TensorFlow训练CNN自动编码器时输入与标签形状不匹配问题求助

问题描述 #

在使用TensorFlow构建CNN自动编码器时，输入数据为shape=(419,128,128)的numpy数组sp_pics，运行autoencoder.fit()时出现以下错误：

ValueError: logits and labels must have the same shape, received ((1, 124, 124, 1) vs (1, 128, 128, 1)).

原代码如下：

input_img = Input(shape=(128, 128, 1))

# Encoder
x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling2D((2, 2), padding='same')(x)

# Decoder
x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, (3, 3), activation='relu')(x)
x = UpSampling2D((2, 2))(x)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)

# Define Auto Encoder Model
autoencoder = Model(input_img, decoded)

autoencoder.compile(optimizer='adam', loss='binary_crossentropy')

# Train the Autoencoder
data_input = np.expand_dims(sp_pics, axis=-1)
autoencoder.fit(data_input, data_input, epochs=50, batch_size=32)

encoder = Model(input_img, encoded)
encoded_imgs = encoder.predict(sp_pics[0,:,:])

print(encoded_imgs.shape[1])

问题原因 #

错误中的124来自解码器的尺寸计算偏差：

1. 编码器经过3次带padding='same'的MaxPooling2D，输入尺寸从128依次变为64→32→16，最终encoded的shape为(16,16,8)
2. 解码器前两次UpSampling2D后，尺寸从16恢复为32→64
3. 此时经过未设置padding='same'的Conv2D(3,3)，尺寸会被压缩为64 - 3 + 1 = 62
4. 最后一次UpSampling2D后，尺寸变为62×2=124，导致输出shape为(124,124,1)，与输入的(128,128,1)不匹配，触发报错。

修复方案 #

给解码器中缺失padding='same'的Conv2D层添加该参数，确保每一步卷积后尺寸保持不变，最终输出与输入形状一致。同时修正编码器预测时的维度错误：

import numpy as np
from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D
from tensorflow.keras.models import Model

input_img = Input(shape=(128, 128, 1))

# Encoder
x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling2D((2, 2), padding='same')(x)

# Decoder - 修复Conv2D的padding参数
x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, (3, 3), activation='relu', padding='same')(x)  # 添加padding='same'
x = UpSampling2D((2, 2))(x)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)

# Define Auto Encoder Model
autoencoder = Model(input_img, decoded)

autoencoder.compile(optimizer='adam', loss='binary_crossentropy')

# Train the Autoencoder
data_input = np.expand_dims(sp_pics, axis=-1)
autoencoder.fit(data_input, data_input, epochs=50, batch_size=32)

encoder = Model(input_img, encoded)
# 修正输入维度：扩展为模型要求的4维格式(批量, 高, 宽, 通道)
encoded_imgs = encoder.predict(np.expand_dims(sp_pics[0,:,:], axis=(0, -1)))

print(encoded_imgs.shape[1])  # 输出应为16

内容的提问来源于stack exchange，提问作者Manfred