猫狗分类CNN模型输入层尺寸不兼容错误排查求助

我正在编写简单CNN模型对本地train目录中的猫狗图片分类，当前代码如下：

import numpy as np
import cv2 as cv
import tensorflow.keras as keras
import os
from sklearn.preprocessing import LabelEncoder
from tensorflow.keras.utils import to_categorical
from tensorflow.keras import layers , models
from sklearn.model_selection import train_test_split

images_vector =[]
images_label =[]

fileNames = os.listdir('train')

for i , f_name in enumerate(fileNames) :
    image = cv.imread('train/' + f_name)
    
    image = cv.resize(image , (50,50))
    image = image/255.0
    image = image.flatten()  # 这里把图片展平成一维了

    images_vector.append(image)
    images_label.append(f_name.split('.')[0]) 

    if i%10000 == 0 : 
        print(f" [INFO ] : {i} images are processed...")

labelEncoder  = LabelEncoder()
images_label = labelEncoder.fit_transform(images_label)

images_label = to_categorical(images_label)

# 补全原代码的语法错误
X_train , X_test , y_train , y_test = train_test_split(images_vector ,images_label  , random_state=40 , train_size=0.8)

print('X_train: ' + str(np.array(X_train).shape))
print('Y_train: ' + str(y_train.shape))
print('X_test:  '  + str(np.array(X_test).shape))
print('Y_test:  '  + str(y_test.shape))

运行以下模型构建代码后：

net = models.Sequential([
    layers.Conv2D(32 , (3,3) , activation='relu' , input_shape = (1,7500)) ,
    layers.MaxPooling2D(2,2),
    layers.Conv2D(64 , (3,3) , activation='relu'),
    layers.Flatten(),
    layers.Dense(2 , activation='softmax')
])

net.summary()

出现错误：

ValueError: Input 0 of layer "conv2d_96" is incompatible with the layer: expected min_ndim=4, found ndim=3. Full shape received: (None, 1, 7500)

错误核心是你把图片展平成了一维数组，而CNN需要保留图片的空间维度（高度、宽度、通道数），同时输入形状定义不符合Conv2D层要求。具体修正步骤如下：

1. 保留图片空间维度，取消展平 #

CNN的Conv2D层需要处理4D张量（批量大小, 高度, 宽度, 通道数），因此读取图片后不要用flatten()，保留(50,50,3)的形状（cv.imread读取的是3通道彩色图）：

image = cv.imread('train/' + f_name)
# 跳过损坏图片
if image is None:
    print(f"跳过损坏图片: {f_name}")
    continue
image = cv.resize(image , (50,50))
image = image/255.0  # 仅做归一化，不展平
images_vector.append(image)

2. 将列表转换为numpy数组 #

原代码中images_vector是Python列表，无shape属性，需转为numpy数组才能被Keras正确识别：

images_vector = np.array(images_vector)

3. 修正模型输入形状 #

Conv2D的input_shape需对应单张图片的维度：(高度, 宽度, 通道数)，即(50,50,3)，批量维度由Keras自动添加：

net = models.Sequential([
    layers.Conv2D(32 , (3,3) , activation='relu' , input_shape = (50,50,3)) ,
    layers.MaxPooling2D(2,2),
    layers.Conv2D(64 , (3,3) , activation='relu'),
    layers.MaxPooling2D(2,2),  # 添加池化层压缩特征图尺寸
    layers.Flatten(),
    layers.Dense(128, activation='relu'),  # 新增全连接层提升模型能力
    layers.Dense(2 , activation='softmax')
])

4. 修正数据划分后的形状打印 #

X_train和X_test需转为numpy数组才能正确打印形状：

X_train = np.array(X_train)
X_test = np.array(X_test)

print('X_train: ' + str(X_train.shape))  # 输出应为 (样本数,50,50,3)
print('Y_train: ' + str(y_train.shape))
print('X_test:  '  + str(X_test.shape))
print('Y_test:  '  + str(y_test.shape))

修改后的完整代码 #

import numpy as np
import cv2 as cv
import tensorflow.keras as keras
import os
from sklearn.preprocessing import LabelEncoder
from tensorflow.keras.utils import to_categorical
from tensorflow.keras import layers , models
from sklearn.model_selection import train_test_split

images_vector =[]
images_label =[]

fileNames = os.listdir('train')

for i , f_name in enumerate(fileNames) :
    image = cv.imread('train/' + f_name)
    if image is None:
        print(f"跳过损坏图片: {f_name}")
        continue
    
    image = cv.resize(image , (50,50))
    image = image/255.0
    images_vector.append(image)
    images_label.append(f_name.split('.')[0]) 

    if i%10000 == 0 : 
        print(f" [INFO ] : {i} images are processed...")

images_vector = np.array(images_vector)

labelEncoder  = LabelEncoder()
images_label = labelEncoder.fit_transform(images_label)
images_label = to_categorical(images_label)

X_train , X_test , y_train , y_test = train_test_split(images_vector ,images_label  , random_state=40 , train_size=0.8)

print('X_train: ' + str(X_train.shape))
print('Y_train: ' + str(y_train.shape))
print('X_test:  '  + str(X_test.shape))
print('Y_test:  '  + str(y_test.shape))

net = models.Sequential([
    layers.Conv2D(32 , (3,3) , activation='relu' , input_shape = (50,50,3)) ,
    layers.MaxPooling2D(2,2),
    layers.Conv2D(64 , (3,3) , activation='relu'),
    layers.MaxPooling2D(2,2),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dense(2 , activation='softmax')
])

net.summary()

修改后输入数据形状为(样本数,50,50,3)，符合Conv2D层要求的4D张量，模型可正常构建。

内容的提问来源于stack exchange，提问作者Ahmad Badpey