用预训练的resnet实现压缩和激励块

Question

我试图用任何经过预先训练的resnet架构来实现SE块。

self.encoder = torchvision.models.resnet34(pretrained=pretrained)

self.conv1 = nn.Sequential(
    self.encoder.conv1,
    self.encoder.bn1,
    self.encoder.relu
    SeBlock(64,2),
    nn.MaxPool2d(2, 2) )

self.conv2 = nn.Sequential(
    self.encoder.layer1,
    SeBlock(64,2),
    )

self.conv3 = nn.Sequential(
    self.encoder.layer2,
    SeBlock(128,2),
    )

self.conv4 = nn.Sequential(
    self.encoder.layer3,
    SeBlock(256,2),
    )

self.conv5 = nn.Sequential(
    self.encoder.layer4,
    SeBlock(512,2),
    )

以上是我的代码，但不起作用。查看githubs，他们从头开始构建resnet，并归纳SE块，然后为resnet的层加载model.state_dict()权重，并训练剩下的模型。我只需要知道什么是正确的程序使用SE块与预先培训的resnet。？

谢谢！

Answer 1

我不常使用PyTorch，但您可以使用TensorFlow-keras实现相同的功能。这很简单。首先，用您想要的输入大小从keras.applications加载预训练的模型。然后，keras.applications.VGG19(include_top = False, weights = 'imagenet', input_shape = (50, 50, 3)).从模型加载中选择加载训练层。根据您的意愿，在各层之间应用SENET注意。示例：

import tensorflow as tf
import numpy as np
from tensorflow.keras.layers import Activation, Conv2D, Input, BatchNormalization, Reshape, GlobalAveragePooling2D, AveragePooling2D, GlobalMaxPooling2D, Flatten , ReLU, Layer, Dense
from tensorflow.keras.activations import sigmoid, softmax, relu, tanh
from tensorflow.keras import Sequential

class SENET_Attn(Layer):
    """
    Channel Attention Block as reported in SENET
    """
    def __init__(self,out_dim, ratio, layer_name="SENET"):
        super(SENET_Attn, self).__init__()
        self.out_dim = out_dim
        self.ratio = ratio
        self.layer_name = layer_name
            
    def build(self, ratio, layer_name="SENET"):
        self.Global_Average_Pooling = GlobalAveragePooling2D(keepdims= True)
        self.Fully_connected_1_1 = Dense(units= self.out_dim/self.ratio, name=self.layer_name+'_fully_connected1',
                                kernel_initializer="glorot_uniform")
        self.Relu = ReLU()
        self.Fully_connected_2 = Dense(units=self.out_dim, name=layer_name+'_fully_connected2', activation = "tanh")
        self.Sigmoid = Activation("sigmoid")
        
    def call(self, inputs):
        inputs = tf.cast(inputs, dtype = "float32")
        squeeze = self.Global_Average_Pooling(inputs)
        excitation = self.Fully_connected_1_1(squeeze)
        excitation = self.Relu(excitation)
        excitation = self.Fully_connected_2(excitation)
        excitation =  self.Sigmoid(excitation)
        excitation = tf.reshape(excitation, [-1,1,1,self.out_dim])
        
        scale = inputs * excitation
        return scale
Vgg = VGG19(include_top = False,input_shape = (50,50,3))

# SENET-Attn VGG19
ratio =16

input_layer = Input(shape=(50,50,3))
out = Vgg.layers[1](input_layer) # Block one of VGG19
out = Vgg.layers[2](out)
out = Vgg.layers[3](out)

out = Vgg.layers[4](out) # Block two of VGG19
out = Vgg.layers[5](out)
out = Vgg.layers[6](out)

#out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[7](out) # Block three of VGG19
out = Vgg.layers[8](out)
out = Vgg.layers[9](out)
out = Vgg.layers[10](out)
out = Vgg.layers[11](out)
 
out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[12](out) # Block four of VGG19
out = Vgg.layers[13](out)
out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[14](out)
out = Vgg.layers[15](out)
out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[16](out)

out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[17](out) #Block five of VGG19

out = Vgg.layers[18](out)
out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[19](out)
out = Vgg.layers[20](out)
out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
out = Vgg.layers[21](out)

out = SENET_Attn(out.shape[-1], ratio, )(out) # SENET Attention
flatten = Flatten()(out)
out = Dense(100)(flatten)
out = tf.keras.layers.ReLU()(out)
out = Dense(100)(out)
out = tf.keras.layers.ReLU()(out)
out = Dense(2)(out)
out = tf.keras.layers.Softmax()(out)

model = Model(input_layer, out)
model.compile(optimizer= tf.keras.optimizers.SGD(), loss= tf.keras.losses.CategoricalCrossentropy()
              , metrics=['accuracy'])
model.summary()