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

import torch
from torch import nn

input = torch.tensor([[[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                       [0, 1, 2, 3, 1],
                       [1, 2, 1, 0, 0],
                       [5, 2, 3, 1, 1],
                       [2, 1, 0, 1, 1]]], dtype=torch.float32)
print(input.shape)
input = torch.reshape(input, (-1, 3, 5, 5))
print(input.shape)


class ZiDingYi(nn.Module):
    def __init__(self):
        super(ZiDingYi, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)

    def forward(self, x):
        x = self.conv1(x)
        return x


zidingyi = ZiDingYi()
print(zidingyi)

output = zidingyi(input)
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。

input = torch.tensor([[[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                       [0, 1, 2, 3, 1],
                       [1, 2, 1, 0, 0],
                       [5, 2, 3, 1, 1],
                       [2, 1, 0, 1, 1]]], dtype=torch.float32)
print(input.shape)
input = torch.reshape(input, (-1, 3, 5, 5))
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__()

class ZiDingYi(nn.Module):
    def __init__(self):
        super(ZiDingYi, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)

    def forward(self, x):
        x = self.conv1(x)
        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：

torch.Size([3, 5, 5])
torch.Size([1, 3, 5, 5])
ZiDingYi(
  (conv1): Conv2d(3, 6, kernel_size=(3, 3), stride=(1, 1))
)
tensor([[[[ 0.7182,  2.4224,  0.9817],
          [ 1.7839,  2.5734,  1.2106],
          [ 1.8459,  1.5189,  1.4425]],

         [[ 0.1448, -0.0430,  0.0554],
          [-0.7607, -0.3164, -1.1233],
          [ 0.9714,  0.2079,  1.0055]],

         [[-0.4258, -0.7102, -0.6917],
          [-0.3210, -0.1598, -0.8339],
          [-2.5572, -1.2985, -1.3247]],

         [[ 0.4376, -0.4646, -1.1285],
          [ 0.1640,  0.0108,  0.2103],
          [-0.5246, -0.5199,  0.1915]],

         [[ 1.2390,  1.2742,  1.0452],
          [ 0.5466,  0.6744,  1.8741],
          [-0.0987, -0.1915, -1.2585]],

         [[-0.7630,  0.5245, -0.3805],
          [-0.2918,  0.1074, -0.2167],
          [ 0.1180,  0.4456,  0.2908]]]], grad_fn=)

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

import torch
import torchvision
from torch import nn
from torch.utils.data import DataLoader

input = torch.tensor([[[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                       [0, 1, 2, 3, 1],
                       [1, 2, 1, 0, 0],
                       [5, 2, 3, 1, 1],
                       [2, 1, 0, 1, 1]]], dtype=torch.float32)
print(input.shape)
input = torch.reshape(input, (-1, 3, 5, 5))
print(input.shape)


class ZiDingYi(nn.Module):
    def __init__(self):
        super(ZiDingYi, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)

    def forward(self, x):
        x = self.conv1(x)
        return x


zidingyi = ZiDingYi()
print(zidingyi)

output = zidingyi(input)
print(output)

dataset_transforms = torchvision.transforms.Compose([torchvision.transforms.ToTensor(),
                                                     torchvision.transforms.Normalize((0.4915, 0.4823, 0.4468),
                                                                          (0.2470, 0.2435, 0.2616))])
train_dataset = torchvision.datasets.CIFAR10(root="./dataset", train=True, download=True, transform=dataset_transforms)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=True, num_workers=0, drop_last=False)

for data in train_dataloader:
    imgs, labels = data
    output2 = zidingyi(imgs)
    print(imgs.shape)
    print(output2.shape)

结果：

torch.Size([3, 5, 5])
torch.Size([1, 3, 5, 5])
ZiDingYi(
  (conv1): Conv2d(3, 6, kernel_size=(3, 3), stride=(1, 1))
)
tensor([[[[-0.3747,  1.3237, -0.4478],
          [ 0.7121,  1.4943,  0.4299],
          [ 0.4241, -0.6195, -0.0987]],

         [[-0.7136, -1.2816, -1.1319],
          [-1.2081, -1.2827, -0.4157],
          [-0.6310, -0.6770, -0.0811]],

         [[-0.8652,  0.0879, -0.4910],
          [-1.2453,  0.2760,  0.0546],
          [-0.8033, -0.5222, -0.3930]],

         [[-0.1145,  0.3100, -0.8689],
          [ 1.0125,  0.6360, -0.2982],
          [-0.6484, -1.0439, -0.3103]],

         [[-0.3132, -0.5373, -1.3907],
          [ 0.4211, -1.0112, -0.5164],
          [-1.5543, -0.2601, -0.3318]],

         [[-0.4129, -1.6505, -0.0648],
          [ 0.1058,  0.1305,  0.3511],
          [ 0.6991, -0.3641,  0.6684]]]], grad_fn=)
Files already downloaded and verified
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
torch.Size([64, 3, 32, 32])
torch.Size([64, 6, 30, 30])
————————————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。

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

tensorboard查看卷积后的图片

import torch
import torchvision
from torch import nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

input = torch.tensor([[[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                      [0, 1, 2, 3, 1],
                      [1, 2, 1, 0, 0],
                      [5, 2, 3, 1, 1],
                      [2, 1, 0, 1, 1]],
                      [[1, 2, 0, 3, 1],
                       [0, 1, 2, 3, 1],
                       [1, 2, 1, 0, 0],
                       [5, 2, 3, 1, 1],
                       [2, 1, 0, 1, 1]]], dtype=torch.float32)
print(input.shape)
input = torch.reshape(input, (-1, 3, 5, 5))
print(input.shape)


class ZiDingYi(nn.Module):
    def __init__(self):
        super(ZiDingYi, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=(1, 1), padding=0)

    def forward(self, x):
        x = self.conv1(x)
        return x


zidingyi = ZiDingYi()
print(zidingyi)

output = zidingyi(input)
print(output)

dataset_transforms = torchvision.transforms.Compose([torchvision.transforms.ToTensor(),
                                                     torchvision.transforms.Normalize((0.4915, 0.4823, 0.4468),
                                                                          (0.2470, 0.2435, 0.2616))])
train_dataset = torchvision.datasets.CIFAR10(root="./dataset", train=True, download=True, transform=dataset_transforms)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=True, num_workers=0, drop_last=False)

writer = SummaryWriter("./logs")
step = 0
for data in train_dataloader:
    imgs, labels = data
    output2 = zidingyi(imgs)
    print(imgs.shape)
    # torch.Size([64, 3, 32, 32])
    print(output2.shape)
    # torch.Size([64, 6, 30, 30])
    # 这个是显示不出来的，需要reshape一下，转化为3通道
    writer.add_images(tag="input", img_tensor=imgs, global_step=step)
    output_reshape = torch.reshape(output2, (-1, 3, 30, 30))
    writer.add_images(tag="output", img_tensor=output_reshape, global_step=step)
    step += 1
writer.close()

进入Terminal终端

tensorboard --logdir=logs

(pytorch) D:\project\pytorch_learn>tensorboard --logdir=logs
TensorFlow installation not found - running with reduced feature set.
Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
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。