训练深度学习 model 时出错

Question

所以我设计了一个CNN并用以下参数编译，

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 function 更改为sparse_categorical_crossentropy ，程序运行良好。

我不明白为什么会这样。

谁能解释这一点以及这些损失函数之间的区别？

Answer 1

问题是， categorical_crossentropy需要一个热编码标签，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label th 元素设置为 1，其他所有内容为 0。

另一方面， sparse_categorical_crossentropy直接使用 integer 标签（因为这里的用例是大量的类，所以单热编码的 label 会浪费 ZCD69B4957619BF060个零。） 我相信，但我无法证实这一点， categorical_crossentropy比其稀疏对应物运行得更快。

对于您的情况，我建议使用 26 个类，使用非稀疏版本并将您的标签转换为一次性编码，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：除非您有理由将float64用于图像，否则我会切换到float32 （它将数据集所需的 memory 减半，并且 model 可能会将它们转换为float32作为第一个操作）

Answer 2

很简单，对于您的 output 类是整数 sparse_categorical_crosentropy 的分类问题，用于标签转换为一个热编码标签的分类问题，我们使用 categorical_crosentropy。