CNN 构成模块

1import torch
2import cv2
3
4weight = torch.zeros(1, 3, 5, 5)  # (N_kernels, N_channels, W, H)
5bias = torch.zeros(1)
6weight[0, 0, :, 0: 2] = 90   # 左侧绿色通道条带
7weight[0, 1, :, 0: 2] = 120
8weight[0, 2, :, 0: 2] = 35
9weight[0, 0, :, 2:] = 75  # 右侧黄色通道条带
10weight[0, 1, :, 2:] = 160
11weight[0, 2, :, 2:] = 160
12
13cv2.imwrite("./output/0_kernel_weight.png", cv2.resize(torch.moveaxis(weight.squeeze(0), 0, 2).int().numpy(), dsize=(300, 300), interpolation=cv2.INTER_NEAREST))
14# 标准化权重
15weight = (weight - torch.mean(weight))/torch.std(weight)
16bias[0] = 0.

1# ... 续上
2# 使用设计好的权重和偏置创建一个二维卷积层
3conv = torch.nn.Conv2d(3, 1, kernel_size=(5, 5), padding='same')
4# 张量必须包装为 torch.nn.Parameter 对象
5conv.weight = torch.nn.Parameter(weight)
6conv.bias = torch.nn.Parameter(bias)
7
8# 加载图像
9original_img = cv2.imread("./images/electronics/rpi_back.jpg")
10cv2.imwrite("./output/1_original.png", original_img)
11
12# 将图像转换为张量
13input_tsr = torch.from_numpy(original_img).float()/255.0  # (H, W, C)
14input_tsr = torch.moveaxis(input_tsr, 2, 0)  # (C, H, W)
15
16# 在输入张量上执行卷积运算
17convolution_tsr = conv(input_tsr)
18
19# 保存卷积结果图像
20convolution_img = convolution_tsr.squeeze(0).detach().numpy()  # (H, W)
21cv2.imwrite("./output/2_convolution.png", 127 + 10 * convolution_img)

1import torch
2
3input_tsr = torch.randn(1, 6, 2, 2)  # (N, C, H, W)
4print(f"input_tsr = {input_tsr}")
5dropout2d = torch.nn.Dropout2d(p=0.5)  # p 表示将被丢弃的通道比例
6print(f"dropout2d.training = {dropout2d.training}")
7output_tsr = dropout2d(input_tsr)  # 将输入张量传入 Dropout2d 层
8print(f"output_tsr = {output_tsr}")

1import torch
2
3input_tsr = torch.randn(1, 6, 2, 2)  # (N, C, H, W)
4print(f"input_tsr = {input_tsr}")
5dropout2d = torch.nn.Dropout2d(p=0.5)
6dropout2d.training = False
7output_tsr = dropout2d(input_tsr)
8print(f"output_tsr = {output_tsr}")

1import torch
2
3input_tsr = torch.randn(3, 1, 2, 2)
4print(f"input_tsr = {input_tsr}")
5batch_norm = torch.nn.BatchNorm2d(num_features=1)  # num_features 是输入张量的通道数
6output_tsr = batch_norm(input_tsr)
7print(f"output_tsr = {output_tsr}")

1import torch
2
3input_tsr = torch.randn(1, 1, 4, 4)
4print(f"input_tsr = \n{input_tsr}")
5maxpool2d = torch.nn.MaxPool2d(kernel_size=(2, 2))
6output_tsr = maxpool2d(input_tsr)
7print(f"output_tsr = \n{output_tsr}")

1import torch
2
3input_tsr = torch.randn(3, 1, 2, 2)
4print(f"input_tsr = {input_tsr}")
5print(f"input_tsr.shape = {input_tsr.shape}")
6output_tsr = input_tsr.view(-1, 1 * 2 * 2)  # 第一个参数为 -1 时，
7    # 输出维度会自动调整以保持元素总数不变
8print(f"output_tsr = {output_tsr}")
9print(f"output_tsr.shape = {output_tsr.shape}")