Remove some of the usages of ops.EagerTensor from Keras. Specifically, the no-op usages (v1 keras code never runs when there are EagerTensors, so it never needs to be checked for v1 only code)
PiperOrigin-RevId: 334210496
Change-Id: I0b2e070dcf38e3ba349241a5fb8941c47e7ee6ff
diff --git a/tensorflow/python/keras/engine/base_layer_test.py b/tensorflow/python/keras/engine/base_layer_test.py
index 321eefc..98c852e 100644
--- a/tensorflow/python/keras/engine/base_layer_test.py
+++ b/tensorflow/python/keras/engine/base_layer_test.py
@@ -789,7 +789,6 @@
with ops.Graph().as_default():
x1 = array_ops.ones((3, 3))
x2 = array_ops.ones((3, 3))
- self.assertIsInstance(x2, ops.EagerTensor)
with self.assertRaisesRegex(TypeError, 'Graph tensors'):
math_ops.matmul(x1, x2)
diff --git a/tensorflow/python/keras/engine/base_preprocessing_layer.py b/tensorflow/python/keras/engine/base_preprocessing_layer.py
index 0818825..318323b 100644
--- a/tensorflow/python/keras/engine/base_preprocessing_layer.py
+++ b/tensorflow/python/keras/engine/base_preprocessing_layer.py
@@ -271,7 +271,7 @@
values, default_value=sparse_default_value)
values = K.get_value(dense_tensor)
- if isinstance(values, (ops.EagerTensor, ops.Tensor)):
+ if isinstance(values, ops.Tensor):
values = K.get_value(values)
# We may get passed a ndarray or the code above may give us a ndarray.
diff --git a/tensorflow/python/keras/engine/training_arrays_v1.py b/tensorflow/python/keras/engine/training_arrays_v1.py
index df4b8cd..ad9b37f 100644
--- a/tensorflow/python/keras/engine/training_arrays_v1.py
+++ b/tensorflow/python/keras/engine/training_arrays_v1.py
@@ -153,10 +153,6 @@
use_steps = is_dataset or steps_per_epoch is not None
do_validation = val_inputs is not None
- # Convert Eager Tensors to NumPy arrays to support batching/shuffling.
- inputs, targets, sample_weights = training_utils_v1. \
- convert_eager_tensors_to_numpy((inputs, targets, sample_weights))
-
# Prepare input data.
inputs = input_iterator or inputs
if validation_in_fit and prepared_feed_values_from_dataset:
diff --git a/tensorflow/python/keras/engine/training_utils_v1.py b/tensorflow/python/keras/engine/training_utils_v1.py
index aa0de60..c198bad 100644
--- a/tensorflow/python/keras/engine/training_utils_v1.py
+++ b/tensorflow/python/keras/engine/training_utils_v1.py
@@ -186,9 +186,6 @@
else:
self.results = np.concatenate(self.results, axis=0)
- if isinstance(self.results, ops.EagerTensor):
- self.results = self.results._numpy() # pylint: disable=protected-access
-
_COPY_THREADS = 4
_COPY_POOL = None
@@ -250,8 +247,6 @@
# initialization is effectively instantaneous.
shape = (self.num_samples,) + batch_element.shape[1:]
dtype = batch_element.dtype
- if isinstance(batch_element, ops.EagerTensor):
- dtype = dtype.as_numpy_dtype
self.results = np.empty(shape=shape, dtype=dtype)
@@ -328,7 +323,7 @@
# or a composite tensor's Value object. In either case, we can't
# allocate an array to hold the object - we'll handle it later.
self.results.append(ConcatAggregator(self.batch_size))
- elif isinstance(batch_element, (np.ndarray, ops.EagerTensor)):
+ elif isinstance(batch_element, np.ndarray):
self.results.append(
(ConcatAggregator(self.batch_size) if self.use_steps else
SliceAggregator(self.num_samples, self.batch_size)))
@@ -866,12 +861,13 @@
'you should pass a 2D sample_weight array.')
else:
if sample_weight is not None and len(sample_weight.shape) != 1:
- raise ValueError('Found a sample_weight array with shape {}. In order to '
- 'use timestep-wise sample weights, you should specify '
- 'sample_weight_mode="temporal" in compile(); found "{}" '
- 'instead. If you just mean to use sample-wise weights, '
- 'make sure your sample_weight array is 1D.'
- .format(sample_weight.shape, sample_weight_mode))
+ raise ValueError(
+ 'Found a sample_weight array with shape {}. In order to '
+ 'use timestep-wise sample weights, you should specify '
+ 'sample_weight_mode="temporal" in compile(); founssd "{}" '
+ 'instead. If you just mean to use sample-wise weights, '
+ 'make sure your sample_weight array is 1D.'.format(
+ sample_weight.shape, sample_weight_mode))
if sample_weight is not None:
if len(sample_weight.shape) > len(y.shape):
@@ -1628,7 +1624,7 @@
if v.ndim == 1:
v = np.expand_dims(v, 1)
- if isinstance(v, (np.ndarray, ops.EagerTensor)):
+ if isinstance(v, np.ndarray):
# We fix the placeholder shape except the batch size.
# This is suboptimal, but it is the best we can do with the info
# we have. The user should call `model._set_inputs(placeholders)`
@@ -1675,25 +1671,6 @@
return ['output_%d' % (i + 1) for i in range(len(outputs_list))]
-def convert_eager_tensors_to_numpy(structure):
- """Convert every EagerTensor in `structure` to NumPy.
-
- Arguments:
- structure: An arbitrary structure of elements to be converted to NumPy
- arrays.
-
- Returns:
- An identical structure with EagerTensors converted to NumPy arrays.
- """
-
- def _convert(element):
- if isinstance(element, ops.EagerTensor):
- return element.numpy()
- return element
-
- return nest.map_structure(_convert, structure)
-
-
def should_run_validation(validation_freq, epoch):
"""Checks if validation should be run this epoch.
diff --git a/tensorflow/python/keras/engine/training_v1.py b/tensorflow/python/keras/engine/training_v1.py
index 69a60e0..54969bb 100644
--- a/tensorflow/python/keras/engine/training_v1.py
+++ b/tensorflow/python/keras/engine/training_v1.py
@@ -3141,7 +3141,7 @@
def _is_symbolic_tensor(x):
- return tensor_util.is_tensor(x) and not isinstance(x, ops.EagerTensor)
+ return tensor_util.is_tensor(x)
def _convert_scipy_sparse_tensor(value, expected_input):
