{NameError}name 'uses_learning_phase' is not defined

[valendin]

When using in an RNN, I get {NameError}name 'uses_learning_phase' is not defined when applying a Dense layer on the output of an LSTM.

Tensorflow 1.14

[digantamisra98]

Hi. Can you please provide the code so that I can replicate the error you're receiving along with the log. I haven't used Mish as of now on a LSTM so I'll have to check to validate the same. Meanwhile, I'm referring to these issue forums here:

1. NameError name 'uses_learning_phase' is not defined when using Keras TimeDistributed with batch_shape on Input tensorflow/tensorflow#23425
2. Problem with TimeDistributed() and Learning Phase keras-team/keras#4178

[digantamisra98]

@valendin can you try using Mish from the Tensorflow Addons repository. Implementation is given here - https://github.com/tensorflow/addons/tree/master/tensorflow_addons/activations

[valendin]

@digantamisra98 will do, have just moved to TF 2.0 so will have to retest

[qo4on]

using Mish from the Tensorflow Addons repository. Implementation is given here - https://github.com/tensorflow/addons/tree/master/tensorflow_addons/activations

There is no mish activation:

dir(tfa.activations)
['__builtins__',
 '__cached__',
 '__doc__',
 '__file__',
 '__loader__',
 '__name__',
 '__package__',
 '__path__',
 '__spec__',
 'absolute_import',
 'division',
 'gelu',
 'hardshrink',
 'lisht',
 'print_function',
 'sparsemax',
 'tanhshrink']

How can I import it to TF 2.0?

[digantamisra98]

@qo4on have you updated TFA? Because Mish is here - https://github.com/tensorflow/addons/blob/master/tensorflow_addons/activations/mish.py

[qo4on]

I installed latest tfa in Colab:

!pip install tensorflow-addons
import tensorflow_addons as tfa

Maybe I should have used nightly build !pip install tfa-nightly but I didn't find how to import it after the installation. I never used tfa before. Do you know how can I import tfa-nightly?

[digantamisra98]

@qo4on I just noticed that they're yet to put out their next release which is why Mish isn't a part of the package even though it's up on the source. You can directly install from the source (Instruction on their Readme) or else wait for the release for this month. I'll put forth an inquiry on when the next release is coming. Thanks

[digantamisra98]

@qo4on here's the update - tensorflow/addons#683 (comment)

[qo4on]

Eventually I found the way to use Mish from the Tensorflow Addons in Colab:

!pip install tf-nightly==2.1.0.dev20191029
!pip install tfa-nightly
import tensorflow_addons as tfa
dir(tfa.activations)
['__builtins__',
 '__cached__',
 '__doc__',
 '__file__',
 '__loader__',
 '__name__',
 '__package__',
 '__path__',
 '__spec__',
 'absolute_import',
 'division',
 'gelu',
 'hardshrink',
 'lisht',
 'mish',
 'print_function',
 'rrelu',
 'softshrink',
 'sparsemax',
 'tanhshrink']

[digantamisra98]

@qo4on Good to know. Let me know if there is any other issues regarding Mish.
@valendin closing this issue as of now, I believe using Mish from TFA will solve the error. Feel free to re-open it if the issue persists. Thanks!