import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import struct
import torch.optim as optim
from PIL import Image
from matplotlib import pyplot as plt
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.train_images = self.load_images('sample/train-images.idx3-ubyte')
self.train_labels, self.train_labels_vector = self.load_labels('sample/train-labels.idx1-ubyte')
self.test_images = self.load_images('sample/t10k-images.idx3-ubyte')
self.test_labels, self.test_labels_vector = self.load_labels('sample/t10k-labels.idx1-ubyte')
@staticmethod
def load_images(file_name):
with open(file_name, 'rb') as bin_file:
buffers = bin_file.read()
magic, num, rows, cols = struct.unpack_from('>IIII', buffers, 0)
bits = num * rows * cols
images = struct.unpack_from('>' + str(bits) + 'B', buffers, struct.calcsize('>IIII'))
images = np.reshape(images, [num, rows * cols])
images = torch.tensor(images, dtype=torch.float)
return images
@staticmethod
def load_labels(file_name):
with open(file_name, 'rb') as bin_file:
buffers = bin_file.read()
magic, num = struct.unpack_from('>II', buffers, 0)
labels = struct.unpack_from('>' + str(num) + "B", buffers, struct.calcsize('>II'))
labels = np.reshape(labels, [num])
vector = list()
for label in labels:
output = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
output[label] = 1
vector.append(output)
vector = torch.tensor(vector, dtype=torch.float)
labels = torch.tensor(labels, dtype=torch.float)
return labels, vector
def show(self, offset=0):
test_output = torch.max(self(torch.tensor(self.test_images[0 + offset:30 + offset], dtype=torch.float)), 1)
fig = plt.figure(figsize=(8, 8))
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
for i in range(30):
images = np.reshape(self.test_images[i + offset], [28, 28])
ax = fig.add_subplot(6, 5, i + 1, xticks=[], yticks=[])
ax.imshow(images, cmap=plt.cm.binary, interpolation='nearest')
ax.text(0, 7, str(test_output[1].numpy()[i]))
plt.show()
def train_model(self, epoch):
print("Train method undefined!")
pass
def test_accuracy(self):
test_output = torch.max(net(torch.tensor(self.test_images, dtype=torch.float)), 1)
accuracy = (test_output[1].numpy() == self.test_labels.numpy()).sum() / len(self.test_labels)
print(accuracy)
def load_image(self, image_name):
image = Image.open(image_name)
image = np.dot(image, [0.299, 0.587, 0.114])
image = np.ones(shape=(28, 28)) * 255 - image # 负片话，让白色为0，黑色为[0,255]
image = torch.tensor(image, dtype=torch.float)
return image
class FullNet(Net):
def __init__(self):
super(FullNet, self).__init__()
self.fc1 = nn.Linear(784, 300)
self.fc2 = nn.Linear(300, 10)
def forward(self, x):
x = F.tanh(self.fc1(x))
x = F.softmax(self.fc2(x))
return x
def train_model(self, epoch, rate=0.1):
criterion = nn.MSELoss()
for i in range(epoch):
output = self(self.train_images)
loss = criterion(output, self.train_labels_vector)
loss.backward()
for f in self.parameters():
f.data.sub_(f.grad.data * rate)
def test(self, image_name):
image = self.load_image(image_name).view(1, 28 * 28)
return torch.max(self(image), 1)[1].numpy()
class CNNNet(Net):
def __init__(self):
super(CNNNet, self).__init__()
# 输入图像channel：1；输出channel：6；5x5卷积核
self.conv1 = nn.Conv2d(1, 6, (5, 5))
self.conv2 = nn.Conv2d(6, 16, (5, 5))
# an affine operation: y = Wx + b
self.fc1 = nn.Linear(16 * 4 * 4, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 10)
@staticmethod
def load_images(file_name):
images = super(CNNNet, CNNNet).load_images(file_name)
return images.view(-1, 1, 28, 28)
def forward(self, x):
# 2x2 Max pooling
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
# 如果是方阵,则可以只使用一个数字进行定义
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = x.view(-1, self.num_flat_features(x))
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def num_flat_features(self, x):
size = x.size()[1:] # 除去批处理维度的其他所有维度
num_features = 1
for s in size:
num_features *= s
return num_features
def train_model(self, epoch):
optimizer = optim.Adam(self.parameters())
criterion = nn.MSELoss()
for i in range(epoch):
optimizer.zero_grad()
output = self(self.train_images)
loss = criterion(output, self.train_labels_vector)
loss.backward()
optimizer.step()
def test(self, image_name):
image = self.load_image(image_name).view(1, 1, 28, 28)
return torch.max(self(image), 1)[1].numpy()
if __name__ == '__main__':
net = torch.load('cc.model')
# net = FullNet()
# net = CNNNet()
# for i in range(10):
# net.train_model(10)
# net.test_accuracy()
print(net.test("test.jpg"))
# net.show()
# torch.save(net, 'cc.model')
