Why is Google's Colab Pro with GPU so slow?

Question

I used the colab pro to train cifar-10 data. But it's very slow, even spent 2 more times than CPU on colab pro and my own PC. These data are on google dirve, and I used pytorch to train. Could someone help me?

I also tried changing to accelerator in colab to 'None', but it's also faster than 'GPU'. It's very strange?

And here is my code:

label_names = ["airplane", "automobile", "bird", 
               "cat", "deer", "dog", "frog", "horse", "ship", "truck"]

label_dict = {}
for idx, name in enumerate(label_names):
    label_dict[name] = idx

def default_loader(path):
    return Image.open(path).convert("RGB")

train_transform = transforms.Compose(
    [
        transforms.RandomResizedCrop((28, 28)), # 原来是32x32
        transforms.RandomHorizontalFlip(),
        transforms.RandomVerticalFlip(),
        transforms.RandomRotation(90),
        transforms.RandomGrayscale(0.1),
        transforms.ColorJitter(0.3, 0.3, 0.3, 0.3),
        transforms.ToTensor()
    ]
)

# 数据的读取和增强
class MyDataset():
    def __init__(self, im_list, transform=None, loader=default_loader):
        super(MyDataset, self).__init__()
        imgs = []
        for im_item in im_list:
            im_label_name = im_item.split('/')[-2]
            imgs.append([im_item, label_dict[im_label_name]])
        
        self.imgs = imgs
        self.transform = transform
        self.loader = loader
        
    def __getitem__(self, index):
        im_path, im_label = self.imgs[index]
        im_data = self.loader(im_path)
        
        if self.transform is not None:
            im_data = self.transform(im_data)
        
        return im_label, im_data
        
    def __len__(self):
        return len(self.imgs)

im_train_list = glob.glob("./drive/MyDrive/cifar-10-batches-py/train/*/*png")
im_test_list = glob.glob("./drive/MyDrive/cifar-10-batches-py/test/*/*png")

# Enhance
train_dataset = MyDataset(im_list=im_train_list, transform=train_transform) 
test_dataset = MyDataset(im_list=im_test_list, transform=transforms.ToTensor())

train_data_loader = DataLoader(dataset=train_dataset, 
                               batch_size=128, shuffle=True, num_workers=4)

test_data_loader = DataLoader(dataset=test_dataset, 
                               batch_size=128, shuffle=False, num_workers=4)

# VGGnet
class VGGnet(nn.Module):
    def __init__(self):
        super(VGGnet, self).__init__()
        
        # 3 * 28 * 28
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU()
        )
        # 14 * 14
        
        self.maxpooling1 = nn.MaxPool2d(kernel_size=2, stride=2)
        
        # 下采样之后channel通常会翻倍
        self.conv2_1 = nn.Sequential(
            nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(128),
            nn.ReLU()
        )
        
        self.conv2_2 = nn.Sequential(
            nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(128),
            nn.ReLU()
        )
        # 7 * 7        
        self.maxpooling2 = nn.MaxPool2d(kernel_size=2, stride=2)
        
        self.conv3_1 = nn.Sequential(
            nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU()
        )
        
        self.conv3_2 = nn.Sequential(
            nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU()
        )      
        # 4 * 4
        self.maxpooling3 = nn.MaxPool2d(kernel_size=2, stride=2, padding=1)
        
        self.conv4_1 = nn.Sequential(
            nn.Conv2d(256, 512, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU()
        )
        
        self.conv4_2 = nn.Sequential(
            nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU()
        )         
        
        # 2* 2
        self.maxpooling4 = nn.MaxPool2d(kernel_size=2, stride=2)
        
        # 全连接 (一个FC层)
        self.fc = nn.Linear(512 * 4, 10)
        
    def forward(self, x):
        batchsize = x.size(0)
        out = self.conv1(x)
        out = self.maxpooling1(out)
        
        out = self.conv2_1(out)
        out = self.conv2_2(out)
        out = self.maxpooling2(out)
        
        out = self.conv3_1(out)
        out = self.conv3_2(out)
        out = self.maxpooling3(out)
        
        out = self.conv4_1(out)
        out = self.conv4_2(out)
        out = self.maxpooling4(out)
        
        out = out.view(batchsize, -1)
        out = self.fc(out)
        out = F.log_softmax(out, dim=1)
        
        return out


device = torch.device("cuda")

epoch_num = 10
lr = 0.01
net = VGGnet().to(device)

# loss
loss_func = nn.CrossEntropyLoss()

# optimizer
optimizer = torch.optim.Adam(net.parameters(), lr=lr)

# lr
torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.9)       

%%time
for epoch in range(epoch_num):
    print(epoch)
    net.train()
    for i, data in enumerate(train_data_loader):
        labels, inputs = data
        labels, inputs = labels.to(device), inputs.to(device)
        
        outputs = net(inputs)
        loss = loss_func(outputs, labels)
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
#         print("loss is {}".format(loss.item()))
        
        _, pred = torch.max(outputs.data, dim=1)
        correct = pred.eq(labels.data).sum()
        
        batch_size = inputs.size(0)
        print("step = {}; loss = {}; mini-batch correct = {}"
              .format(i, loss.item(), 100. * correct/batch_size) )

        

Answer 1

I knew the reason! When I put the data to local, not google dirve, it's very very fast! (Although I still don't understand why the diff of CPU is not big (colab "None" vc PC CPU))

!mkdir train_local
!wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
!tar zxvf cifar-10-python.tar.gz -C train_local
!rm cifar-10-python.tar.gz