当适合我的模型时，我得到 ValueError: Input 0 of layer sequence is incompatible with the layer

Question

我正在尝试使用 mnist 数据集来模拟 LeNet，

我正在做以下事情，

import tensorflow as tf
import tensorflow_datasets as tfds
from tensorflow.keras import layers, models
from tensorflow.keras.utils import plot_model
from tensorflow.keras.optimizers import SGD

import matplotlib.pyplot as plt
import numpy as np

# Import dataset
(train_ds, test_ds), info_ds = tfds.load('mnist', split=['train','test'], 
                               as_supervised = True,
                               with_info=True,
                               batch_size = -1)

train_images, train_labels = tfds.as_numpy(train_ds)
test_images, test_labels = tfds.as_numpy(test_ds)

# Split test to obtain validation dataset
val_size = int(len(test_images) * 0.8)
val_images, test_images = test_images[:val_size], test_images[val_size:]
val_labels, test_labels = test_labels[:val_size], test_labels[val_size:]

# Normalizing images between 0 to 1
train_images, test_images = train_images / 255.0, test_images / 255.0

# Create the model
model = models.Sequential()
model.add(layers.Conv2D(filters=6, kernel_size=(5,5), activation='relu', input_shape=(32,32,3)))
model.add(layers.MaxPooling2D(pool_size=(2,2)))
model.add(layers.Conv2D(filters=16, kernel_size=(5,5), activation='relu'))
model.add(layers.MaxPooling2D(pool_size=(2,2)))
model.add(layers.Flatten())
model.add(layers.Dense(120,activation='relu'))
model.add(layers.Dense(84,activation='relu'))
model.add(layers.Dense(10,activation='softmax'))

# Compile
opt = SGD(learning_rate=0.1)
model.compile(optimizer=opt,
               loss='categorical_crossentropy',
               metrics=['accuracy'])

# Fit
history = model.fit(train_images, train_labels,
                    epochs=10, batch_size=128,
                    validation_data=(val_images, val_labels),
                    verbose=2)

合适的时候，我得到这个错误：

ValueError：层顺序的输入 0 与层不兼容：输入形状的预期轴 -1 具有值 3，但接收到形状为 (None, 28, 28, 1) 的输入

这意味着我必须重塑 mi 图像？

我想也许我必须像这样将标签转换为分类标签

from tensorflow.keras.utils import to_categorical

train_labels = to_categorical(train_labels)
test_labels = to_categorical(test_labels)

但随后又出现同样的错误，

ValueError: 层序贯_1 的输入 0 与层不兼容：输入形状的预期轴 -1 具有值 3，但接收到形状为 (None, 28, 28, 1) 的输入

有人可以帮我理解吗？

非常感谢！

Answer 1

您的代码中有两个问题。比如

• （问题 1）您在模型中设置了 input_shpae = (32,32,3)，而 mnist 样本为 (28, 28, 1)。如果您检查样品形状，您会看到：

train_images.shape, train_labels.shape
((60000, 28, 28, 1), (60000,))

但是您将输入形状定义为与模型定义中的不同。

# current: and not OK, according to the sample shapes 
...kernel_size=(5,5), activation='relu', input_shape=(32,32,3))) 

# should be, otherwise resize your input 
...kernel_size=(5,5), activation='relu', input_shape=(28,28,3))) 

• （问题 2）输入标签是整数（而不是 one-hot 编码），检查 train_labels[:5] 但您设置了 categorical_crossentropy，而它应该是sparse_categorical_crossentropy 用于整数目标。

current
model.compile(optimizer=opt,
               loss='categorical_crossentropy', # when labels are one-hot encoded 
               metrics=['accuracy'])

should be 
model.compile(optimizer=opt,
               loss='sparse_categorical_crossentropy',  # when label are integers 
               metrics=['accuracy'])

现在，正如您稍后提到的，您已尝试使用to_categorical 对目标标签进行 one-hot 编码，在这种情况下，您可以使用 categorical_crossentropy 作为损失函数。