experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Relax the check for state_size

Browse Source 
The behaviour of `hasattr` is to evaluate the state_size member. In the case of `tfa.seq2seq.AttentionWrapper`, that is a @property member that is built at graph runtime after calling `setup_memory`, thus `hasattr` returns an error when using AttentionWrapper with dynamic memories.

More details: https://github.com/tensorflow/addons/issues/680
This commit is contained in:
George Sterpu 2019-11-21 13:59:30 +00:00 committed by GitHub
parent c5eafbfebd
commit 4ab6a520c9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/python/keras/layers/recurrent.py
View File

@ -82,7 +82,7 @@ class StackedRNNCells(Layer):
if not hasattr(cell, 'call'):
raise ValueError('All cells must have a `call` method. '
'received cells:', cells)
if not hasattr(cell, 'state_size'):
if not ('state_size' in dir(cell) or hasattr(cell, 'state_size')):
raise ValueError('All cells must have a '
'`state_size` attribute. '
'received cells:', cells)
@ -391,7 +391,7 @@ class RNN(Layer):
if not hasattr(cell, 'call'):
raise ValueError('`cell` should have a `call` method. '
'The RNN was passed:', cell)
if not hasattr(cell, 'state_size'):
if not ('state_size' in dir(cell) or hasattr(cell, 'state_size')):
raise ValueError('The RNN cell should have '
'an attribute `state_size` '
'(tuple of integers, '