ValueError：使用张量流在神经网络中从 1 中减去 2 导致的负维度大小

Question

我正在尝试编写一个模型来对 mnist 数据集中的手写数字进行分类。我目前使用的模型如下所示：

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"),
    tf.keras.layers.Dense(4000,activation = 'sigmoid'),
    tf.keras.layers.MaxPooling2D(),
    tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"),
    tf.keras.layers.Dense(4000,activation = 'sigmoid'),
    tf.keras.layers.MaxPooling2D(),
    tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"),
    tf.keras.layers.Dense(4000,activation = 'sigmoid'),
    tf.keras.layers.MaxPooling2D(),
    tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"),
    tf.keras.layers.Dense(4000,activation = 'sigmoid'),
    tf.keras.layers.MaxPooling2D(),
    tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"),
    tf.keras.layers.Dense(4000,activation = 'sigmoid'),
    tf.keras.layers.MaxPooling2D(),
    tf.keras.layers.Flatten(input_shape = (28,28)),
    tf.keras.layers.Dense(10,'softmax')
])

它应该是一个具有 5 层的卷积深度神经网络，我在模型的最后一行不断收到此错误消息。

ValueError: Negative dimension size caused by subtracting 2 from 1 for '{{node max_pooling2d_9/MaxPool}} = MaxPool[T=DT_FLOAT, data_format="NHWC", ksize=[1, 2, 2, 1], padding="VALID", strides=[1, 2, 2, 1]](dense_10/Sigmoid)' with input shapes: [?,1,1,4000].`

我不明白错误或如何修复它。有人可以帮帮我吗？我已经将填充设置为“相同”，但它没有修复它。如果有帮助，我也在使用 google colab。

Answer 1

这里可能需要你。要显示您的代码，您需要将其括在四个连续的特殊字符中。字符to use is the one to the left of 1 on your keyboard. Put 4 of these characters in a row. Then skip a line and enter code. When done with code create a new line. On it put four of those 连续字符。这封装了您的代码，显示它会在您发布问题时正确显示。您发布的格式化代码如下所示。

model = tf.keras.models.Sequential([ 
tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"), 
tf.keras.layers.Dense(4000,activation = 'sigmoid'), tf.keras.layers.MaxPooling2D(), tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"), tf.keras.layers.Dense(4000,activation = 'sigmoid'), 
tf.keras.layers.MaxPooling2D(), 
tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"), tf.keras.layers.Dense(4000,activation = 'sigmoid'),
tf.keras.layers.MaxPooling2D(), 
tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"), tf.keras.layers.Dense(4000,activation = 'sigmoid'), 
tf.keras.layers.MaxPooling2D(), 
tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"), tf.keras.layers.Dense(4000,activation = 'sigmoid'), 
tf.keras.layers.MaxPooling2D(), tf.keras.layers.Flatten(input_shape = (28,28)), tf.keras.layers.Dense(10,'softmax') ] )  

它应该是一个具有 5 层的卷积深度神经网络，我在模型的最后一行不断收到此错误消息。 ValueError: '{{node max_pooling2d_9/MaxPool}} = MaxPoolT=DT_FLOAT, data_format="NHWC", ksize=[1, 2, 2, 1], padding="VALID", strides=[1, 2, 2, 1]'

Answer 2

在您的 keras 模型中，input_shape 参数仅用于您的第一层并初始化您的网络。

因为您使用MaxPooling2D，所以您的图像缩小了2。如果您想了解您的网络，请尝试print(model.summary())。因此，在某一时刻，您将 1x1x4000 图像缩小了 2 倍。例如，尝试在网络中深入一些：

model = tf.keras.models.Sequential([
        tf.keras.layers.Conv2D(8,(3,3),input_shape = (28,28,1),padding = "same"),
        tf.keras.layers.Dense(4000,activation = 'sigmoid'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(8,(3,3),padding = "same"),
        tf.keras.layers.Dense(4000,activation = 'sigmoid'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(8,(3,3),padding = "same"),
        tf.keras.layers.Dense(4000,activation = 'sigmoid'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(8,(3,3),padding = "same"),
        tf.keras.layers.Dense(4000,activation = 'sigmoid'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(10,'softmax')])

结构如下：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_35 (Conv2D)           (None, 28, 28, 8)         80        
_________________________________________________________________
dense_39 (Dense)             (None, 28, 28, 4000)      36000     
_________________________________________________________________
max_pooling2d_34 (MaxPooling (None, 14, 14, 4000)      0         
_________________________________________________________________
conv2d_36 (Conv2D)           (None, 14, 14, 8)         288008    
_________________________________________________________________
dense_40 (Dense)             (None, 14, 14, 4000)      36000     
_________________________________________________________________
max_pooling2d_35 (MaxPooling (None, 7, 7, 4000)        0         
_________________________________________________________________
conv2d_37 (Conv2D)           (None, 7, 7, 8)           288008    
_________________________________________________________________
dense_41 (Dense)             (None, 7, 7, 4000)        36000     
_________________________________________________________________
max_pooling2d_36 (MaxPooling (None, 3, 3, 4000)        0         
_________________________________________________________________
conv2d_38 (Conv2D)           (None, 3, 3, 8)           288008    
_________________________________________________________________
dense_42 (Dense)             (None, 3, 3, 4000)        36000     
_________________________________________________________________
max_pooling2d_37 (MaxPooling (None, 1, 1, 4000)        0         
_________________________________________________________________
flatten_5 (Flatten)          (None, 4000)              0         
_________________________________________________________________
dense_43 (Dense)             (None, 10)                40010     
=================================================================
Total params: 1,048,114
Trainable params: 1,048,114
Non-trainable params: 0
_________________________________________________________________

最后，你的网络没有优化好，留给你去寻找和尝试更好的架构。