训练图像分类CNN时遭遇数据基数模糊问题求助

问题背景 #

训练性别分类CNN模型时触发「Ambiguous data cardinality」（数据基数模糊）报错，已确认图像数据集与标签集的样本数量匹配，但无法定位问题根源。相关代码如下：

import numpy as np
import os, shutil
import tensorflow as tf
from PIL import Image
import matplotlib.pyplot as plt
import glob

dataset_path = '/Users/myusername/Downloads/Male_and_Female_face_dataset'

male_faces = os.path.join(dataset_path, 'Male_Faces')
female_faces = os.path.join(dataset_path, 'Female_Faces')

batch_size = 32
img_height = 180
img_width = 180

plt.figure(figsize=(10, 10))


image_files = os.listdir(male_faces)

# Loop for displaying only the first 9 images
for i in range(9):
    img_path = os.path.join(male_faces, image_files[i])
    
    # Read the image using PIL and convert it to a NumPy array
    img_array = np.asarray(Image.open(img_path))
    
    ax = plt.subplot(3, 3, i+1)
    plt.imshow(img_array.astype("uint8"))
    plt.title(f"Male {i+1}")  # You can modify this title as needed
    plt.axis('off')

plt.show()

plt.figure(figsize=(10, 10))


image_files = os.listdir(female_faces)

# Loop for displaying only the first 9 images
for i in range(9):
    img_path = os.path.join(female_faces, image_files[i])
    
    # Read the image using PIL and convert it to a NumPy array
    img_array = np.asarray(Image.open(img_path))
    
    ax = plt.subplot(3, 3, i+1)
    plt.imshow(img_array.astype("uint8"))
    plt.title(f"female {i+1}")  # You can modify this title as needed
    plt.axis('off')

plt.show()

male_file_types = set()
female_file_types = set()

for img in os.listdir(male_faces):
    img_path = os.path.join(male_faces, img)
    file_extension = os.path.splitext(img_path)[1]
    
    if file_extension not in male_file_types:
        male_file_types.add(file_extension)

male_file_types

for img in os.listdir(female_faces):
    img_path = os.path.join(female_faces, img)
    file_extension = os.path.splitext(img_path)[1]
    
    if file_extension not in female_file_types:
        female_file_types.add(file_extension)
    
female_file_types

male_file_types

female_file_types

male_images = []
female_images = []
labels = []

for file in os.listdir(male_faces):
    file_path = os.path.join(male_faces, file)
    file = np.asarray(Image.open(file_path))
    male_images.append(file)
    labels.append('male')   
        
for file in os.listdir(female_faces):
    file_path = os.path.join(female_faces, file)
    file = np.asarray(Image.open(file_path))
    female_images.append(file)
    labels.append('female')

len(male_images)

len(female_images)

X = male_images + female_images
y = labels

len(X)

len(y)

from sklearn.model_selection import train_test_split

print(y[0])

one_hot_labels = np.array([[1, 0] if label == 'female' else [0, 1] for label in y])

print(one_hot_labels)

y = one_hot_labels

for i in range(5):
    print(X[i].shape)

import cv2

X_resized = []

for array in X:
    X_resize = cv2.resize(array, (200, 200), interpolation=cv2.INTER_AREA)
    X_resized.append(X_resize)
        

X = X_resized

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=0.25, random_state=1)

model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_shape=(200, 200, 3)),
    tf.keras.layers.Conv2D(32, kernel_size=(3, 3), activation='relu'),
    tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(64, activation='relu'),
    tf.keras.layers.Dense(2, activation='softmax')
])

# Compile the model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

# Train the model
history = model.fit(X_train, y_train, epochs=10, batch_size=32, validation_data=(X_val, y_val))

# Evaluate the model on the validation set
loss, accuracy = model.evaluate(X_val, y_val)
print(f'Validation accuracy: {accuracy}')

问题根源 #

你虽然调整了所有图像的尺寸为(200,200,3)，但X_train、X_val始终是Python列表，而非NumPy数组。TensorFlow的model.fit需要输入具有明确形状的张量（NumPy数组或TensorFlow张量），列表无法让框架正确推断数据的整体维度和样本基数，即使每个列表元素的形状一致，列表本身没有统一的形状属性，最终触发基数模糊报错。

解决方案 #

在划分数据集后，将图像列表转换为NumPy数组，添加以下代码：

# 将训练、验证、测试集转换为NumPy数组
X_train = np.array(X_train)
X_val = np.array(X_val)
X_test = np.array(X_test)

转换后可以检查数组形状，确保符合模型输入要求：

print(X_train.shape)  # 预期输出：(样本数, 200, 200, 3)
print(y_train.shape)  # 预期输出：(样本数, 2)

完成上述修改后，TensorFlow就能正确识别数据的基数和维度，正常启动训练。

内容的提问来源于stack exchange，提问作者gusifer98