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

Fix bug where TensorFlowOpLayer would autocast inputs.

Browse Source 
PiperOrigin-RevId: 261395761
This commit is contained in:
Reed Wanderman-Milne 2019-08-02 14:47:35 -07:00 committed by TensorFlower Gardener
parent 6fba1efce3
commit 686c123392
2 changed files with 14 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/python/keras/engine/base_layer.py
View File

@ -2488,8 +2488,11 @@ class TensorFlowOpLayer(Layer):
constants=None,
trainable=True,
dtype=None):
# Pass autocast=False, as if inputs are cast, input types might not match
# Operation type.
super(TensorFlowOpLayer, self).__init__(
name=_TF_OP_LAYER_NAME_PREFIX + name, trainable=trainable, dtype=dtype)
name=_TF_OP_LAYER_NAME_PREFIX + name, trainable=trainable, dtype=dtype,
autocast=False)
if not isinstance(node_def, bytes):
node_def = node_def.encode('utf-8')
self.node_def = node_def_pb2.NodeDef.FromString(node_def)

10
tensorflow/python/keras/layers/tensorflow_op_layer_test.py
View File

@ -126,6 +126,15 @@ def _reuse_op():
return keras.Model(inputs, outputs)
def _float64_op():
inputs = keras.Input(shape=(10,))
x = keras.layers.Dense(10, dtype='float64')(inputs)
x = gen_nn_ops.relu(x)
assert x.dtype == 'float64', 'x has dtype: %s' % x.dtype
outputs = keras.layers.Dense(10)(x)
return keras.Model(inputs, outputs)
class LayerWithLayer(keras.layers.Layer):
def build(self, input_shape):
@ -179,6 +188,7 @@ class AutoLambdaTest(keras_parameterized.TestCase):
('op_with_tensor_list', _op_with_tensor_list),
('add_n', _add_n),
('_reuse_op', _reuse_op),
('_float64_op', _float64_op),
('_inner_layer', _inner_layer),
('_reuse_ancillary_layer', _reuse_ancillary_layer),
)