重塑Tensorflow ImageDataGenerator

Question

我正在尝试使用ImageDataGenerator创建我的训练数据集。有两个不同的文件夹，其中包含不同标签的图像。当我开始训练的时候，我得到了一个关于形状的错误：

ValueError: Shapes (None, 1) and (None, 2) are incompatible

我该如何解决这个问题？每个图像的原始大小是100x100，它们是RGB形式的。

 datagen = ImageDataGenerator(
    rescale = 1./255,
    rotation_range = 5,
    zoom_range = (0.95, 0.95),
    horizontal_flip = False,
    vertical_flip = False,
    data_format = 'channels_last',
    validation_split = 0.0,
    dtype = tf.float32
    )

train_generator = datagen.flow_from_directory('sign_data/', 
                                              target_size=(48,48),
                                              batch_size=4, color_mode='grayscale', 
                                              class_mode='sparse', shuffle=True, 
                                              subset='training', seed=123) 

model = Sequential()
model.add(Conv2D(16, 3, data_format="channels_last", padding='same', input_shape=(48,48,1)))
model.add(Conv2D(32, 3, padding='same'))
model.add(MaxPooling2D())
model.add(Flatten())
model.add(Dense(2))

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

hist = model.fit(train_generator, epochs=10)

我也尝试使用target_size=(48,48,1)

train_generator = datagen.flow_from_directory('sign_data/', 
                                          target_size=(48,48,1),
                                          batch_size=4, color_mode='grayscale', 
                                          class_mode='sparse', shuffle=True, 
                                          subset='training', seed=123)

错误是

ValueError: could not broadcast input array from shape (48,48,1) into shape (48,48,1,1)

Answer 1

来自评论

如果使用sparse类模式，则需要使用sparse categorical crossentropy loss (转译自Dr.Snoopy)

工作代码如下所示

datagen = ImageDataGenerator(
    rescale = 1./255,
    rotation_range = 5,
    zoom_range = (0.95, 0.95),
    horizontal_flip = False,
    vertical_flip = False,
    data_format = 'channels_last',
    validation_split = 0.0,
    dtype = tf.float32
    )

train_generator = datagen.flow_from_directory('sign_data/', 
                                              target_size=(48,48),
                                              batch_size=4, color_mode='grayscale', 
                                              class_mode='sparse', shuffle=True, 
                                              subset='training', seed=123) 

model = Sequential()
model.add(Conv2D(16, 3, data_format="channels_last", padding='same', input_shape=(48,48,1)))
model.add(Conv2D(32, 3, padding='same'))
model.add(MaxPooling2D())
model.add(Flatten())
model.add(Dense(2))

model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

hist = model.fit(train_generator, epochs=10)