def testContainsIndexedSlices_PerReplica(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
per_replica = value_lib.PerReplica((t0, t1))
self.assertTrue(cross_device_utils.contains_indexed_slices(per_replica))
def testContainsIndexedSlices_PerDevice(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
per_device = value_lib.PerDevice({"/gpu:0": t0, "/cpu:0": t1})
self.assertTrue(cross_tower_utils.contains_indexed_slices(per_device))
def testContainsIndexedSlices_PerReplica(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
per_replica = value_lib.PerReplica({"/gpu:0": t0, "/cpu:0": t1})
self.assertTrue(cross_device_utils.contains_indexed_slices(per_replica))
def testMultipleGradients(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
total = constant_op.constant([[1., 2.], [5, 6], [10., 12.]])
result = gradients_impl._AggregateIndexedSlicesGradients([t0, t1])
self._assert_indexed_slices_equal(total, result)
def testAggregateIndexedSlices(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
total = constant_op.constant([[1., 2.], [5, 6], [10., 12.]])
result = cross_device_utils.aggregate_tensors_or_indexed_slices([t0, t1])
self.assertIsInstance(result, ops.IndexedSlices)
self._assert_values_equal(total, result)
def testMultipleGradients(self):
t0 = math_ops._as_indexed_slices(constant_op.constant(
[[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(constant_op.constant(
[[0., 0.], [5, 6], [7., 8.]]))
total = constant_op.constant(
[[1., 2.], [5, 6], [10., 12.]])
result = gradients_impl._AggregateIndexedSlicesGradients([t0, t1])
self._assert_indexed_slices_equal(total, result)
def testAggregateIndexedSlices(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
total = constant_op.constant([[1., 2.], [5, 6], [10., 12.]])
result = cross_device_utils.aggregate_tensors_or_indexed_slices([t0, t1])
self.assertIsInstance(result, ops.IndexedSlices)
self._assert_values_equal(total, result)
def testContainsIndexedSlices_PerDeviceMapOutput(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
per_device = value_lib.PerDevice({
"/gpu:0": value_lib.MapOutput([t0]),
"/cpu:0": value_lib.MapOutput([t1])})
self.assertTrue(cross_tower_utils.contains_indexed_slices(per_device))
def testContainsIndexedSlices_PerReplicaMapOutput(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
per_replica = value_lib.PerReplica({
"/gpu:0": value_lib.MapOutput([t0]),
"/cpu:0": value_lib.MapOutput([t1])})
self.assertTrue(cross_tower_utils.contains_indexed_slices(per_replica))
def testDivideIndexedSlices(self):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
n = 2
expected = constant_op.constant([[0.5, 1.], [0, 0], [1.5, 2.]])
result = cross_device_utils.divide_by_n_tensors_or_indexed_slices(t, n)
self.assertIsInstance(result, ops.IndexedSlices)
self._assert_values_equal(expected, result)
def testDivideIndexedSlices(self):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
n = 2
expected = constant_op.constant([[0.5, 1.], [0, 0], [1.5, 2.]])
result = cross_device_utils.divide_by_n_tensors_or_indexed_slices(t, n)
self.assertIsInstance(result, ops.IndexedSlices)
self._assert_values_equal(expected, result)
def testIndexedSlicesToTensor(self):
with self.test_session():
np_val = np.random.rand(4, 4, 4, 4).astype(np.float32)
c = constant_op.constant(np_val)
c_sparse = math_ops._as_indexed_slices(c)
self.assertAllEqual(np_val.shape, c_sparse.dense_shape.eval())
c_dense = math_ops.mul(c_sparse, 1.0)
self.assertAllClose(np_val, c_dense.eval())
def _make_output_composite_tensors_match(true_graph, false_graph):
"""Modifies true_graph and false_graph so they have the same output signature.
Currently the only transformation implemented is turning a Tensor into an
equivalent IndexedSlices if the other branch returns an IndexedSlices.
Updates {true,false}_graph.{outputs,structured_outputs}.
Args:
true_graph: FuncGraph
false_graph: FuncGraph
Raises:
TypeError: if a pair of outputs cannot be rewritten.
"""
# Note: since this is only used for gradient graphs, we do not expect the
# outputs to be structured (e.g. nested lists), and thus do not need to use
# nest.flatten, etc.
true_outputs = list(true_graph.structured_outputs)
false_outputs = list(false_graph.structured_outputs)
assert len(true_outputs) == len(false_outputs)
for idx, (true_out,
false_out) in enumerate(zip(true_outputs, false_outputs)):
if type(true_out) == type(false_out): # pylint: disable=unidiomatic-typecheck
continue
if (isinstance(true_out, ops.IndexedSlices)
and isinstance(false_out, ops.Tensor)):
with false_graph.as_default():
false_outputs[idx] = math_ops._as_indexed_slices(false_out)
elif (isinstance(true_out, ops.Tensor)
and isinstance(false_out, ops.IndexedSlices)):
with true_graph.as_default():
true_outputs[idx] = math_ops._as_indexed_slices(true_out)
else:
raise TypeError("Cannot reconcile tf.cond %i-th outputs:\n"
" true_fn returned: %s\n"
" false_fn returned: %s" %
(idx, true_out, false_out))
true_graph.structured_outputs = true_outputs
true_graph.outputs = func_graph_module.flatten(true_outputs)
false_graph.structured_outputs = false_outputs
false_graph.outputs = func_graph_module.flatten(false_outputs)
def _make_output_composite_tensors_match(true_graph, false_graph):
"""Rewrites {true,false}_graph's outputs to use the same _TensorLike classes.
Currently the only transformation implemented is turning a Tensor into an
equivalent IndexedSlices if the other branch returns an IndexedSlices.
Updates {true,false}_graph.{outputs,structured_outputs}.
Args:
true_graph: FuncGraph
false_graph: FuncGraph
Raises:
TypeError: if a pair of outputs cannot be rewritten.
"""
# Note: since this is only used for gradient graphs, we do not expect the
# outputs to be structured (e.g. nested lists), and thus do not need to use
# nest.flatten, etc.
true_outputs = list(true_graph.structured_outputs)
false_outputs = list(false_graph.structured_outputs)
assert len(true_outputs) == len(false_outputs)
for idx, (true_out, false_out) in enumerate(zip(true_outputs, false_outputs)):
if type(true_out) == type(false_out): # pylint: disable=unidiomatic-typecheck
continue
if (isinstance(true_out, ops.IndexedSlices) and
isinstance(false_out, ops.Tensor)):
with false_graph.as_default():
false_outputs[idx] = math_ops._as_indexed_slices(false_out)
elif (isinstance(true_out, ops.Tensor) and
isinstance(false_out, ops.IndexedSlices)):
with true_graph.as_default():
true_outputs[idx] = math_ops._as_indexed_slices(true_out)
else:
raise TypeError(
"Cannot reconcile tf.cond %i-th outputs:\n"
" true_fn returned: %s\n"
" false_fn returned: %s" % (idx, true_out, false_out))
true_graph.structured_outputs = true_outputs
true_graph.outputs = func_graph_module.flatten(true_outputs)
false_graph.structured_outputs = false_outputs
false_graph.outputs = func_graph_module.flatten(false_outputs)
def testInt64Indices(self):
with self.test_session():
np_val = np.random.rand(4, 4, 4, 4).astype(np.float32)
c = constant_op.constant(np_val)
c_sparse = math_ops._as_indexed_slices(c)
c_sparse = ops.IndexedSlices(
c_sparse.values, math_ops.cast(c_sparse.indices, dtypes.int64),
c_sparse.dense_shape)
self.assertAllEqual(np_val.shape, c_sparse.dense_shape.eval())
c_dense = math_ops.mul(c_sparse, 1.0)
self.assertAllClose(np_val, c_dense.eval())
def testCopyIndexedSlices(self):
with ops.device("/cpu:0"):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
destination = "/gpu:0"
result = cross_device_utils.copy_tensor_or_indexed_slices_to_device(
t, destination)
self.assertIsInstance(result, ops.IndexedSlices)
self._assert_values_equal(t, result)
self.assertEqual(device_util.resolve(destination),
device_util.resolve(result.device))
def testCopyIndexedSlices(self):
with ops.device("/cpu:0"):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
destination = "/gpu:0"
result = cross_device_utils.copy_tensor_or_indexed_slices_to_device(
t, destination)
self.assertIsInstance(result, ops.IndexedSlices)
self._assert_values_equal(t, result)
self.assertEqual(
device_util.resolve(destination), device_util.resolve(result.device))
def _make_output_composite_tensors_match(op_type, branch_graphs):
"""Modifies each branch_graph's outputs to have the same output signature.
Currently the only transformation implemented is turning a Tensor into an
equivalent IndexedSlices if the other branch returns an IndexedSlices.
Updates branch_graph.{outputs,structured_outputs} for each branch_graph in
branch_graphs.
Args:
op_type: _COND or _CASE
branch_graphs: `list` of `FuncGraph`
Raises:
TypeError: if a set of outputs cannot be rewritten.
"""
# Note: since this is only used for gradient graphs, we do not expect the
# outputs to be structured (e.g. nested lists), and thus do not need to use
# nest.flatten, etc.
assert branch_graphs
branch_outputs = [g.structured_outputs for g in branch_graphs]
outputs_per_branch = list(len(outs) for outs in branch_outputs)
assert len(set(outputs_per_branch)) == 1, outputs_per_branch
for output_idx, branch_outs in enumerate(zip(*branch_outputs)):
if len(set(type(out) for out in branch_outs)) == 1:
continue
if not any(isinstance(out, ops.IndexedSlices) for out in branch_outs):
continue
for branch_idx, branch_out in enumerate(branch_outs):
if isinstance(branch_out, ops.IndexedSlices):
continue
elif isinstance(branch_out, ops.Tensor):
with branch_graphs[branch_idx].as_default():
branch_outputs[branch_idx][
output_idx] = math_ops._as_indexed_slices(branch_out)
else:
raise TypeError(
"Cannot reconcile {op_name} {output_idx}-th outputs:\n"
" outputs from all branches: {outputs}".format(
op_name="tf.cond"
if op_type == _COND else "tf.switch_case",
output_idx=output_idx,
outputs=branch_outs))
for branch_graph, branch_outs in zip(branch_graphs, branch_outputs):
branch_graph.structured_outputs = branch_outs
branch_graph.outputs = [
t for t in func_graph_module.flatten(branch_outs) if t is not None
]
def testIndexedSlicesToTensorList(self):
with self.test_session():
numpy_list = []
dense_list = []
sparse_list = []
for _ in range(3):
np_val = np.random.rand(4, 4, 4, 4).astype(np.float32)
c = constant_op.constant(np_val)
c_sparse = math_ops._as_indexed_slices(c)
numpy_list.append(np_val)
dense_list.append(c)
sparse_list.append(c_sparse)
packed_dense = array_ops.pack(dense_list)
packed_sparse = array_ops.pack(sparse_list)
self.assertAllClose(packed_dense.eval(), packed_sparse.eval())
def testContainsIndexedSlices_List(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
self.assertTrue(cross_tower_utils.contains_indexed_slices([t0, t1]))
def testContainsIndexedSlices_List(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
self.assertTrue(cross_tower_utils.contains_indexed_slices([t0, t1]))
def testOneGradient(self):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
result = gradients_impl._AggregateIndexedSlicesGradients([t])
self._assert_indexed_slices_equal(t, result)
def false_fn():
return (None, None, None,
math_ops._as_indexed_slices(constant_op.constant([2.])))
def true_fn():
return (None, None, None,
math_ops._as_indexed_slices(constant_op.constant([1.])))
def testIsIndexedSlices(self):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
self.assertTrue(cross_device_utils.contains_indexed_slices(t))
def testContainsIndexedSlices_Tuple(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
self.assertTrue(cross_device_utils.contains_indexed_slices((t0, t1)))
def testIsIndexedSlices(self):
t = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
self.assertTrue(cross_device_utils.contains_indexed_slices(t))
def testContainsIndexedSlices_Tuple(self):
t0 = math_ops._as_indexed_slices(
constant_op.constant([[1., 2.], [0, 0], [3., 4.]]))
t1 = math_ops._as_indexed_slices(
constant_op.constant([[0., 0.], [5, 6], [7., 8.]]))
self.assertTrue(cross_device_utils.contains_indexed_slices((t0, t1)))
def testOneGradient(self):
t = math_ops._as_indexed_slices(constant_op.constant(
[[1., 2.], [0, 0], [3., 4.]]))
result = gradients_impl._AggregateIndexedSlicesGradients([t])
self._assert_indexed_slices_equal(t, result)