import torch
import torchvision
import torchvision.transforms as transforms

transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4,
                                          shuffle=True, num_workers=2)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                       download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=4,
                                         shuffle=False, num_workers=2)

classes = ('plane', 'car', 'bird', 'cat',
           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

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

def make_sparse(in_dim, out_dim, size, mode):
    # Always compress the more compressible dimension
    invert = False
    if out_dim > in_dim:
        invert = True
        out_dim, in_dim = in_dim, out_dim

    mask = torch.zeros(out_dim, in_dim)
    assert(mode in ['Expander', 'Group'])

    if mode == 'Expander':
        minstd = -1
        for _ in range(20):
            m = torch.zeros(out_dim, in_dim)
            for i in range(out_dim):
                x = torch.randperm(in_dim)
                m[i][x[:size]] = 1
            std = m.sum(dim=0).std()
            if minstd == -1 or minstd > std:
                minstd = std
                mask = m

    elif mode == 'Group':
        assert(in_dim%size==0 and out_dim%size==0)
        for i in range(out_dim):
            for j in range(in_dim):
                if (i//size == j//size):
                    mask[i][j] == 1

    if invert:
        mask = mask.t()

    return mask

class SparseConv2d(torch.nn.Module):
    def __init__(self, inWCin, inWCout, kernel_size, stride=1, padding=0, dilation=1, sparse_size=0, sparse_mode='Expander'):
        super(SparseConv2d, self).__init__()
        self.kernel_size = kernel_size
        self.stride = stride
        self.padding = padding
        self.dilation = dilation
        self.out_channels = inWCout

        mask = make_sparse(in_dim=inWCin, out_dim=inWCout, size=sparse_size, mode=sparse_mode)
        weight = torch.zeros((inWCout, inWCin))
        weight = 0.01*torch.nn.init.kaiming_normal_(weight)
        weight = torch.mul(weight, mask)
        weight = weight.unsqueeze(2).unsqueeze(3).repeat(1, 1, kernel_size, kernel_size)
        weight = weight.view(weight.size(0), -1)
        weight = weight.to_sparse().cuda()
        self.weight = torch.nn.Parameter(weight, requires_grad=True)

    def forward(self, x):
        out = (x.size(2)+2*self.padding-self.dilation*(self.kernel_size-1)-1)//self.stride+1
        x_unf = torch.nn.functional.unfold(x, (self.kernel_size, self.kernel_size)).transpose(1,2)
        x_unf = torch.sparse.mm(self.weight, x_unf.reshape(-1, x_unf.size(2)).t()).t().reshape(x.size(0),-1,self.out_channels).transpose(1,2)
        x_unf = x_unf.view(x_unf.size(0), x_unf.size(1), out, out)
        return x_unf


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5, bias=False)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5, bias=False)
        self.fc1 = nn.Conv2d(16, 120, 5, bias=False)
        self.fc2 = nn.Conv2d(120, 84, 1, bias=False)
        self.fc3 = nn.Conv2d(84, 10, 1, bias=False)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        x = x.view(-1, 10)
        return x

class Net2(nn.Module):
    def __init__(self):
        super(Net2, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5, bias=False)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = SparseConv2d(6, 16, 5, sparse_size=12)
        self.fc1 = SparseConv2d(16, 120, 5, sparse_size=80)
        self.fc2 = SparseConv2d(120, 84, 1, sparse_size=64)
        self.fc3 = nn.Conv2d(84, 10, 1, bias=False)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        x = x.view(-1, 10)
        return x

#net = Net().cuda()
net = Net2().cuda()

import torch.optim as optim

criterion = nn.CrossEntropyLoss().cuda()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

for epoch in range(2):  # loop over the dataset multiple times

    running_loss = 0.0
    for i, data in enumerate(trainloader, 0):
        # get the inputs; data is a list of [inputs, labels]
        inputs, labels = data
        inputs, labels = inputs.cuda(), labels.cuda()

        # zero the parameter gradients
        optimizer.zero_grad()

        # forward + backward + optimize
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        # print statistics
        running_loss += loss.item()
        if i % 2000 == 1999:    # print every 2000 mini-batches
            print('[%d, %5d] loss: %.3f' %
                  (epoch + 1, i + 1, running_loss / 2000))
            running_loss = 0.0

print('Finished Training')

correct = 0
total = 0
with torch.no_grad():
    for data in testloader:
        images, labels = data
        images, labels = images.cuda(), labels.cuda()
        outputs = net(images)
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

print('Accuracy of the network on the 10000 test images: %d %%' % (
    100 * correct / total))