CV05_深度学习模块之间的缝合教学(1)

发布于:2024-07-11 ⋅ 阅读:(15) ⋅ 点赞:(0)

1.1 在哪里缝

测试文件?(×)

训练文件?(×)

模型文件?(√)

1.2 骨干网络与模块缝合

以Vision Transformer为例,模型文件里有很多类,我们只在最后集大成的那个类里添加模块。

之后后,我们准备好我们要缝合的模块,比如SE Net模块,我们先建立一个测试文件测试能否跑通

import numpy as np
import torch
from torch import nn
from torch.nn import init

class SEAttention(nn.Module):
    # 初始化SE模块,channel为通道数,reduction为降维比率
    def __init__(self, channel=512, reduction=16):
        super().__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)  # 自适应平均池化层,将特征图的空间维度压缩为1x1
        self.fc = nn.Sequential(  # 定义两个全连接层作为激励操作,通过降维和升维调整通道重要性
            nn.Linear(channel, channel // reduction, bias=False),  # 降维,减少参数数量和计算量
            nn.ReLU(inplace=True),  # ReLU激活函数,引入非线性
            nn.Linear(channel // reduction, channel, bias=False),  # 升维,恢复到原始通道数
            nn.Sigmoid()  # Sigmoid激活函数,输出每个通道的重要性系数
        )

    # 权重初始化方法
    def init_weights(self):
        for m in self.modules():  # 遍历模块中的所有子模块
            if isinstance(m, nn.Conv2d):  # 对于卷积层
                init.kaiming_normal_(m.weight, mode='fan_out')  # 使用Kaiming初始化方法初始化权重
                if m.bias is not None:
                    init.constant_(m.bias, 0)  # 如果有偏置项,则初始化为0
            elif isinstance(m, nn.BatchNorm2d):  # 对于批归一化层
                init.constant_(m.weight, 1)  # 权重初始化为1
                init.constant_(m.bias, 0)  # 偏置初始化为0
            elif isinstance(m, nn.Linear):  # 对于全连接层
                init.normal_(m.weight, std=0.001)  # 权重使用正态分布初始化
                if m.bias is not None:
                    init.constant_(m.bias, 0)  # 偏置初始化为0

    # 前向传播方法
    def forward(self, x):
        b, c, _, _ = x.size()  # 获取输入x的批量大小b和通道数c
        y = self.avg_pool(x).view(b, c)  # 通过自适应平均池化层后,调整形状以匹配全连接层的输入
        y = self.fc(y).view(b, c, 1, 1)  # 通过全连接层计算通道重要性,调整形状以匹配原始特征图的形状
        return x * y.expand_as(x)  # 将通道重要性系数应用到原始特征图上,进行特征重新校准

# 示例使用
if __name__ == '__main__':
    input = torch.randn(50, 512, 7, 7)  # 随机生成一个输入特征图
    se = SEAttention(channel=512, reduction=8)  # 实例化SE模块,设置降维比率为8
    output = se(input)  # 将输入特征图通过SE模块进行处理
    print(output.shape)  # 打印处理后的特征图形状,验证SE模块的作用

打印处理后的形状,我们这里要注意,缝合模块时只需要注意第一维,也就是这个channel,要和骨干网络保持一致,只要你把输入输出的通道数对齐,那么这个通道数就可以缝合成功。

把模块复制进骨干网络中:

然后进行缝合,在缝合之前要先测试通道是否匹配,不然肯定报错。

如何验证通道数

我们找到骨干网络前向传播的部分,在你想加入这个模块地方print(x.shape)即可。运行训练文件:

放在最前面:

通道数为3(8为batch size)。

将模块添加进骨干网络

在骨干网络的init函数下添加:(ctrl+p可查看参数)通道数与之前查的对齐。

在前向传播中添加:

看看是否正常运行:

正常运行,说明模块缝合成功!

打印缝合后的模型结构

该操作在模型文件中进行。

VisionTransformer(
  (patch_embed): PatchEmbed(
    (proj): Conv2d(3, 768, kernel_size=(16, 16), stride=(16, 16))
    (norm): Identity()
  )
  (pos_drop): Dropout(p=0.0, inplace=False)
  (blocks): Sequential(
    (0): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (1): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (2): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (3): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (4): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (5): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (6): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (7): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (8): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (9): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (10): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
    (11): Block(
      (norm1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (attn): Attention(
        (qkv): Linear(in_features=768, out_features=2304, bias=True)
        (attn_drop): Dropout(p=0.0, inplace=False)
        (proj): Linear(in_features=768, out_features=768, bias=True)
        (proj_drop): Dropout(p=0.0, inplace=False)
      )
      (drop_path): Identity()
      (norm2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
      (mlp): Mlp(
        (fc1): Linear(in_features=768, out_features=3072, bias=True)
        (act): GELU()
        (fc2): Linear(in_features=3072, out_features=768, bias=True)
        (drop): Dropout(p=0.0, inplace=False)
      )
    )
  )
  (norm): LayerNorm((768,), eps=1e-06, elementwise_affine=True)
  (pre_logits): Sequential(
    (fc): Linear(in_features=768, out_features=768, bias=True)
    (act): Tanh()
  )
  (head): Linear(in_features=768, out_features=21843, bias=True)
  (se): SEAttention(
    (avg_pool): AdaptiveAvgPool2d(output_size=1)
    (fc): Sequential(
      (0): Linear(in_features=3, out_features=0, bias=False)
      (1): ReLU(inplace=True)
      (2): Linear(in_features=0, out_features=3, bias=False)
      (3): Sigmoid()
    )
  )
)

我们可以看到多了一个SEAttention,说明模块缝合进去了!

1.3 模块之间缝合

以SENet和ECA模块为例。

串联模块

方式1

同1.2。照猫画虎。(注意通道数保持一致)

打印模型结构:

ECAAttention(
  (gap): AdaptiveAvgPool2d(output_size=1)
  (conv): Conv1d(1, 1, kernel_size=(3,), stride=(1,), padding=(1,))
  (sigmoid): Sigmoid()
  (se): SEAttention(
    (avg_pool): AdaptiveAvgPool2d(output_size=1)
    (fc): Sequential(
      (0): Linear(in_features=64, out_features=4, bias=False)
      (1): ReLU(inplace=True)
      (2): Linear(in_features=4, out_features=64, bias=False)
      (3): Sigmoid()
   )))

 方式2

我们定义一个串联函数,将模块之间串联起来:

实例化查看一下模型结构

输出结果:

torch.Size([1, 63, 64, 64]) torch.Size([1, 63, 64, 64])
Cascade(
  (se): SEAttention(
    (avg_pool): AdaptiveAvgPool2d(output_size=1)
    (fc): Sequential(
      (0): Linear(in_features=63, out_features=3, bias=False)
      (1): ReLU(inplace=True)
      (2): Linear(in_features=3, out_features=63, bias=False)
      (3): Sigmoid()
    )
  )
  (eca): ECAAttention(
    (gap): AdaptiveAvgPool2d(output_size=1)
    (conv): Conv1d(1, 1, kernel_size=(63,), stride=(1,), padding=(31,))
    (sigmoid): Sigmoid()
  )
)

并联模块

对于并联模块,方法有很多种,两个两个模块输出的张量可以:

(1)逐元素相加(2)逐元素相乘(3)concat拼接(4)等等

输出结果:

torch.Size([1, 63, 64, 64]) torch.Size([1, 126, 64, 64])
Cascade(
  (se): SEAttention(
    (avg_pool): AdaptiveAvgPool2d(output_size=1)
    (fc): Sequential(
      (0): Linear(in_features=63, out_features=3, bias=False)
      (1): ReLU(inplace=True)
      (2): Linear(in_features=3, out_features=63, bias=False)
      (3): Sigmoid()
    )
  )
  (eca): ECAAttention(
    (gap): AdaptiveAvgPool2d(output_size=1)
    (conv): Conv1d(1, 1, kernel_size=(63,), stride=(1,), padding=(31,))
    (sigmoid): Sigmoid()
  )
)

1.4 思考 

我们不要拘泥于只串联获并联,可以将二者结合,多个模块中,部分模块并联后又与其他模块串联,等等。。这种排列组合之后,总会有一个你想要的模型!!!