• 3.pytorch学习：conv2d——2d卷积

    目录

    自建一个tensor理解卷积

    加入图像数据集

    tensorboard查看卷积后的图片

    自建一个tensor理解卷积

    1. import torch
    2. from torch import nn
    3.  
    4. input = torch.tensor([[[1, 2, 0, 3, 1],
    5.                       [0, 1, 2, 3, 1],
    6.                       [1, 2, 1, 0, 0],
    7.                       [5, 2, 3, 1, 1],
    8.                       [2, 1, 0, 1, 1]],
    9.                       [[1, 2, 0, 3, 1],
    10.                       [0, 1, 2, 3, 1],
    11.                       [1, 2, 1, 0, 0],
    12.                       [5, 2, 3, 1, 1],
    13.                       [2, 1, 0, 1, 1]],
    14.                       [[1, 2, 0, 3, 1],
    15.                        [0, 1, 2, 3, 1],
    16.                        [1, 2, 1, 0, 0],
    17.                        [5, 2, 3, 1, 1],
    18.                        [2, 1, 0, 1, 1]]], dtype=torch.float32)
    19. print(input.shape)
    20. input = torch.reshape(input, (-1, 3, 5, 5))
    21. print(input.shape)
    22.  
    23.  
    24. class ZiDingYi(nn.Module):
    25.     def __init__(self):
    26.         super(ZiDingYi, self).__init__()
    27.         self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)
    28.  
    29.     def forward(self, x):
    30.         x = self.conv1(x)
    31.         return x
    32.  
    33.  
    34. zidingyi = ZiDingYi()
    35. print(zidingyi)
    36.  
    37. output = zidingyi(input)
    38. print(output)

    自创一个tensor张量，为了与图像相符合，采用了三通道、高与宽的形式。

    1. input = torch.tensor([[[1, 2, 0, 3, 1],
    2.                       [0, 1, 2, 3, 1],
    3.                       [1, 2, 1, 0, 0],
    4.                       [5, 2, 3, 1, 1],
    5.                       [2, 1, 0, 1, 1]],
    6.                       [[1, 2, 0, 3, 1],
    7.                       [0, 1, 2, 3, 1],
    8.                       [1, 2, 1, 0, 0],
    9.                       [5, 2, 3, 1, 1],
    10.                       [2, 1, 0, 1, 1]],
    11.                       [[1, 2, 0, 3, 1],
    12.                        [0, 1, 2, 3, 1],
    13.                        [1, 2, 1, 0, 0],
    14.                        [5, 2, 3, 1, 1],
    15.                        [2, 1, 0, 1, 1]]], dtype=torch.float32)
    16. print(input.shape)
    17. input = torch.reshape(input, (-1, 3, 5, 5))
    18. print(input.shape)

    注意的点1：

    输入的整数系统默认识别为长整型（long），这种数据类型神经网络不能接收，必须强制转换为浮点型，其中torch下给出了浮点型的数据类型。

    dtype=torch.float32

    注意的点2：

    不知道新增维度的值时可以给-1让系统自己运算。

    input = torch.reshape(input, (-1, 3, 5, 5))

     继承nn.Module父类，利用super()调用父类中的__init__()

    1. class ZiDingYi(nn.Module):
    2.     def __init__(self):
    3.         super(ZiDingYi, self).__init__()
    4.         self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)
    5.  
    6.     def forward(self, x):
    7.         x = self.conv1(x)
    8.         return x

    其中2d卷积的参数，收到的通道数，输出的通道数，卷积核大小（可以给int，也可以给元组,此处给的int），步长（可以给int，也可以给元组，此处给的元组），填充大小。

    forward()函数负责利用__init__()的属性，传入输入值并返回输出值。

    结果：

    1. torch.Size([3, 5, 5])
    2. torch.Size([1, 3, 5, 5])
    3. ZiDingYi(
    4.   (conv1): Conv2d(3, 6, kernel_size=(3, 3), stride=(1, 1))
    5. )
    6. tensor([[[[ 0.7182,  2.4224,  0.9817],
    7.           [ 1.7839,  2.5734,  1.2106],
    8.           [ 1.8459,  1.5189,  1.4425]],
    9.  
    10.          [[ 0.1448, -0.0430,  0.0554],
    11.           [-0.7607, -0.3164, -1.1233],
    12.           [ 0.9714,  0.2079,  1.0055]],
    13.  
    14.          [[-0.4258, -0.7102, -0.6917],
    15.           [-0.3210, -0.1598, -0.8339],
    16.           [-2.5572, -1.2985, -1.3247]],
    17.  
    18.          [[ 0.4376, -0.4646, -1.1285],
    19.           [ 0.1640,  0.0108,  0.2103],
    20.           [-0.5246, -0.5199,  0.1915]],
    21.  
    22.          [[ 1.2390,  1.2742,  1.0452],
    23.           [ 0.5466,  0.6744,  1.8741],
    24.           [-0.0987, -0.1915, -1.2585]],
    25.  
    26.          [[-0.7630,  0.5245, -0.3805],
    27.           [-0.2918,  0.1074, -0.2167],
    28.           [ 0.1180,  0.4456,  0.2908]]]], grad_fn=)
    29.  
    30. Process finished with exit code 0

    观察结果符合特征图大小计算公式：

    输出高=（(输入高+2×填充大小-卷积核大小）/步长大小）+1

    加入图像数据集

    1. import torch
    2. import torchvision
    3. from torch import nn
    4. from torch.utils.data import DataLoader
    5.  
    6. input = torch.tensor([[[1, 2, 0, 3, 1],
    7.                       [0, 1, 2, 3, 1],
    8.                       [1, 2, 1, 0, 0],
    9.                       [5, 2, 3, 1, 1],
    10.                       [2, 1, 0, 1, 1]],
    11.                       [[1, 2, 0, 3, 1],
    12.                       [0, 1, 2, 3, 1],
    13.                       [1, 2, 1, 0, 0],
    14.                       [5, 2, 3, 1, 1],
    15.                       [2, 1, 0, 1, 1]],
    16.                       [[1, 2, 0, 3, 1],
    17.                        [0, 1, 2, 3, 1],
    18.                        [1, 2, 1, 0, 0],
    19.                        [5, 2, 3, 1, 1],
    20.                        [2, 1, 0, 1, 1]]], dtype=torch.float32)
    21. print(input.shape)
    22. input = torch.reshape(input, (-1, 3, 5, 5))
    23. print(input.shape)
    24.  
    25.  
    26. class ZiDingYi(nn.Module):
    27.     def __init__(self):
    28.         super(ZiDingYi, self).__init__()
    29.         self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)
    30.  
    31.     def forward(self, x):
    32.         x = self.conv1(x)
    33.         return x
    34.  
    35.  
    36. zidingyi = ZiDingYi()
    37. print(zidingyi)
    38.  
    39. output = zidingyi(input)
    40. print(output)
    41.  
    42. dataset_transforms = torchvision.transforms.Compose([torchvision.transforms.ToTensor(),
    43.                                                      torchvision.transforms.Normalize((0.4915, 0.4823, 0.4468),
    44.                                                                           (0.2470, 0.2435, 0.2616))])
    45. train_dataset = torchvision.datasets.CIFAR10(root="./dataset", train=True, download=True, transform=dataset_transforms)
    46. train_dataloader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=True, num_workers=0, drop_last=False)
    47.  
    48. for data in train_dataloader:
    49.     imgs, labels = data
    50.     output2 = zidingyi(imgs)
    51.     print(imgs.shape)
    52.     print(output2.shape)

    结果：

    1. torch.Size([3, 5, 5])
    2. torch.Size([1, 3, 5, 5])
    3. ZiDingYi(
    4.   (conv1): Conv2d(3, 6, kernel_size=(3, 3), stride=(1, 1))
    5. )
    6. tensor([[[[-0.3747,  1.3237, -0.4478],
    7.           [ 0.7121,  1.4943,  0.4299],
    8.           [ 0.4241, -0.6195, -0.0987]],
    9.  
    10.          [[-0.7136, -1.2816, -1.1319],
    11.           [-1.2081, -1.2827, -0.4157],
    12.           [-0.6310, -0.6770, -0.0811]],
    13.  
    14.          [[-0.8652,  0.0879, -0.4910],
    15.           [-1.2453,  0.2760,  0.0546],
    16.           [-0.8033, -0.5222, -0.3930]],
    17.  
    18.          [[-0.1145,  0.3100, -0.8689],
    19.           [ 1.0125,  0.6360, -0.2982],
    20.           [-0.6484, -1.0439, -0.3103]],
    21.  
    22.          [[-0.3132, -0.5373, -1.3907],
    23.           [ 0.4211, -1.0112, -0.5164],
    24.           [-1.5543, -0.2601, -0.3318]],
    25.  
    26.          [[-0.4129, -1.6505, -0.0648],
    27.           [ 0.1058,  0.1305,  0.3511],
    28.           [ 0.6991, -0.3641,  0.6684]]]], grad_fn=)
    29. Files already downloaded and verified
    30. torch.Size([64, 3, 32, 32])
    31. torch.Size([64, 6, 30, 30])
    32. torch.Size([64, 3, 32, 32])
    33. torch.Size([64, 6, 30, 30])
    34. torch.Size([64, 3, 32, 32])
    35. torch.Size([64, 6, 30, 30])
    36. torch.Size([64, 3, 32, 32])
    37. torch.Size([64, 6, 30, 30])
    38. ————————————后面省略——————————————

     其中64为批处理大小，通道数由3转化为6，特征图从32高宽转为30高宽，符合计算公式。

    1. torch.Size([64, 3, 32, 32])
    2. torch.Size([64, 6, 30, 30])

    tensorboard查看卷积后的图片

    1. import torch
    2. import torchvision
    3. from torch import nn
    4. from torch.utils.data import DataLoader
    5. from torch.utils.tensorboard import SummaryWriter
    6.  
    7. input = torch.tensor([[[1, 2, 0, 3, 1],
    8.                       [0, 1, 2, 3, 1],
    9.                       [1, 2, 1, 0, 0],
    10.                       [5, 2, 3, 1, 1],
    11.                       [2, 1, 0, 1, 1]],
    12.                       [[1, 2, 0, 3, 1],
    13.                       [0, 1, 2, 3, 1],
    14.                       [1, 2, 1, 0, 0],
    15.                       [5, 2, 3, 1, 1],
    16.                       [2, 1, 0, 1, 1]],
    17.                       [[1, 2, 0, 3, 1],
    18.                        [0, 1, 2, 3, 1],
    19.                        [1, 2, 1, 0, 0],
    20.                        [5, 2, 3, 1, 1],
    21.                        [2, 1, 0, 1, 1]]], dtype=torch.float32)
    22. print(input.shape)
    23. input = torch.reshape(input, (-1, 3, 5, 5))
    24. print(input.shape)
    25.  
    26.  
    27. class ZiDingYi(nn.Module):
    28.     def __init__(self):
    29.         super(ZiDingYi, self).__init__()
    30.         self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)
    31.  
    32.     def forward(self, x):
    33.         x = self.conv1(x)
    34.         return x
    35.  
    36.  
    37. zidingyi = ZiDingYi()
    38. print(zidingyi)
    39.  
    40. output = zidingyi(input)
    41. print(output)
    42.  
    43. dataset_transforms = torchvision.transforms.Compose([torchvision.transforms.ToTensor(),
    44.                                                      torchvision.transforms.Normalize((0.4915, 0.4823, 0.4468),
    45.                                                                           (0.2470, 0.2435, 0.2616))])
    46. train_dataset = torchvision.datasets.CIFAR10(root="./dataset", train=True, download=True, transform=dataset_transforms)
    47. train_dataloader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=True, num_workers=0, drop_last=False)
    48.  
    49. writer = SummaryWriter("./logs")
    50. step = 0
    51. for data in train_dataloader:
    52.     imgs, labels = data
    53.     output2 = zidingyi(imgs)
    54.     print(imgs.shape)
    55.     # torch.Size([64, 3, 32, 32])
    56.     print(output2.shape)
    57.     # torch.Size([64, 6, 30, 30])
    58.     # 这个是显示不出来的，需要reshape一下，转化为3通道
    59.     writer.add_images(tag="input", img_tensor=imgs, global_step=step)
    60.     output_reshape = torch.reshape(output2, (-1, 3, 30, 30))
    61.     writer.add_images(tag="output", img_tensor=output_reshape, global_step=step)
    62.     step += 1
    63. writer.close()

    进入Terminal终端

    tensorboard --logdir=logs

    1. (pytorch) D:\project\pytorch_learn>tensorboard --logdir=logs
    2. TensorFlow installation not found - running with reduced feature set.
    3. Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
    4. TensorBoard 2.10.1 at http://localhost:6006/ (Press CTRL+C to quit)

     

    input是归一化后的图片，归一化后灰度值会有负数，显示的时候负数置0，变黑色。

    output是卷积处理后的图片，每个卷积核可以视为一个滤波器。

    其中同一行会有两个形状的图片，那是因为一个图片三个通道，经过卷积操作变成了6个通道，但是由于图片只能三个通道的时候显示出来（为了在tensorboard上观察），故进行了reshape操作：

    output_reshape = torch.reshape(output2, (-1, 3, 30, 30))

    (64,6,30,30)reshape成了(128,3,30,30)

    6通道变成了两个3通道，由两个图片显示出来。

  • 相关阅读:
    S7-200 SMART与ABB ACS580变频器进行MODBUS RTU通信的具体方法示例
    优先级队列（堆）【Java】
    C++算法：二叉树的序列化与反序列化
    初探工厂抽象模式
    简单聊聊 分布式
    C. String Transformation 1(图的思想)
    easy-monitor3.0 nodejs性能监控和分析工具
    IDEA中 lombok不生效解决方法
    win下的开发环境变量记录
    《信息学奥赛一本通·初赛真题解析》
  • 原文地址：https://blog.csdn.net/wzfafabga/article/details/127695823