训练深度学习模型时出错

Question

所以我设计了一个 CNN 并使用以下参数进行编译，

training_file_loc = "8-SignLanguageMNIST/sign_mnist_train.csv"
testing_file_loc = "8-SignLanguageMNIST/sign_mnist_test.csv"

def getData(filename):
    images = []
    labels = []
    with open(filename) as csv_file:
        file = csv.reader(csv_file, delimiter = ",")
        next(file, None)
        
        for row in file:
            label = row[0]
            data = row[1:]
            img = np.array(data).reshape(28,28)
            
            images.append(img)
            labels.append(label)
        
        images = np.array(images).astype("float64")
        labels = np.array(labels).astype("float64")
        
    return images, labels

training_images, training_labels = getData(training_file_loc)
testing_images, testing_labels = getData(testing_file_loc)

print(training_images.shape, training_labels.shape)
print(testing_images.shape, testing_labels.shape)

training_images = np.expand_dims(training_images, axis = 3)
testing_images = np.expand_dims(testing_images, axis = 3)

training_datagen = ImageDataGenerator(
    rescale = 1/255,
    rotation_range = 45,
    width_shift_range = 0.2,
    height_shift_range = 0.2,
    shear_range = 0.2,
    zoom_range = 0.2,
    horizontal_flip = True,
    fill_mode = "nearest"
)

training_generator = training_datagen.flow(
    training_images,
    training_labels,
    batch_size = 64,
)


validation_datagen = ImageDataGenerator(
    rescale = 1/255,
    rotation_range = 45,
    width_shift_range = 0.2,
    height_shift_range = 0.2,
    shear_range = 0.2,
    zoom_range = 0.2,
    horizontal_flip = True,
    fill_mode = "nearest"
)

validation_generator = training_datagen.flow(
    testing_images,
    testing_labels,
    batch_size = 64,
)

model = tf.keras.Sequential([
    keras.layers.Conv2D(16, (3, 3), input_shape = (28, 28, 1), activation = "relu"),
    keras.layers.MaxPooling2D(2, 2),
    keras.layers.Conv2D(32, (3, 3), activation = "relu"),
    keras.layers.MaxPooling2D(2, 2),
    keras.layers.Flatten(),
    keras.layers.Dense(256, activation = "relu"),
    keras.layers.Dropout(0.25),
    keras.layers.Dense(512, activation = "relu"),
    keras.layers.Dropout(0.25),
    keras.layers.Dense(26, activation = "softmax")
])

model.compile(
    loss = "categorical_crossentropy",
    optimizer = RMSprop(lr = 0.001),
    metrics = ["accuracy"]
)

但是，当我运行 model.fit() 时，出现以下错误，

ValueError: Shapes (None, 1) and (None, 24) are incompatible

将损失函数更改为sparse_categorical_crossentropy后，程序运行正常。

我不明白为什么会这样。

谁能解释一下这个以及这些损失函数之间的区别？

Answer 1

问题是，categorical_crossentropy 需要单热编码标签，这意味着，对于每个样本，它需要一个长度为 num_classes 的张量，其中设置了第 label 个元素为 1，其他一切为 0。

另一方面，sparse_categorical_crossentropy直接使用整数标签（因为这里的用例是大量的类，所以one-hot-encoded标签会浪费内存很多零）。我相信，但我无法证实这一点，categorical_crossentropy 比它的稀疏对应物更快运行。

对于你的情况，使用 26 类我建议使用非稀疏版本并将你的标签转换为单热编码，如下所示：

def getData(filename):
    images = []
    labels = []
    with open(filename) as csv_file:
        file = csv.reader(csv_file, delimiter = ",")
        next(file, None)
        
        for row in file:
            label = row[0]
            data = row[1:]
            img = np.array(data).reshape(28,28)
            
            images.append(img)
            labels.append(label)
        
        images = np.array(images).astype("float64")
        labels = np.array(labels).astype("float64")
        
    return images, tf.keras.utils.to_categorical(labels, num_classes=26) # you can omit num_classes to have it computed from the data

旁注：除非你有理由对图像使用 float64，否则我会切换到 float32（它将数据集所需的内存减半，并且模型可能会将它们转换为 float32 作为第一个操作）

Answer 2

简单，对于输出类为整数 sparse_categorical_crosentropy 的分类问题，对于标签在一个热编码标签中转换的分类问题，我们使用 categorical_crosentropy.

训练深度学习模型时出错

Error while training a deep learning model

python

machine-learning

deep-learning

keras

tensorflow2.0