概述
笔者使用pytorch框架构建神经网络,使用matplotlib显示网络分类过程,生动的演示了网络拟合过程
先看效果图:
最终训练300次后拟合效果:
代码流程
1.环境导入
import matplotlib.pyplot as plt
import numpy as np
import os
import torch
os.environ['KMP_DUPLICATE_LIB_OK']='True' #设置环境变量,避免出现库错误的问题
2.生成数据
batch_size = 20
data_len = 2
max = 80 #数据最大值为80
x1 = np.random.randint(0,max,(batch_size,data_len)) / max #随机生成0~80的数据,并且除以最大值,进行归一化
y1 = np.ones((batch_size,1)) #标签值为1
data1 = np.concatenate((x1,y1),axis=1) #合并数据,将标签与数据合并一起,方面后面随机打乱
x2 = np.random.randint(0,max,(batch_size,data_len)) / max #随机生成0~80的数据,并且除以最大值,进行归一化
y2 = np.zeros((batch_size,1)) #标签值为0
data2 = np.concatenate((x2,y2),axis=1) #合并数据,将标签与数据合并一起,方面后面随机打乱
data = np.concatenate((data1,data2))
np.random.shuffle(data) #将数据打乱
xb = data[:,0:data_len]
xb = torch.tensor(xb,dtype=torch.float)
yb = data[:,data_len:data_len+1]
yb = torch.tensor(yb,dtype=torch.long).reshape(batch_size * 2) #将标签值转成pytorch张量
"""生成测试数据,即生成(80 * 80)的数据,包含所有的数据组合"""
test_data = np.array([0,0]).reshape(1,2)
for x in range(max):
for y in range(max):
xy = np.array([x,y]).reshape(1,2)
test_data = np.append(test_data,xy,axis=0) #合并数据,不使用concatenate,加快速度
test_data = test_data / max
test_data = torch.tensor(test_data,dtype=torch.float)
print(data1)
print(data2)
3.构建网络
定义网络:
因为是二维平面的演示,因此数据维度是2,数据分别对应x坐标,y坐标
因此网络输入维度为(batch_size,2)
class LinearModel(torch.nn.Module):
def __init__(self):
super(LinearModel, self).__init__()
self.nInput = 2
self.fc = torch.nn.Sequential(
torch.nn.Linear(self.nInput, 50),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(50, 50),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(50, 2)
)
def forward(self, x):
x = x.view(x.size(0), -1) # 在进入全连接层之前需要把数据拉直Flatten
x = self.fc(x)
return x
定义网络:
model = LinearModel()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),lr=0.01)
model.train(mode=True)
4.显示拟合过程
训练过程中,使用matplotlib显示拟合过程
plt.ion() #启动交互模式
for epochs in range(400):
optimizer.zero_grad() #梯度清除
outputs = model(xb) #训练
loss = criterion(outputs, yb) #求损失
loss.backward() #求导
optimizer.step() #反向传播,参数更新
# correct = 0
# total = 0
# model.eval()
# with torch.no_grad(): # 下面的代码就不会再计算梯度
# outputs = model(xb)
# loss = criterion(outputs, yb) #求损失
# _, predicted = torch.max(outputs.data, dim=1) # _为每一行的最大值,predicted表示每一行最大值的下标
# total += yb.size(0)
# correct += (predicted == yb).sum().item()
# # if(100 * correct / total == 100):
# # print('--Accuracy on test set:%0.2f %%,loss=%f' % (100 * correct / total,loss))
# acc = (100 * correct / total)
# print('epoch=%d, total=%d, correct=%d, last loss=%f, Accuracy:%0.2f %%' % (epochs,total,correct,loss,acc))
print("epoch:",epochs,"loss",loss.item())
outputs = model(test_data)
_, predicted = torch.max(outputs.data, dim=1) # _为每一行的最大值,predicted表示每一行最大值的下标
plt.cla() #清除显示,避免数据累加,每次画越来越多的数据,加快速度
boolIndex = (predicted == 1) #创建bool索引,方便提取出标签值为1的对应数据
plt.scatter(test_data[boolIndex][:,0],test_data[boolIndex][:,1],color='red')
boolIndex = (predicted == 0)
plt.scatter(test_data[boolIndex][:,0],test_data[boolIndex][:,1],color='blue')
plt.scatter(x1[:,0],x1[:,1],color = 'black')
plt.scatter(x2[:,0],x2[:,1],color = 'yellow')
plt.pause(0.001)
plt.pause(0.001)
plt.ioff()
plt.show()#显示最后拟合结果
plt.close()
最终效果图
附上完整代码,可以拷贝直接运行
import matplotlib.pyplot as plt
import numpy as np
import os
import torch
os.environ['KMP_DUPLICATE_LIB_OK']='True'
class LinearModel(torch.nn.Module):
def __init__(self):#输入图片的宽高
super(LinearModel, self).__init__()
self.nInput = 2
self.fc = torch.nn.Sequential(
torch.nn.Linear(self.nInput, 50),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(50, 50),
torch.nn.ReLU(inplace=True),
torch.nn.Linear(50, 2)
)
def forward(self, x):
x = x.view(x.size(0), -1) # 在进入全连接层之前需要把数据拉直Flatten
x = self.fc(x)
return x
batch_size = 20
data_len = 2
max = 80 #数据最大值为80
x1 = np.random.randint(0,max,(batch_size,data_len)) / max #随机生成0~80的数据,并且除以最大值,进行归一化
y1 = np.ones((batch_size,1))
data1 = np.concatenate((x1,y1),axis=1) #合并数据,将标签与数据合并一起,方面后面随机打乱
x2 = np.random.randint(0,max,(batch_size,data_len)) / max #随机生成0~80的数据,并且除以最大值,进行归一化
y2 = np.zeros((batch_size,1))
data2 = np.concatenate((x2,y2),axis=1) #合并数据,将标签与数据合并一起,方面后面随机打乱
data = np.concatenate((data1,data2))
np.random.shuffle(data) #将数据打乱
xb = data[:,0:data_len]
xb = torch.tensor(xb,dtype=torch.float)
yb = data[:,data_len:data_len+1]
yb = torch.tensor(yb,dtype=torch.long).reshape(batch_size * 2) #将标签值转成pytorch张量
"""生成测试数据,即生成(80 * 80)的数据,包含所有的数据组合"""
test_data = np.array([0,0]).reshape(1,2)
for x in range(max):
for y in range(max):
xy = np.array([x,y]).reshape(1,2)
test_data = np.append(test_data,xy,axis=0) #合并数据,不使用concatenate,加快速度
test_data = test_data / max
test_data = torch.tensor(test_data,dtype=torch.float)
print(data1)
print(data2)
plt.ion()
model = LinearModel()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),lr=0.01)
model.train(mode=True)
for epochs in range(400):
optimizer.zero_grad() #梯度清除
outputs = model(xb) #训练
loss = criterion(outputs, yb) #求损失
loss.backward() #求导
optimizer.step() #反向传播,参数更新
# correct = 0
# total = 0
# model.eval()
# with torch.no_grad(): # 下面的代码就不会再计算梯度
# outputs = model(xb)
# loss = criterion(outputs, yb) #求损失
# _, predicted = torch.max(outputs.data, dim=1) # _为每一行的最大值,predicted表示每一行最大值的下标
# total += yb.size(0)
# correct += (predicted == yb).sum().item()
# # if(100 * correct / total == 100):
# # print('--Accuracy on test set:%0.2f %%,loss=%f' % (100 * correct / total,loss))
# acc = (100 * correct / total)
# print('epoch=%d, total=%d, correct=%d, last loss=%f, Accuracy:%0.2f %%' % (epochs,total,correct,loss,acc))
print("epoch:",epochs,"loss",loss.item())
outputs = model(test_data)
_, predicted = torch.max(outputs.data, dim=1) # _为每一行的最大值,predicted表示每一行最大值的下标
plt.cla()
boolIndex = (predicted == 1)
plt.scatter(test_data[boolIndex][:,0],test_data[boolIndex][:,1],color='red')
boolIndex = (predicted == 0)
plt.scatter(test_data[boolIndex][:,0],test_data[boolIndex][:,1],color='blue')
plt.scatter(x1[:,0],x1[:,1],color = 'black')
plt.scatter(x2[:,0],x2[:,1],color = 'yellow')
plt.pause(0.001)
plt.pause(0.0001)
plt.ioff()
plt.show()
plt.close()
喜欢点个关注吧!