python - 使用 fit_generator 不匹配的形状时出错(Keras)

Question

我正在尝试构建一个简单的分类 CNN，它将使用以下代码将一组 1233 张图像分为 4 个类别:

unclassified_datagen = keras.preprocessing.image.ImageDataGenerator(
    rescale=1. / 255,
    horizontal_flip=True
)
unclassified_generator = train_datagen.flow_from_directory(
    'data/unclassified',
    target_size=(120, 120),
    batch_size=1233,
    class_mode='input',
    shuffle=False,
)

model_unclassified = keras.Sequential()
model_unclassified.add(layers.Conv2D(1233, (3, 3), input_shape=(120, 120, 3), padding="SAME"))
model_unclassified.add(layers.Dense(64, activation='relu'))
model_unclassified.add(layers.Dense(4, activation='sigmoid'))

model_unclassified.compile(loss='sparse_categorical_crossentropy',
                           optimizer='rmsprop',
                           metrics=['accuracy'])
model_unclassified.fit_generator(unclassified_generator, epochs=1)

但我收到以下错误: ValueError: Error when checking target: expected dense_2 to have shape (120, 120, 1) but got array with shape (120, 120, 3)
我究竟做错了什么？

Answer 1

您应该添加 Flatten层，因为 Conv2D为每个样本返回 3D 数组:

model_unclassified = keras.Sequential()
model_unclassified.add(layers.Conv2D(1233, (3, 3), input_shape=(120, 120, 3), padding="SAME"))
model_unclassified.add(layers.Flatten())
model_unclassified.add(layers.Dense(64, activation='relu'))
model_unclassified.add(layers.Dense(4, activation='sigmoid'))