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

Adds an unsafe internal fastpath for oplayer construction in TF1 graphs due to specific requests.

Browse Source 
PiperOrigin-RevId: 342905795
Change-Id: I8385e5730a060b460e0e156886a3bdd0cbc3070c
This commit is contained in:
Tomer Kaftan 2020-11-17 11:24:19 -08:00 committed by TensorFlower Gardener
parent b2d1af50b9
commit da495851f9
1 changed files with 14 additions and 1 deletions

15
tensorflow/python/keras/engine/base_layer_utils.py
View File

@ -189,6 +189,19 @@ def create_keras_history(tensors):
return created_layers
# Unsafe Internal attribute.
# If True, Keras will not evaluate the constant-foldable inputs to tf op
# layers in TF1 graphs. This *might* speed up model construction time in
# certain settings, but it means
# the models will not be serializable/deserializable via get_config
# (Only via Savedmodels). It may also change the semantics of whether
# generated random numbers are generated once and re-used, or recomputed
# each time.
# Note: This path triggers for TPUEstimators / xla compiled graphs regardless
# of this setting.
_UNSAFE_GRAPH_OP_LAYER_CREATION = False
def _create_keras_history_helper(tensors, processed_ops, created_layers):
"""Helper method for `create_keras_history`.
@ -238,7 +251,7 @@ def _create_keras_history_helper(tensors, processed_ops, created_layers):
not ops.executing_eagerly_outside_functions())
using_xla = control_flow_util.GraphOrParentsInXlaContext(
ops.get_default_graph())
if ds_with_session or using_xla:
if ds_with_session or using_xla or _UNSAFE_GRAPH_OP_LAYER_CREATION:
# In Legacy Graph mode, evaluating here makes Session be
# configured improperly. The downside of this is that saving
# via `get_config` breaks, but SavedModel still works.