utils.py

import torch
import torch.nn as nn
import torch.nn.functional as F


def get_kl_loss(model, args, epoch):
    kl_loss = 0.0
    for layer in model.modules():
        if hasattr(layer, "tensor"):
            kl_loss += layer.tensor.get_kl_divergence_to_prior()
    kl_mult = args.kl_multiplier * torch.clamp(
        torch.tensor((
                (epoch - args.no_kl_epochs) / args.warmup_epochs)), 0.0, 1.0)
    """
    print("KL loss ",kl_loss.item())
    print("KL Mult ",kl_mult.item())
    """
    return kl_loss * kl_mult.to(kl_loss.device)


def get_net(args):
    if args.model_type in ['CP', 'TensorTrain', 'TensorTrainMatrix', 'Tucker']:
        return get_TensorizedNet(args)
    else:
        return Net()


def train(args, model, device, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)

        if args.rank_loss:
            ard_loss = get_kl_loss(model, args, epoch)
            loss += ard_loss

        loss.backward()
        optimizer.step()
        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.item()))
            if args.dry_run:
                break


def test(model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item()  # sum up batch loss
            pred = output.argmax(dim=1, keepdim=True)  # get the index of the max log-probability
            correct += pred.eq(target.view_as(pred)).sum().item()

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


def get_TensorizedNet(args):
    if args.model_type == 'full':
        fc1 = nn.Linear(784, 512)
        fc2 = nn.Linear(512, 10)

    else:
        if args.model_type == 'TensorTrainMatrix':
            shape1 = [[4, 7, 4, 7], [4, 4, 8, 4]]
            shape2 = [[16, 32], [2, 5]]

        else:
            shape1 = [28, 28, 16, 32]
            shape2 = [32, 16, 10]

        fc1 = TensorizedLinear(784, 512, shape=shape1, tensor_type=args.model_type, max_rank=args.rank,
                               em_stepsize=args.em_stepsize)
        fc2 = TensorizedLinear(512, 10, shape=shape2, tensor_type=args.model_type, max_rank=args.rank,
                               em_stepsize=args.em_stepsize)

    return TensorizedNet(fc1, fc2)


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.dropout = nn.Dropout(0.5)
        self.fc1 = nn.Linear(784, 512)
        self.fc2 = nn.Linear(512, 10)
        self.relu = nn.ReLU()

    def forward(self, x):
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc2(x)
        output = F.log_softmax(x, dim=1)
        return output


class TensorizedNet(nn.Module):
    def __init__(self, fc1, fc2):
        super(TensorizedNet, self).__init__()
        self.dropout = nn.Dropout(0.5)
        self.add_module('fc1', fc1)
        self.add_module('fc2', fc2)
        self.relu = nn.ReLU()

    def forward(self, x):
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc2(x)
        output = F.log_softmax(x, dim=1)
        return