使用TensorFlow时遇ValueError：Conv2D输入维度不兼容

错误信息 #

Input 0 of layer "conv2d_1" is incompatible with the layer: expected min_ndim=4, found ndim=2. Full shape received: (None, 6).

数据与模型背景 #

• 输入数据：
  • img_array：形状(332240, 6, 160, 160)，共332240个样本，每个样本包含6张160×160的单通道图像
  • value_array：形状(332240, 6)，每个样本对应6维辅助数值数据
  • target_array：形状(332240, 1)，回归任务的目标输出
• 模型目标：对每个样本的6张图像分别做卷积处理，展平后与辅助数值数据拼接，再通过全连接层完成回归。

错误原因 #

模型定义时创建了6个独立的Input(shape=(160,160,1))作为图像输入，但训练时传入的是一个包含6张图的张量img_array，导致模型误将辅助数据value_array传入了第一个卷积层。而卷积层Conv2D要求输入为4维张量（批量大小 + 高 + 宽 + 通道数），但实际收到的是2维的辅助数据(None,6)，因此触发维度不兼容错误。

同时代码存在两处语法错误：

• image_inputs = [Input(shape=(160,160,1) for _ in range(6)] 缺少右括号，应改为image_inputs = [Input(shape=(160,160,1)) for _ in range(6)]
• array_input = Input(shape=(6,) 缺少右括号，应改为array_input = Input(shape=(6,))

修正方案 #

方案1：拆分输入数据适配原模型结构 #

将img_array拆分为6个独立的图像张量，每个张量形状为(332240,160,160,1)，再与辅助数据一起传入模型训练：

import numpy as np
from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, Flatten, Concatenate, Dense
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import SGD

# 拆分图像输入
img_inputs = [img_array[:, i, :, :, np.newaxis] for i in range(6)]

# 修正语法错误后的原模型结构
image_inputs = [Input(shape=(160,160,1)) for _ in range(6)]
array_input = Input(shape=(6,))

conv_outputs = []
for imgs in image_inputs:
    x = Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform')(imgs)
    x = MaxPooling2D((2, 2))(x)
    x = Flatten()(x)
    conv_outputs.append(x)

conv_concat = Concatenate()(conv_outputs)
concatenated_inputs = Concatenate()([conv_concat, array_input])

for units in (100,10):
    concatenated_inputs = Dense(units, activation='relu')(concatenated_inputs)

output_layer = Dense(1, activation='linear')(concatenated_inputs)
final_model = Model(inputs=[image_inputs, array_input], outputs=output_layer)

# 编译模型（假设optimizer已定义）
if optimizer == 'sgd':
    sgd_optimizer = SGD(learning_rate=0.01, momentum=0.9)
    optimizer = sgd_optimizer
final_model.compile(optimizer=optimizer, loss='mean_squared_error', metrics=['mean_squared_error','mean_absolute_error'])

# 传入拆分后的图像输入和辅助数据
final_model.fit([*img_inputs, value_array], target_array, validation_split=0.2, verbose=1, batch_size=32)

方案2：修改模型结构适配原始输入形状 #

用单个Input接收包含6张图的张量，通过Lambda层拆分处理，代码更简洁高效：

import tensorflow as tf
from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, Flatten, Concatenate, Dense, Lambda
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import SGD

# 定义输入：图像输入为(6,160,160)，添加通道维度后变为(6,160,160,1)
image_input = Input(shape=(6,160,160))
# 添加通道维度
image_input_expanded = Lambda(lambda x: tf.expand_dims(x, axis=-1))(image_input)
# 拆分为6个(160,160,1)的张量
split_images = Lambda(lambda x: tf.split(x, num_or_size_splits=6, axis=1))(image_input_expanded)
# 去掉多余维度，得到6个(None,160,160,1)的张量
split_images = [Lambda(lambda x: tf.squeeze(x, axis=1))(img) for img in split_images]

array_input = Input(shape=(6,))

conv_outputs = []
for imgs in split_images:
    x = Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform')(imgs)
    x = MaxPooling2D((2, 2))(x)
    x = Flatten()(x)
    conv_outputs.append(x)

conv_concat = Concatenate()(conv_outputs)
concatenated_inputs = Concatenate()([conv_concat, array_input])

for units in (100,10):
    concatenated_inputs = Dense(units, activation='relu')(concatenated_inputs)

output_layer = Dense(1, activation='linear')(concatenated_inputs)
final_model = Model(inputs=[image_input, array_input], outputs=output_layer)

# 编译模型
if optimizer == 'sgd':
    sgd_optimizer = SGD(learning_rate=0.01, momentum=0.9)
    optimizer = sgd_optimizer
final_model.compile(optimizer=optimizer, loss='mean_squared_error', metrics=['mean_squared_error','mean_absolute_error'])

# 直接传入原始输入数据
final_model.fit([img_array, value_array], target_array, validation_split=0.2, verbose=1, batch_size=32)

内容的提问来源于stack exchange，提问作者may7even