Getting learning_phase at the activation functions level

I modified the ReLU activation function to print the value of the learning phase, expecting that during the training (fit) and testing (evaluate, predict) it will be different. In fact, according to the documentation, the value during the training should be 1 whereas during the testing 0 is expected.

def relu(x, alpha=0., max_value=None): print(K.learning_phase()) return K.relu(x, alpha=alpha, max_value=max_value) During the training I get: Tensor(“activation_1/keras_learning_phase:0”, dtype=bool) and during testing I don’t get anything at all.

How does it make sense? JM.

[jmlipman]

Thanks!

activation._uses_learning_phase was the key 😉

[zhhongzhi]

@mahnerak Thanks! I noticed that I have wrongly edited it. And the original error reported is:

`Using TensorFlow backend.
Traceback (most recent call last):
  File "/mydata/wp/SimpleQAMine/fullSimpleQA/relation_reranking/learn_get_intern_valuse.py", line 60, in <module>
    y = Dense(units=10, activation=custom_activation) (h)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/Keras-2.0.2-py2.7.egg/keras/engine/topology.py", line 578, in __call__
    output = self.call(inputs, **kwargs)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/Keras-2.0.2-py2.7.egg/keras/layers/core.py", line 845, in call
    output = self.activation(output)
  File "/mydata/wp/SimpleQAMine/fullSimpleQA/relation_reranking/learn_get_intern_valuse.py", line 54, in custom_activation
    activation = x + K.in_train_phase(0.0, 1000.0)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/Keras-2.0.2-py2.7.egg/keras/backend/tensorflow_backend.py", line 2451, in in_train_phase
    x = switch(training, x, alt)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/Keras-2.0.2-py2.7.egg/keras/backend/tensorflow_backend.py", line 2410, in switch
    else_expression_fn)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 1741, in cond
    orig_res, res_t = context_t.BuildCondBranch(fn1)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 1668, in BuildCondBranch
    real_v = self._ProcessOutputTensor(v)
  File "/home/hongzhi/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/control_flow_ops.py", line 1626, in _ProcessOutputTensor
    if val.name not in self._values:
AttributeError: 'float' object has no attribute 'name'`

Thank you for your attention.