CNN学习

gaohaikuo1 · 发表于 3 天前

import torch
import torch.nn as nn
import torch.nn.functional as F
# 设置随机种子确保可重复性
torch.manual_seed(42)
class SimpleCNN(nn.Module):
def __init__(self):
super(SimpleCNN, self).__init__()
# 手动初始化卷积层参数
self.conv1 = nn.Conv2d(1, 2, kernel_size=2, stride=2, bias=True)
self.conv1.weight.data = torch.tensor([
[[[0.1, 0.2], [0.3, 0.4]]], # 第一个滤波器
[[[0.5, 0.6], [0.7, 0.8]]] # 第二个滤波器
], dtype=torch.float32)
self.conv1.bias.data = torch.tensor([0.1, 0.2], dtype=torch.float32)
# 手动初始化全连接层参数
self.fc = nn.Linear(2, 3, bias=True)
self.fc.weight.data = torch.tensor([[0.1, 0.2],
[0.3, 0.4],
[0.5, 0.6]], dtype=torch.float32)
self.fc.bias.data = torch.tensor([0.1, 0.2, 0.3], dtype=torch.float32)
def forward(self, x):
# 卷积层
conv_out = self.conv1(x)
print(f"1. 卷积层输出: {conv_out.detach().squeeze().numpy().round(4)}")
# ReLU激活
relu_out = F.relu(conv_out)
print(f"2. ReLU激活输出: {relu_out.detach().squeeze().numpy().round(4)}")
# 展平特征图
flattened = relu_out.view(relu_out.size(0), -1)
# 全连接层
fc_out = self.fc(flattened)
print(f"3. 全连接层输出: {fc_out.detach().squeeze().numpy().round(4)}")
return fc_out
# 创建模型
model = SimpleCNN()
# 输入数据 (batch_size=1, channels=1, height=2, width=2)
x = torch.tensor([[[[1.0, 0.0],
[0.0, 1.0]]]])
# 真实标签 (类别索引为2)
y = torch.tensor([1])
for i in range(20):
# 前向传播
print("================= 前向传播 =================!!!!!!!!!!!!!!!!!!!!")
output = model(x)
# 计算损失 (交叉熵损失)
criterion = nn.CrossEntropyLoss()
loss = criterion(output, y)
# 计算softmax概率
probs = F.softmax(output, dim=1)
print(f"4. Softmax概率: {probs.detach().squeeze().numpy().round(4)}")
print(f"5. 损失值: {loss.item():.4f}")
# 反向传播
print("\n================= 反向传播 =================")
model.zero_grad()
loss.backward()
# 打印梯度
print("6. 卷积层权重梯度:")
print(model.conv1.weight.grad.detach().squeeze().numpy().round(4))
print("卷积层偏置梯度:", model.conv1.bias.grad.detach().numpy().round(4))
print("\n7. 全连接层权重梯度:")
print(model.fc.weight.grad.detach().numpy().round(4))
print("全连接层偏置梯度:", model.fc.bias.grad.detach().numpy().round(4))
# 参数更新 (学习率0.1)
print("\n================= 参数更新 (学习率0.1) =================")
with torch.no_grad():
# 更新卷积层参数
model.conv1.weight -= 0.1 * model.conv1.weight.grad
model.conv1.bias -= 0.1 * model.conv1.bias.grad
# 更新全连接层参数
model.fc.weight -= 0.1 * model.fc.weight.grad
model.fc.bias -= 0.1 * model.fc.bias.grad
print("8. 更新后的卷积层权重:")
print(model.conv1.weight.detach().squeeze().numpy().round(4))
print("更新后的卷积层偏置:", model.conv1.bias.detach().numpy().round(4))
print("\n9. 更新后的全连接层权重:")
print(model.fc.weight.detach().numpy().round(4))
print("更新后的全连接层偏置:", model.fc.bias.detach().numpy().round(4))
print("================= 前向传播 =================")
output = model(x)

复制代码

		自动登录	找回密码
密码			立即注册