使用"Flatten"或"Reshape"在keras中获得未知输入形状的1D输出

Question

我想使用keras层Flatten()或Reshape((-1,))在模型的末尾输出一维矢量[0,0,1,0,0, ... ,0,0,1,0].

可悲的是,由于我未知的输入形状存在问题:
input_shape=(4, None, 1))).

所以通常输入形状介于两者之间[batch_size, 4, 64, 1],[batch_size, 4, 256, 1]输出应该是batch_size x未知维度(对于上面的第一个例子:[batch_size, 64]和对于secound [batch_size, 256]).

我的模型看起来像:

model = Sequential()
model.add(Convolution2D(32, (4, 32), padding='same', input_shape=(4, None, 1)))
model.add(BatchNormalization())
model.add(LeakyReLU())
model.add(Convolution2D(1, (1, 2), strides=(4, 1), padding='same'))
model.add(Activation('sigmoid'))
# model.add(Reshape((-1,))) produces the error
# int() argument must be a string, a bytes-like object or a number, not 'NoneType' 
model.compile(loss='binary_crossentropy', optimizer='adadelta')

所以我当前的输出形状是[batchsize,1,unknown dimension,1].例如,这不允许我使用class_weights "ValueError: class_weight not supported for 3+ dimensional targets.".

当我使用灵活的输入形状时,是否可以使用类似的东西Flatten()或者Reshape((1,))在keras(带有张量流后端的2.0.4)中展平我的3维输出？

非常感谢!

Answer 1

您可以尝试K.batch_flatten()包裹在Lambda图层中.输出形状K.batch_flatten()在运行时动态确定.

model.add(Lambda(lambda x: K.batch_flatten(x)))
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_5 (Conv2D)            (None, 4, None, 32)       4128      
_________________________________________________________________
batch_normalization_3 (Batch (None, 4, None, 32)       128       
_________________________________________________________________
leaky_re_lu_3 (LeakyReLU)    (None, 4, None, 32)       0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 1, None, 1)        65        
_________________________________________________________________
activation_3 (Activation)    (None, 1, None, 1)        0         
_________________________________________________________________
lambda_5 (Lambda)            (None, None)              0         
=================================================================
Total params: 4,321
Trainable params: 4,257
Non-trainable params: 64
_________________________________________________________________


X = np.random.rand(32, 4, 256, 1)
print(model.predict(X).shape)
(32, 256)

X = np.random.rand(32, 4, 64, 1)
print(model.predict(X).shape)
(32, 64)