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，我得到了以下错误，
ValueError: Shapes (None, 1) and (None, 24) are incompatible
将损失函数改为sparse_categorical_crossentropy后，程序运行良好。
我不明白为什么会发生这种事。
有人能解释一下这一点，以及这些损失函数之间的区别吗？
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，其中第label个元素设置为1，其他元素都为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，所以一个热编码标签会浪费大量零的内存(。我相信，但我不能证实这一点，categorical_crossentropy比稀疏的对应程序运行得更快。
对于您的情况，对于26个类，我建议使用非稀疏版本，并将您的标签转换为一个热编码的标签，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，模型可能会将它们转换为float32，作为第一个操作(
Simple，对于输出类为整数的分类问题，使用sparse_categorical_crosentropy，对于那些将标签转换为一个热编码标签的问题，我们使用categorical_crosentropy。