pytorch如何利用ResNet18進行手寫數字識別

利用ResNet18進行手寫數字識別

先寫resnet18.py

代碼如下:

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


class ResBlk(nn.Module):
    """
    resnet block
    """

    def __init__(self, ch_in, ch_out, stride=1):
        """

        :param ch_in:
        :param ch_out:
        """
        super(ResBlk, self).__init__()

        self.conv1 = nn.Conv2d(ch_in, ch_out, kernel_size=3, stride=stride, padding=1)
        self.bn1 = nn.BatchNorm2d(ch_out)
        self.conv2 = nn.Conv2d(ch_out, ch_out, kernel_size=3, stride=1, padding=1)
        self.bn2 = nn.BatchNorm2d(ch_out)

        self.extra = nn.Sequential()

        if ch_out != ch_in:
            # [b, ch_in, h, w] => [b, ch_out, h, w]
            self.extra = nn.Sequential(
                nn.Conv2d(ch_in, ch_out, kernel_size=1, stride=stride),
                nn.BatchNorm2d(ch_out)
            )

    def forward(self, x):
        """

        :param x: [b, ch, h, w]
        :return:
        """
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))

        # short cut
        # extra module:[b, ch_in, h, w] => [b, ch_out, h, w]
        # element-wise add:
        out = self.extra(x) + out
        out = F.relu(out)

        return out


class ResNet18(nn.Module):
    def __init__(self):
        super(ResNet18, self).__init__()

        self.conv1 = nn.Sequential(
            nn.Conv2d(1, 64, kernel_size=3, stride=3, padding=0),
            nn.BatchNorm2d(64)
        )
        # followed 4 blocks

        # [b, 64, h, w] => [b, 128, h, w]
        self.blk1 = ResBlk(64, 128, stride=2)

        # [b, 128, h, w] => [b, 256, h, w]
        self.blk2 = ResBlk(128, 256, stride=2)

        # [b, 256, h, w] => [b, 512, h, w]
        self.blk3 = ResBlk(256, 512, stride=2)

        # [b, 512, h, w] => [b, 512, h, w]
        self.blk4 = ResBlk(512, 512, stride=2)

        self.outlayer = nn.Linear(512 * 1 * 1, 10)

    def forward(self, x):
        """

        :param x:
        :return:
        """
        # [b, 1, h, w] => [b, 64, h, w]
        x = F.relu(self.conv1(x))

        # [b, 64, h, w] => [b, 512, h, w]
        x = self.blk1(x)
        x = self.blk2(x)
        x = self.blk3(x)
        x = self.blk4(x)

        # print(x.shape) # [b, 512, 1, 1]
        # 意思就是不管之前的特征圖尺寸為多少,隻要設置為(1,1),那麼最終特征圖大小都為(1,1)
        # [b, 512, h, w] => [b, 512, 1, 1]
        x = F.adaptive_avg_pool2d(x, [1, 1])
        x = x.view(x.size(0), -1)
        x = self.outlayer(x)

        return x


def main():
    blk = ResBlk(1, 128, stride=4)
    tmp = torch.randn(512, 1, 28, 28)
    out = blk(tmp)
    print('blk', out.shape)

    model = ResNet18()
    x = torch.randn(512, 1, 28, 28)
    out = model(x)
    print('resnet', out.shape)
    print(model)


if __name__ == '__main__':
    main()

再寫繪圖utils.py

代碼如下

import torch
from matplotlib import pyplot as plt

device = torch.device('cuda')


def plot_curve(data):
    fig = plt.figure()
    plt.plot(range(len(data)), data, color='blue')
    plt.legend(['value'], loc='upper right')
    plt.xlabel('step')
    plt.ylabel('value')
    plt.show()


def plot_image(img, label, name):
    fig = plt.figure()
    for i in range(6):
        plt.subplot(2, 3, i + 1)
        plt.tight_layout()
        plt.imshow(img[i][0] * 0.3081 + 0.1307, cmap='gray', interpolation='none')
        plt.title("{}: {}".format(name, label[i].item()))
        plt.xticks([])
        plt.yticks([])
    plt.show()


def one_hot(label, depth=10):
    out = torch.zeros(label.size(0), depth).cuda()
    idx = label.view(-1, 1)
    out.scatter_(dim=1, index=idx, value=1)
    return out

最後是主函數mnist_train.py

代碼如下:

import torch
from torch import nn
from torch.nn import functional as F
from torch import optim
from resnet18 import ResNet18

import torchvision
from matplotlib import pyplot as plt

from utils import plot_image, plot_curve, one_hot

batch_size = 512

# 加載數據
train_loader = torch.utils.data.DataLoader(
    torchvision.datasets.MNIST('mnist_data', train=True, download=True,
                               transform=torchvision.transforms.Compose([
                                   torchvision.transforms.ToTensor(),
                                   torchvision.transforms.Normalize(
                                       (0.1307,), (0.3081,))
                               ])),
    batch_size=batch_size, shuffle=True)

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

# 在裝載完成後,我們可以選取其中一個批次的數據進行預覽
x, y = next(iter(train_loader))

# x:[512, 1, 28, 28], y:[512]
print(x.shape, y.shape, x.min(), x.max())
plot_image(x, y, 'image sample')

device = torch.device('cuda')

net = ResNet18().to(device)

optimizer = optim.SGD(net.parameters(), lr=0.01, momentum=0.9)

train_loss = []

for epoch in range(5):

    # 訓練
    net.train()
    for batch_idx, (x, y) in enumerate(train_loader):

        # x: [b, 1, 28, 28], y: [512]
        # [b, 1, 28, 28] => [b, 10]
        x, y = x.to(device), y.to(device)
        out = net(x)
        # [b, 10]
        y_onehot = one_hot(y)
        # loss = mse(out, y_onehot)
        loss = F.mse_loss(out, y_onehot).to(device)
        # 先給梯度清0
        optimizer.zero_grad()
        loss.backward()
        # w' = w - lr*grad
        optimizer.step()

        train_loss.append(loss.item())

        if batch_idx % 10 == 0:
            print(epoch, batch_idx, loss.item())

plot_curve(train_loss)
# we get optimal [w1, b1, w2, b2, w3, b3]

# 測試
net.eval()
total_correct = 0
for x, y in test_loader:
    x, y = x.cuda(), y.cuda()
    out = net(x)
    # out: [b, 10] => pred: [b]
    pred = out.argmax(dim=1)
    correct = pred.eq(y).sum().float().item()
    total_correct += correct

total_num = len(test_loader.dataset)
acc = total_correct / total_num
print('test acc:', acc)

x, y = next(iter(test_loader))
x, y = x.cuda(), y.cuda()
out = net(x)
pred = out.argmax(dim=1)
x = x.cpu()
pred = pred.cpu()
plot_image(x, pred, 'test')

結果為:

4 90 0.009581390768289566
4 100 0.010348389856517315
4 110 0.01111914124339819
test acc: 0.9703

運行時註意把模型和參數放在GPU裡,這樣節省時間,此代碼作為測試代碼,僅供參考。

總結

以上為個人經驗,希望能給大傢一個參考,也希望大傢多多支持WalkonNet。

推薦閱讀: