仅由 conv2d 层组成的 keras 中的训练网络

Question

我在 mnist 上用 keras 训练了我自己的模型。我只有 conv2d 层，因为我想在小图像（mnist：28x28 px）上训练网络，然后对 1920x1080 的大图像进行推理。

我的形状（用于训练）：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv1 (Conv2D)               (None, 28, 28, 64)        640       
_________________________________________________________________
batch_normalization_117 (Bat (None, 28, 28, 64)        256       
_________________________________________________________________
leaky_re_lu_117 (LeakyReLU)  (None, 28, 28, 64)        0         
_________________________________________________________________
max_pooling2d_119 (MaxPoolin (None, 14, 14, 64)        0         
_________________________________________________________________
conv2 (Conv2D)               (None, 14, 14, 128)       73856     
_________________________________________________________________
batch_normalization_118 (Bat (None, 14, 14, 128)       512       
_________________________________________________________________
leaky_re_lu_118 (LeakyReLU)  (None, 14, 14, 128)       0         
_________________________________________________________________
max_pooling2d_120 (MaxPoolin (None, 7, 7, 128)         0         
_________________________________________________________________
conv3 (Conv2D)               (None, 7, 7, 256)         295168    
_________________________________________________________________
batch_normalization_119 (Bat (None, 7, 7, 256)         1024      
_________________________________________________________________
leaky_re_lu_119 (LeakyReLU)  (None, 7, 7, 256)         0         
_________________________________________________________________
max_pooling2d_121 (MaxPoolin (None, 4, 4, 256)         0         
_________________________________________________________________
conv4 (Conv2D)               (None, 4, 4, 128)         295040    
_________________________________________________________________
batch_normalization_120 (Bat (None, 4, 4, 128)         512       
_________________________________________________________________
leaky_re_lu_120 (LeakyReLU)  (None, 4, 4, 128)         0         
_________________________________________________________________
max_pooling2d_122 (MaxPoolin (None, 2, 2, 128)         0         
_________________________________________________________________
conv5 (Conv2D)               (None, 1, 1, 10)          5130      
=================================================================
Total params: 672,138
Trainable params: 670,986
Non-trainable params: 1,152

推理形状：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv1 (Conv2D)               (None, 1920, 1080, 64)    640       
_________________________________________________________________
batch_normalization_113 (Bat (None, 1920, 1080, 64)    256       
_________________________________________________________________
leaky_re_lu_113 (LeakyReLU)  (None, 1920, 1080, 64)    0         
_________________________________________________________________
max_pooling2d_115 (MaxPoolin (None, 960, 540, 64)      0         
_________________________________________________________________
conv2 (Conv2D)               (None, 960, 540, 128)     73856     
_________________________________________________________________
batch_normalization_114 (Bat (None, 960, 540, 128)     512       
_________________________________________________________________
leaky_re_lu_114 (LeakyReLU)  (None, 960, 540, 128)     0         
_________________________________________________________________
max_pooling2d_116 (MaxPoolin (None, 480, 270, 128)     0         
_________________________________________________________________
conv3 (Conv2D)               (None, 480, 270, 256)     295168    
_________________________________________________________________
batch_normalization_115 (Bat (None, 480, 270, 256)     1024      
_________________________________________________________________
leaky_re_lu_115 (LeakyReLU)  (None, 480, 270, 256)     0         
_________________________________________________________________
max_pooling2d_117 (MaxPoolin (None, 240, 135, 256)     0         
_________________________________________________________________
conv4 (Conv2D)               (None, 240, 135, 128)     295040    
_________________________________________________________________
batch_normalization_116 (Bat (None, 240, 135, 128)     512       
_________________________________________________________________
leaky_re_lu_116 (LeakyReLU)  (None, 240, 135, 128)     0         
_________________________________________________________________
max_pooling2d_118 (MaxPoolin (None, 120, 68, 128)      0         
_________________________________________________________________
conv5 (Conv2D)               (None, 119, 67, 10)       5130      
=================================================================
Total params: 672,138
Trainable params: 670,986
Non-trainable params: 1,152

这里的目标是用我的输出类的尺寸创建一个卷积图像，它代表我的大图像中的滑动窗口以进行推理。

但是 keras 不会让我训练，因为在最后一层它会减少我的前一层输出的形状（从 (batch,x,y,channels) 到 (batch,channels)）：

ValueError: Error when checking target: expected conv5 to have 4 dimensions, but got array with shape (48000, 10)

形状需要是 (48000, 1, 1, 10) !!!我能做些什么来防止这种情况发生？当我引入flatten和dense时，我以后不能用它来推断大图像吗？

感谢您的时间和帮助。

Answer 1

为了能够在不同的输入大小上进行训练和测试，您应该做两件事：

1. 引入None作为输入维度。
2. 使用GlobalAveragePooling2D 和Conv2D 层，过滤器大小等于类别数。

以下示例代码可以创建一个模型，以对具有任何输入大小的图像进行训练和推理（假设最大池化和跨步不会导致负维度）。

from keras import layers, Model

my_input = layers.Input(shape=(None, None, 1))

x = layers.Conv2D(filters=32, kernel_size=3, strides=1)(my_input)
x = layers.BatchNormalization()(x)
x = layers.MaxPooling2D()(x)
x = layers.Conv2D(filters=64, kernel_size=3, strides=1)(x)
x = layers.BatchNormalization()(x)
x = layers.MaxPooling2D()(x)

out = layers.Conv2D(filters=10, kernel_size=1, strides=1)(x)
out = layers.GlobalAveragePooling2D()(out)
out = layers.Activation('softmax')(out)
model = Model(my_input, out)
model.summary()

模型摘要打印如下：

Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, None, None, 1)     0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, None, None, 32)    320       
_________________________________________________________________
batch_normalization_1 (Batch (None, None, None, 32)    128       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, None, None, 32)    0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, None, None, 64)    18496     
_________________________________________________________________
batch_normalization_2 (Batch (None, None, None, 64)    256       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, None, None, 64)    0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, None, None, 10)    650       
_________________________________________________________________
global_average_pooling2d_1 ( (None, 10)                0         
_________________________________________________________________
activation_1 (Activation)    (None, 10)                0         
=================================================================
Total params: 19,850
Trainable params: 19,658
Non-trainable params: 192