_size, shuffle=True, **kwargs)
args.inners = train_loader.sampler.num_samples # TODO: try m < n

test_loader = torch.utils.data.DataLoader(
    datasets.MNIST('../data', train=False, transform=transforms.Compose([
                       transforms.ToTensor(),
                       transforms.Normalize((0.1307,), (0.3081,))
                   ])),
    batch_size=args.batch_size, shuffle=True, **kwargs)


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
        self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
        self.conv2_drop = nn.Dropout2d()
        self.fc1 = nn.Linear(320, 50)
        self.fc2 = nn.Linear(50, 10)

    def forward(self, x):
        x = F.relu(F.max_pool2d(self.conv1(x), 2))
        #x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
        x = F.relu(F.max_pool2d(self.conv2(x), 2))
        x = x.view(-1, 320)
        x = F.relu(self.fc1(x))
        #x = F.dropout(x, training=self.training)
        x = self.fc2(x)
        return F.log_softmax(x)

model = Net()
model_prev = Net()
if args.cuda:
    model.cuda()
    model_prev.cuda()

sgd = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)
svrg = svrg_solver.SVRG(model.parameters(), model_prev.parameters(), lr=args.lr)
svrg.copy_params()

def train(epoch, opt, inners):
    model.train()
    verbose = True
    for batch_idx, (data, target) in enumerate(train_loader):
        if batch_idx > inners:
            break
        if batch_idx == 3:
            verbose = False
        def closure_sgd():
            sgd.zero_grad()
            output = model(data)
            loss = F.nll_loss(output, target)
            loss.backward()
            return loss

        def closure_svrg():
            # net
            model.zero_grad()
            output = model(data)
            loss = F.nll_loss(output, target)
            loss.backward()
            # net_prev
            model_prev.zero_grad()
            output = model_prev(data)
            loss0 = F.nll_loss(output, target)
            loss0.backward()
            return loss

        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data), Variable(target)

        if opt is 'sgd':
            loss = sgd.step(closure_sgd)
        elif opt is 'svrg':
            loss = svrg.step(closure_svrg, verbose)
        else:
            raise TypeError('not supported optimizer!')

        if batch_idx % args.log_interval == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.data[0]))

def test(epoch):
    model.eval()
    test_loss = 0
    correct = 0
    for data, target in test_loader:
        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data, volatile=True), Variable(target)
        output = model(data)
        test_loss += F.nll_loss(output, target).data[0]
        pred = output.data.max(1)[1] # get the index of the max log-probability
        correct += pred.eq(target.data).cpu().sum()

    test_loss = test_loss
    test_loss /= len(test_loader) # loss function already averages over batch size
    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
        test_loss, correct, len(test_loader.dataset),
        100. * correct / len(test_loader.dataset)))

#import ipdb; ipdb.set_trace()

for epoch in range(1, args.epochs + 1):
    def closure(data, target):
        if args.cuda:
            data, target = data.cuda(), target.cuda()
        data, target = Variable(data), Variable(target)
        model_prev.zero_grad()
        output = model_prev(data)
        loss0 = F.nll_loss(output, target)
        loss0.backward()
        return loss0

    if epoch == -1:
        train(epoch, 'sgd', args.inners)
    else:
        # full grad: g_k = f'(y_k)
        svrg.full_grad(train_loader, closure) # TODO: itertools.islice(train_loader, i_end)

        # one epoch, m steps of v^t = f'_i(x^t) - f'_i(y_k) + g_k, x^t+1 = x^t - lr * v^t
        train(epoch, 'svrg', args.inners)

        # copy params
        svrg.copy_params()

        # acceleration

    # validation
    test(epoch)

svrg_solver.py

from torch.optim.optimizer import Optimizer, required


class SVRG(Optimizer):
    r"""Implements stochastic variance reduction gradient descent (optionally with momentum).

    Args:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float): learning rate

    Example:

    .. note::
    """

    def __init__(self, params, params_prev, lr=required, weight_decay=0):
        defaults = dict(lr=lr, weight_decay=weight_decay)
        super(SVRG, self).__init__(params, defaults)
        for group in self.param_groups:
            group['params_prev'] = list(params_prev)

    def __setstate__(self, state):
        super(SVRG, self).__setstate__(state)

    def step(self, closure=None, verbose=False):
        """Performs a single optimization step.

        Arguments:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            loss = closure() # forward, backward for model and model_prev

        for group in self.param_groups:
            weight_decay = group['weight_decay']

            for p, p0 in zip(group['params'], group['params_prev']):
                if p.grad is None:
                    continue

                state = self.state[p0]
                assert 'full_grad' in state # should be non-empty
                d_p = p.grad.data
                d_p0 = p0.grad.data
                v = d_p - d_p0 + state['full_grad']
                if verbose:
                    print('v.norm = {}, fg.norm = {}'.format(v.norm(), state['full_grad'].norm()))
                    if state['full_grad'].norm() < 1e-12:
                        import ipdb; ipdb.set_trace()

                if weight_decay != 0:
                    v.add_(weight_decay, p.data)
                p.data.add_(-group['lr'], v)

        return loss

    def full_grad(self, data_loader, closure):
        """
        """
        num_samples = data_loader.sampler.num_samples
        for i, (data, target) in enumerate(data_loader):
            # forward, backward -> grad
            closure(data, target)

            for group in self.param_groups:
                for p in group['params_prev']:
                    if p.grad is None:
                        continue
                    grad = p.grad.data
                    state = self.state[p]
                    if 'full_grad' not in state:
                        state['full_grad'] = grad.new().resize_as_(grad).zero_()
                    if i == 0:
                        state['full_grad'].zero_()
                    state['full_grad'].add_(1/float(num_samples), grad)

    def copy_params(self):
        """
        """
        for group in self.param_groups:
            for p, p0 in zip(group['params'], group['params_prev']):
                #if p.grad is None:
                #    continue
                p0.data.copy_(p.data)