在DenseNet201中添加DeformConv2D后出现维度不匹配错误求解决

希望构建修改版DenseNet201模型，在conv1/conv层后添加DeformConv2D（可变形卷积）层，其余层保持不变。编写的代码如下：

import tensorflow as tf
from tensorflow.keras.applications import DenseNet201
from tensorflow.keras.layers import Layer, Input
from tensorflow.keras.models import Model

class DeformConv2D(Layer):
    # Placeholder for the actual implementation
    def __init__(self, out_channels, kernel_size, stride=1, padding='same', **kwargs):
        super(DeformConv2D, self).__init__(**kwargs)
        # Initialize your DeformConv2D layer here

    def call(self, inputs):
        # Implement the forward pass of your DeformConv2D layer here
        return inputs  # Placeholder return

class ModifiedDenseNet201(Model):
    def __init__(self, **kwargs):
        super(ModifiedDenseNet201, self).__init__(**kwargs)
        self.base_model = DenseNet201(include_top=False, input_shape=(224, 224, 3))
        self.deform_conv2d = DeformConv2D(out_channels=64, kernel_size=(3, 3), stride=1, padding='same')
        
        # Freeze the layers of the base model
        for layer in self.base_model.layers:
            layer.trainable = False

    def call(self, inputs):
        x = inputs
        for layer in self.base_model.layers:
            if layer.name == 'conv1/conv':
                x = layer(x)
                x = self.deform_conv2d(x)  # Insert the DeformConv2D layer after 'conv1/conv'
            else:
                x = layer(x)
        return x

# Create an instance of the modified model
input_tensor = Input(shape=(224, 224, 3))
modified_densenet201 = ModifiedDenseNet201()
output = modified_densenet201(input_tensor)

# Create the modified model
model = Model(inputs=input_tensor, outputs=output)

# Summary to verify the model structure
model.summary()

运行时出现错误：

ValueError: Exception encountered when calling layer 'conv2_block1_concat' (type Concatenate).
    
    A merge layer should be called on a list of inputs. Received: inputs=Tensor("modified_dense_net201_1/conv2_block1_2_conv/Conv2D:0", shape=(None, 56, 56, 32), dtype=float32) (not a list of tensors)
    
    Call arguments received by layer 'conv2_block1_concat' (type Concatenate):
      • inputs=tf.Tensor(shape=(None, 56, 56, 32), dtype=float32)


Call arguments received by layer "modified_dense_net201_1" (type ModifiedDenseNet201):
  • inputs=tf.Tensor(shape=(None, 224, 224, 3), dtype=float32)

错误显示concat层输入格式异常，需修改代码解决问题，实现预期模型架构。

错误根源是直接遍历DenseNet的层并逐个调用时，破坏了原模型中Concatenate层依赖的多输入分支结构——原DenseNet的concat层接收来自多个分支的张量列表，但逐个调用层时无法保留这种分支输入关系，导致concat层只收到单个张量而非列表。

需修改以下核心部分：

• 放弃遍历所有层的方式，改为截取原模型到conv1/conv的输出，插入DeformConv2D后，再接上原模型剩余部分
• 确保DeformConv2D的输出通道数与conv1/conv一致，避免后续层维度不匹配
• 特殊处理Concatenate层，手动维护其输入张量列表，替换为经过可变形卷积后的张量

修改后的完整代码：

import tensorflow as tf
from tensorflow.keras.applications import DenseNet201
from tensorflow.keras.layers import Layer, Input
from tensorflow.keras.models import Model

class DeformConv2D(Layer):
    # 可变形卷积占位实现，后续替换为实际逻辑
    def __init__(self, out_channels, kernel_size, stride=1, padding='same', **kwargs):
        super(DeformConv2D, self).__init__(**kwargs)
        # 这里可添加实际可变形卷积的参数初始化，比如偏移量预测卷积
        self.conv = tf.keras.layers.Conv2D(out_channels, kernel_size, stride=stride, padding=padding)

    def call(self, inputs):
        # 示例实现，替换为实际可变形卷积前向逻辑
        return self.conv(inputs)

def build_modified_densenet201():
    # 加载基础DenseNet201，不包含顶层
    base_model = DenseNet201(include_top=False, input_shape=(224, 224, 3))
    
    # 冻结基础模型参数
    for layer in base_model.layers:
        layer.trainable = False
    
    # 获取conv1/conv层的输出张量
    conv1_output = base_model.get_layer('conv1/conv').output
    
    # 插入DeformConv2D层，输出通道数与conv1/conv保持一致（DenseNet201的conv1/conv输出为64通道）
    deform_output = DeformConv2D(out_channels=64, kernel_size=(3, 3), stride=1, padding='same')(conv1_output)
    
    # 从conv1/conv的下一层（conv1/bn）开始，重构后续模型结构
    x = deform_output
    start_process = False
    for layer in base_model.layers:
        if layer.name == 'conv1/bn':
            start_process = True
            x = layer(x)
            continue
        
        if not start_process:
            continue
        
        # 特殊处理Concatenate层：替换原输入中来自conv1/conv的张量为deform_output
        if isinstance(layer, tf.keras.layers.Concatenate):
            concat_inputs = []
            for node in layer._inbound_nodes:
                for input_tensor in node.input_tensors:
                    if input_tensor == conv1_output:
                        concat_inputs.append(deform_output)
                    else:
                        concat_inputs.append(input_tensor)
            x = layer(concat_inputs)
        else:
            x = layer(x)
    
    # 构建完整模型
    model = Model(inputs=base_model.input, outputs=x)
    return model

# 创建修改后的模型并打印结构
modified_model = build_modified_densenet201()
modified_model.summary()

关键修改说明 #

1. 模型拼接逻辑优化：通过截取原模型中间层输出、插入自定义层后再接后续层的方式，保留原模型的分支结构
2. Concatenate层适配：手动维护concat层的输入列表，将原本来自conv1/conv的输入替换为经过可变形卷积后的张量，确保concat层接收正确的输入格式
3. 通道数一致性：DeformConv2D输出通道数设为64，与原conv1/conv输出通道数匹配，避免后续层维度不兼容

内容的提问来源于stack exchange，提问作者shel coop