如何冻结我自己模型的最后一层？

Question

model = ResNet50(weights='imagenet', include_top=False, input_shape=(224, 224, 3))

上面的代码允许我使用 imagenet 的权重，但我想使用我自己的 imagenet 权重，我应该在我的代码中进行哪些更改以允许我只在自己的数据集上训练最后一层？这是我的模型的代码：

def mini_XCEPTION(input_shape, num_classes, l2_regularization=0.01): 正则化 = l2(l2_regularization) ＃ 根据 img_input = 输入（输入形状） x = Conv2D(8, (3, 3), strides=(1, 1), kernel_regularizer=regularization, use_bias=False)(img_input) x = BatchNormalization()(x) x = 激活('relu')(x) x = Conv2D(8, (3, 3), strides=(1, 1), kernel_regularizer=regularization, use_bias=False)(x) x = BatchNormalization()(x) x = 激活('relu')(x) # 模块 1 残差 = Conv2D(16, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) 残差 = BatchNormalization()(残差) x = SeparableConv2D(16, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = 激活('relu')(x) x = SeparableConv2D(16, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x) x = layers.add([x, 残差]) # 模块 2 残差 = Conv2D(32, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) 残差 = BatchNormalization()(残差) x = SeparableConv2D(32, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = 激活('relu')(x) x = SeparableConv2D(32, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x) x = layers.add([x, 残差]) # 模块 3 残差 = Conv2D(64, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) 残差 = BatchNormalization()(残差) x = SeparableConv2D(64, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = 激活('relu')(x) x = SeparableConv2D(64, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x) x = layers.add([x, 残差]) # 模块 4 残差 = Conv2D(128, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) 残差 = BatchNormalization()(残差) x = SeparableConv2D(128, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = 激活('relu')(x) x = SeparableConv2D(128, (3, 3), padding='same', kernel_regularizer=正则化， use_bias=False)(x) x = BatchNormalization()(x) x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x) x = layers.add([x, 残差]) x = Conv2D(num_classes, (3, 3), # kernel_regularizer=正则化， 填充='相同')(x) x = GlobalAveragePooling2D()(x) 输出=激活（'softmax'，名称='预测'）（x） 模型 = 模型（img_input，输出） 返回模型

Answer 1

首先，您为ResNet50 加载保存的权重。之后，您使用相同的架构，并在最后一层：

last_layer=GlobalAveragePooling2D()(x)
last_layer.trainable=True

对于之前的所有层，您都使用trainable=False，例如：

x = Conv2D(8, (3, 3), strides=(1, 1), kernel_regularizer=regularization,
               use_bias=False)(img_input)
x.trainable=False

residual = Conv2D(16, (1, 1), strides=(2, 2),
                      padding='same', use_bias=False)(x)
residual.trainable=False

residual = BatchNormalization()(residual)
residual.trainable=False

x = SeparableConv2D(16, (3, 3), padding='same',
                        kernel_regularizer=regularization,
                        use_bias=False)(x)
x.trainable=False