1. 加载MINST数据集
如果还没有安装torch以及torchvision的,请看文章:Torch安装
安装完成之后,详细的python代码如下:
from torchvision import datasets
from torchvision.transforms import ToTensor, Lambda, Compose
import matplotlib.pyplot as plt
import os
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
# load train data
train_data = datasets.MNIST(
root='data', # save in the directory data
train=True, # True means train data, False means test data
download=True,
transform=ToTensor() # (0, 255) to (0, 1)
)
# load test data
test_data = datasets.MNIST(
root='data',
train=False, #True means train data, False means test data
download=True,
transform=ToTensor()
)
下载成功结果:
添加下列代码显示其中一条数据:
print(train_data.data.size())
print(train_data.targets.size())
plt.imshow(train_data.data[130].numpy(), cmap='gray')
plt.title('%i' % train_data.targets[130])
plt.show()
数据如下图所示:
2. 快速构建CNN网络
python代码:
class CNN(nn.Module):
def __init__(self):
super(CNN, self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(
in_channels=1, # size of input channels
out_channels=16, # size of input channels
kernel_size=(5, 5), # size of filter
stride=(1, 1), # step of filter
padding=2, # padding num
),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2),
)
self.conv2 = nn.Sequential(
nn.Conv2d(16, 32, (5, 5), (1, 1), 2),
nn.ReLU(),
nn.MaxPool2d(2),
)
self.flat = nn.Flatten() # flattern the result
self.out = nn.Linear(32 * 7 * 7, 10)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.flat(x)
out = self.out(x)
return out
如果打印一下网络的架构,使用语句:
model = CNN()
print(model)
显示的结构如下所示:
CNN(
(conv1): Sequential(
(0): Conv2d(1, 16, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv2): Sequential(
(0): Conv2d(16, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(flat): Flatten(start_dim=1, end_dim=-1)
(out): Linear(in_features=1568, out_features=10, bias=True)
)
3. 整体代码(CPU版本)
python代码:
import torch
import torchvision
import torch.nn as nn
import torch.utils.data as Data
import matplotlib.pyplot as plt
import numpy as np
import os
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
# Hyper parameters
EPOCH = 1
BATCH_SIZE = 50
LR = 0.001
DOWNLOAD_MINST = True
# Download training data
train_data = torchvision.datasets.MNIST(
root='./mnist',
train=True,
transform=torchvision.transforms.ToTensor(),
download=DOWNLOAD_MINST
)
#print(train_data.data.size())
#print(train_data.targets.size())
#plt.imshow(train_data.data[130].numpy(), cmap='gray')
#plt.title('%i' % train_data.targets[130])
#plt.show()
#
train_loader = Data.DataLoader(
dataset=train_data,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4
)
# Download testing data
test_data = torchvision.datasets.MNIST(
root='./mnist',
train=False,
)
test_x = torch.unsqueeze(test_data.data, dim=1).type(torch.FloatTensor)[:2000] / 255
test_y = test_data.targets[:2000]
class CNN(nn.Module):
def __init__(self):
super(CNN, self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(
in_channels=1,
out_channels=16,
kernel_size=(5, 5),
stride=(1, 1),
padding=2
),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2),
)
self.conv2 = nn.Sequential(
nn.Conv2d(16, 32, (5, 5), (1, 1), 2),
nn.ReLU(),
nn.MaxPool2d(2)
)
self.flat = nn.Flatten()
self.out = nn.Linear(32 * 7 * 7, 10)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.flat(x)
output = self.out(x)
return output
cnn = CNN()
optimizer = torch.optim.Adam(cnn.parameters(), lr=LR)
loss_fuction = nn.CrossEntropyLoss()
if __name__ == '__main__':
for epoch in range(EPOCH):
for step, (batch_x, batch_y) in enumerate(train_loader):
prediction = cnn(batch_x)
loss = loss_fuction(prediction, batch_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 50 == 0:
test_output = cnn(test_x)
pred_y = torch.max(test_output, 1)[1].data.numpy()
accuracy = (sum(pred_y == np.array(test_y.data)).item()) / test_y.size(0)
print('Epoch:%d' % epoch, end='||')
print('train loss:%.4f' % loss.item(), end='||')
print('test accuracy:%.4f' % accuracy)
test_output = cnn(test_x[:10])
pred_y = torch.max(test_output, 1)[1].data.numpy()
print(pred_y, 'prediction number')
print(test_y[:10].numpy(), 'real number')
运行结果:
Epoch:0||train loss:2.3074||test accuracy:0.2195
Epoch:0||train loss:0.2392||test accuracy:0.7895
Epoch:0||train loss:0.3036||test accuracy:0.8910
Epoch:0||train loss:0.2483||test accuracy:0.9140
Epoch:0||train loss:0.3289||test accuracy:0.9045
Epoch:0||train loss:0.1809||test accuracy:0.9275
Epoch:0||train loss:0.0711||test accuracy:0.9480
Epoch:0||train loss:0.1408||test accuracy:0.9480
Epoch:0||train loss:0.2358||test accuracy:0.9600
Epoch:0||train loss:0.2181||test accuracy:0.9445
Epoch:0||train loss:0.0309||test accuracy:0.9675
Epoch:0||train loss:0.1352||test accuracy:0.9575
Epoch:0||train loss:0.1682||test accuracy:0.9725
Epoch:0||train loss:0.0470||test accuracy:0.9735
Epoch:0||train loss:0.0341||test accuracy:0.9710
Epoch:0||train loss:0.1404||test accuracy:0.9680
Epoch:0||train loss:0.1307||test accuracy:0.9670
Epoch:0||train loss:0.1597||test accuracy:0.9720
Epoch:0||train loss:0.0743||test accuracy:0.9735
Epoch:0||train loss:0.0263||test accuracy:0.9765
Epoch:0||train loss:0.0135||test accuracy:0.9735
Epoch:0||train loss:0.0150||test accuracy:0.9770
Epoch:0||train loss:0.0122||test accuracy:0.9765
Epoch:0||train loss:0.0241||test accuracy:0.9760
[7 2 1 0 4 1 4 9 5 9] prediction number
[7 2 1 0 4 1 4 9 5 9] real number
4. 整体代码(GPU版本)
python代码:
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor
# hyper parameters
LR = 0.002
EPOCH = 5
BATCH_SIZE = 50
train_data = datasets.MNIST(
root='data',
train=True,
download=True,
transform=ToTensor()
)
test_data = datasets.MNIST(
root= 'data',
train=False,
download=True,
transform=ToTensor()
)
test_loader = DataLoader(
dataset=test_data,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
)
train_loader = DataLoader(
dataset=train_data,
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=4,
)
class CNN(nn.Module):
def __init__(self):
super(CNN, self).__init__()
self.conv1 = nn.Sequential(
nn.Conv2d(
in_channels=1,
out_channels=16,
kernel_size=(5, 5),
stride=(1, 1),
padding=2,
),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2),
)
self.conv2 = nn.Sequential(
nn.Conv2d(16, 32, (5, 5), (1, 1), 2),
nn.ReLU(),
nn.MaxPool2d(2),
)
self.flat = nn.Flatten()
self.out = nn.Linear(32 * 7 * 7, 10)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.flat(x)
out = self.out(x)
return out
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = CNN().to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=LR)
loss_function = torch.nn.CrossEntropyLoss()
def train(data_loader, model, loss_function, optimizer):
size = len(data_loader.dataset)
model.train()
for batch, (X, y) in enumerate(data_loader):
X, y = X.to(device), y.to(device)
# Compute prediction error
pred = model(X)
loss = loss_function(pred, y)
# Back propagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
if batch % 100 == 0:
loss, current = loss.item(), batch * len(X)
print(f"loss: {loss:>7f} [{current:>5d}/{size:>5d}]")
def test(data_loader, model, loss_fn):
size = len(data_loader.dataset)
num_batches = len(data_loader)
model.eval()
test_loss, correct = 0, 0
with torch.no_grad():
for X, y in data_loader:
X, y = X.to(device), y.to(device)
pred = model(X)
test_loss += loss_fn(pred, y).item()
correct += (pred.argmax(1) == y).type(torch.float).sum().item()
test_loss /= num_batches
correct /= size
print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")
torch.save(model.state_dict(), "model.pth")
print("Saved PyTorch Model State to model.pth")
def call_model():
for t in range(EPOCH):
print(f"Epoch {t + 1}\n-------------------------------")
train(train_loader, model, loss_function, optimizer)
test(test_loader, model, loss_function)
print("Done!")
if __name__ == '__main__':
call_model()
结果展示:
Epoch 1
-------------------------------
loss: 2.319763 [ 0/60000]
loss: 0.197887 [ 5000/60000]
loss: 0.231820 [10000/60000]
loss: 0.029967 [15000/60000]
loss: 0.193301 [20000/60000]
loss: 0.028721 [25000/60000]
loss: 0.230080 [30000/60000]
loss: 0.030196 [35000/60000]
loss: 0.088837 [40000/60000]
loss: 0.061477 [45000/60000]
loss: 0.021868 [50000/60000]
loss: 0.009769 [55000/60000]
Test Error:
Accuracy: 97.0%, Avg loss: 0.092558
Epoch 2
-------------------------------
loss: 0.062530 [ 0/60000]
loss: 0.037623 [ 5000/60000]
loss: 0.057986 [10000/60000]
loss: 0.003451 [15000/60000]
loss: 0.034381 [20000/60000]
loss: 0.005294 [25000/60000]
loss: 0.171846 [30000/60000]
loss: 0.042936 [35000/60000]
loss: 0.073527 [40000/60000]
loss: 0.039359 [45000/60000]
loss: 0.005118 [50000/60000]
loss: 0.003203 [55000/60000]
Test Error:
Accuracy: 98.6%, Avg loss: 0.041006
Epoch 3
-------------------------------
loss: 0.018981 [ 0/60000]
loss: 0.037027 [ 5000/60000]
loss: 0.030264 [10000/60000]
loss: 0.000278 [15000/60000]
loss: 0.024834 [20000/60000]
loss: 0.003731 [25000/60000]
loss: 0.117729 [30000/60000]
loss: 0.058137 [35000/60000]
loss: 0.068251 [40000/60000]
loss: 0.009537 [45000/60000]
loss: 0.015441 [50000/60000]
loss: 0.000517 [55000/60000]
Test Error:
Accuracy: 98.9%, Avg loss: 0.033202
Epoch 4
-------------------------------
loss: 0.005696 [ 0/60000]
loss: 0.011667 [ 5000/60000]
loss: 0.033686 [10000/60000]
loss: 0.000063 [15000/60000]
loss: 0.021328 [20000/60000]
loss: 0.001276 [25000/60000]
loss: 0.115446 [30000/60000]
loss: 0.013636 [35000/60000]
loss: 0.031956 [40000/60000]
loss: 0.006652 [45000/60000]
loss: 0.006096 [50000/60000]
loss: 0.001458 [55000/60000]
Test Error:
Accuracy: 98.9%, Avg loss: 0.034399
Epoch 5
-------------------------------
loss: 0.001917 [ 0/60000]
loss: 0.004953 [ 5000/60000]
loss: 0.039799 [10000/60000]
loss: 0.000086 [15000/60000]
loss: 0.087789 [20000/60000]
loss: 0.006337 [25000/60000]
loss: 0.046358 [30000/60000]
loss: 0.020176 [35000/60000]
loss: 0.009923 [40000/60000]
loss: 0.008083 [45000/60000]
loss: 0.007701 [50000/60000]
loss: 0.002184 [55000/60000]
Test Error:
Accuracy: 99.0%, Avg loss: 0.034642
Done!
Saved PyTorch Model State to model.pth
5. 保存和加载模型
下面的代码是保存训练好的模型的参数:
torch.save(model.state_dict(), "model.pth")
print("Saved PyTorch Model State to model.pth")
而加载模型的语法则如下所示:
model = CNN()
model.load_state_dict(torch.load("model.pth"))
将训练好的模型去预测test数据集上的种类标签,完整代码如下所示:
def load_model():
model = CNN()
model.load_state_dict(torch.load("model.pth"))
classes = [
"T-shirt/top",
"Trouser",
"Pullover",
"Dress",
"Coat",
"Sandal",
"Shirt",
"Sneaker",
"Bag",
"Ankle boot",
]
model.eval()
inputs, labels = next(iter(test_loader))
with torch.no_grad():
prediction = model(inputs)
accuracy = 0
for i in range(len(prediction)):
predicted, actual = classes[prediction[i].argmax(0)], classes[labels[i]]
print(f'Predicted: "{predicted}", Actual: "{actual}"')
if predicted == actual:
accuracy += 1
print("accuracy:%.4f" % (accuracy / len(prediction)))
if __name__ == '__main__':
load_model()
预测的结果如下所示,准确率为98%:
Predicted: "Sneaker", Actual: "Sneaker"
Predicted: "Pullover", Actual: "Pullover"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "T-shirt/top", Actual: "T-shirt/top"
Predicted: "Coat", Actual: "Coat"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Coat", Actual: "Coat"
Predicted: "Ankle boot", Actual: "Ankle boot"
Predicted: "Sandal", Actual: "Sandal"
Predicted: "Ankle boot", Actual: "Ankle boot"
Predicted: "T-shirt/top", Actual: "T-shirt/top"
Predicted: "Shirt", Actual: "Shirt"
Predicted: "Ankle boot", Actual: "Ankle boot"
Predicted: "T-shirt/top", Actual: "T-shirt/top"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Sandal", Actual: "Sandal"
Predicted: "Ankle boot", Actual: "Ankle boot"
Predicted: "Sneaker", Actual: "Sneaker"
Predicted: "Sandal", Actual: "Dress"
Predicted: "Coat", Actual: "Coat"
Predicted: "Ankle boot", Actual: "Ankle boot"
Predicted: "Shirt", Actual: "Shirt"
Predicted: "Shirt", Actual: "Shirt"
Predicted: "Sandal", Actual: "Sandal"
Predicted: "Coat", Actual: "Coat"
Predicted: "T-shirt/top", Actual: "T-shirt/top"
Predicted: "Sneaker", Actual: "Sneaker"
Predicted: "Coat", Actual: "Coat"
Predicted: "T-shirt/top", Actual: "T-shirt/top"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Dress", Actual: "Dress"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Dress", Actual: "Dress"
Predicted: "Coat", Actual: "Coat"
Predicted: "Sneaker", Actual: "Sneaker"
Predicted: "Pullover", Actual: "Pullover"
Predicted: "Sneaker", Actual: "Sneaker"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Pullover", Actual: "Pullover"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Sneaker", Actual: "Sneaker"
Predicted: "Coat", Actual: "Coat"
Predicted: "Pullover", Actual: "Pullover"
Predicted: "Dress", Actual: "Dress"
Predicted: "Sandal", Actual: "Sandal"
Predicted: "Trouser", Actual: "Trouser"
Predicted: "Pullover", Actual: "Pullover"
Predicted: "Coat", Actual: "Coat"
Predicted: "Coat", Actual: "Coat"
accuracy:0.9800
6. 测试全部测试集的内容
python代码:
def load_model():
model = CNN()
model.load_state_dict(torch.load("model.pth"))
classes = [
"T-shirt/top",
"Trouser",
"Pullover",
"Dress",
"Coat",
"Sandal",
"Shirt",
"Sneaker",
"Bag",
"Ankle boot",
]
model.eval()
with torch.no_grad():
accuracy = 0
for (images, labels) in test_loader:
predict = model(images)
for i in range(len(predict)):
predicted, actual = classes[predict[i].argmax(0)], classes[labels[i]]
print(f'Predicted: "{predicted}", Actual: "{actual}"')
if predicted == actual:
accuracy += 1
print("accuracy:%.4f" % (accuracy / len(test_data.data)))
print('num:%d' % accuracy)
if __name__ == '__main__':
load_model()
测试结果:
...
accuracy:0.9708
num:9708
这里就不全部展开打印出来的结果了,因为总共有10000条数据,只展示最后的结果。