keras更新后出现 Exception: Regularizers cannot be reused

[cescigl]

具体的代码是这个 model2.add(TimeDistributed(Convolution2D(64, 3, 3, border_mode='valid', W_regularizer=l2(0.01), activity_regularizer=activity_l2(0.01))))

如果去掉activity_regularizer就正常了，tf和th都试了，都会报这个异常

[cescigl]

Using TensorFlow backend. Traceback (most recent call last): File "test.py", line 26, in

File "/Users/cesc/anaconda/lib/python2.7/site-packages/keras/engine/topology.py", line 487, in call self.build(input_shapes[0]) File "/Users/cesc/anaconda/lib/python2.7/site-packages/keras/layers/wrappers.py", line 100, in build super(TimeDistributed, self).build() File "/Users/cesc/anaconda/lib/python2.7/site-packages/keras/layers/wrappers.py", line 28, in build regularizer.set_layer(self) File "/Users/cesc/anaconda/lib/python2.7/site-packages/keras/regularizers.py", line 113, in set_layer raise Exception('Regularizers cannot be reused') Exception: Regularizers cannot be reused

[MoyanZitto]

@cescigl 对这个问题不太了解，看了看源代码分析了一下，也许你应该把正则项放到Convolution2D之外，也就是对TimeDistributed Covolution2D正则而不是对其中的一个Convolution2D正则，TimeDistributed的父类是Wrapper，其父类是Layer所以也可以加正则。

猜测问题的原因可能是为一个层施加了多次正则

[cescigl]

@MoyanZitto fchollet在27天前更新了wrapper层。 Fix activity regularization for wrapper layers fchollet committed 27 days ago 添加了相关代码 # properly attribute the current layer to # regularizers that need access to it # (e.g. ActivityRegularizer). for regularizer in self.regularizers: if hasattr(regularizer, 'set_layer'): regularizer.set_layer(self)

不知道会不会有影响

[summerHaha]

@cescigl 您好，我现在遇到的问题跟您一样，TimeDistributed(Dense(5, activation='softmax', activity_regularizer=activity_l1(0.01)))(lstm) ，提示 “Regularizers cannot be reused”，看了下文档，感觉这么用的用法是正确的啊，不知道为什么报错。 所以想请问一下，您这个问题是怎么解决的？谢谢～～

[cescigl]

@summerHaha 我并没有测试新版本的问题，这个问题只在某个版本出现，之前的版本是OK的，我只是没有用regular