我曾天真地尝试实现自己的池化和非池化功能，以便在池化时保留索引，以便在非池化时将值映射到索引。这是代码。
from keras.layers import Input, Layer
from keras.models import Model
from keras.layers.convolutional import Conv2D
import keras.backend as K
class MaxPool2DWithIndices(Layer):
def __init__(self, filter_size=(2,2), strides=(2,2), padding='valid', **kwargs):
super(MaxPool2DWithIndices, self).__init__(**kwargs)
self.filter_size = filter_size
self.strides = strides
self.padding = padding
def __call__(self, inputs):
'''
inputs: (n, H, W, d), ie batch of n inputs of size (H, W, d)
'''
print (inputs.shape)
inputs_shape = inputs.shape #K.array(inputs.shape)
pooled_H = (inputs_shape[1]-self.filter_size[0])//self.strides[0]+1
pooled_W = (inputs_shape[2]-self.filter_size[1])//self.strides[1]+1
#if type(inputs_shape[0]) is not int: inputs_shape[0] = 1
mpooled = np.zeros((inputs_shape[0], pooled_H, pooled_W, inputs_shape[3]))
indices = np.zeros((inputs_shape[0], pooled_H, pooled_W, inputs_shape[3]))
for n in range(0, inputs_shape[0], 2): # for each example
for i in range(0, inputs_shape[1], 2):
for j in range(0, inputs_shape[2], 2):
for k in range(inputs_shape[3]):
mpooled[n, i//2, j//2, k] = np.max(inputs[n, i:i+2, j:j+2, k])
indices[n, i//2, j//2, k] = np.argmax(inputs[n, i:i+2, j:j+2, k])
return [mpooled, indices]
class MaxUnPool2DWithIndices(Layer):
def __init__(self, filter_size=(2,2), strides=(2,2), padding='valid'):
super(MaxUnPool2DWithIndices, self).__init__(**kwargs)
self.indices = indices
self.filter_size = filter_size
self.strides = strides
self.padding = padding
def __call__(self, inputs, indices):
'''
inputs: (n, H, W, d), ie batch of n inputs of size (H, W, d)
'''
inputs_shape = inputs.shape #K.array(inputs.shape)
unpooled_H = (inputs_shape[1]-1)*self.strides[0]+self.filter_size[0]
unpooled_W = (inputs_shape[2]-1)*self.strides[1]+self.filter_size[1]
#if type(inputs_shape[0]) is not int: inputs_shape[0] = 1
max_unpooled = np.zeros((inputs_shape[0], unpooled_H, unpooled_W, inputs_shape[3]))
for n in range(inputs_shape[0]): # for each example
for i in range(0, unpooled_H, 2):
for j in range(0, unpooled_W, 2):
for k in range(inputs_shape[2]):
if self.indices[n, i//2, j//2, k] == 0:
max_unpooled[n, i+0, j+0, k] = inputs[n, i//2, j//2, k]
elif self.indices[n, i//2, j//2, k] == 1:
max_unpooled[n, i+0, j+1, k] = inputs[n, i//2, j//2, k]
elif self.indices[n, i//2, j//2, k] == 2:
max_unpooled[n, i+1, j+0, k] = inputs[n, i//2, j//2, k]
else: # it is 3
max_unpooled[n, i+1, j+1, k] = inputs[n, i//2, j//2, k]
return max_unpooled
以下面的模型为例来复制错误：
input_layer = Input(shape=(1024, 1024, 1))
encoded_out = Conv2D(4, (3, 3), activation='relu', padding="same")(input_layer)
encoded_out, indices1 = MaxPool2DWithIndices(filter_size=(2,2), strides=(2,2), padding='valid')(encoded_out)
decoded_out = MaxUnPool2DWithIndices(filter_size=(2,2), strides=(2,2), padding='valid')(encoded_out, indices1)
decoded_out = Conv2D(4, (3, 3), activation='relu', padding="same")(decoded_out)
model = Model(input_layer, decoded_out)
上述代码输出以下错误：
(None, 1024, 1024, 4)
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-77-5a2cf359f3ac> in <module>
2
3 encoded_out = Conv2D(4, (3, 3), activation='relu', padding="same")(input_layer)
----> 4 encoded_out, indices1 = MaxPool2DWithIndices(filter_size=(2,2), strides=(2,2), padding='valid')(encoded_out)
5
6 decoded_out = MaxUnPool2DWithIndices(filter_size=(2,2), strides=(2,2), padding='valid')(encoded_out, indices1)
<ipython-input-76-8a54043c249e> in __call__(self, inputs)
20 pooled_W = (inputs_shape[2]-self.filter_size[1])//self.strides[1]+1
21 #if type(inputs_shape[0]) is not int: inputs_shape[0] = 1
---> 22 mpooled = np.zeros((inputs_shape[0], pooled_H, pooled_W, inputs_shape[3]))
23 indices = np.zeros((inputs_shape[0], pooled_H, pooled_W, inputs_shape[3]))
24 for n in range(0, inputs_shape[0], 2): # for each example
TypeError: 'NoneType' object cannot be interpreted as an integer
这当然是由于事先不知道批量大小。如何解决此问题？
__call__函数需要一些if-else条件来处理input.shape=(None，H，W，d(或input.shape=(None，None，None(等情况。解决了！