def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
body_graph = _get_body_graph(op)
# Replace None gradients with zeros. This is needed because `grads` could have
# None incoming gradients for the TensorLists. If we pass None's through, the
# custom gradient of TensorListPopBack will create an EmptyTensorList inside
# the FuncGraph which is undesirable.
# TODO(b/80444525): There might be an issue with treating no gradient as zero
# gradient in certain cases. Consider replacing None gradients with Zeros
# for accumulators only.
grads = [
g if g is not None else array_ops.zeros_like(output)
for g, output in zip(grads, op.outputs)
]
body_grad_graph, args = _create_grad_func(
body_graph, grads,
util.unique_grad_fn_name(body_graph.name), op)
intermediate_tensors = _get_intermediates(body_grad_graph)
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(intermediate_tensor.shape))
with body_grad_graph.as_default():
tensor_list_ph = body_grad_graph.capture(tensor_list, whitelisted=True)
# Push the intermediate tensor to the tensor list.
appended_tensor_list = list_ops.tensor_list_push_back(tensor_list_ph,
intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_grad_graph.outputs.append(appended_tensor_list)
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
loop_vars = args + body_grad_graph.external_captures
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name, grad_cond, loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
assert len(loop_vars) == len(body_grad_graph.inputs)
assert len(loop_vars) == len(body_grad_graph.outputs)
assert len(loop_vars) == len(cond_grad_graph.inputs)
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
name="%s_grad" % op.name)
_copy_handle_data(body_grad_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
# outputs[0] is the loop counter.
# outputs[1] is the total number of loop iterations.
return outputs[2:2 + len(op.inputs)]
def While(input_, cond, body, name=None, hostmem=None):
r"""output = input; While (Cond(output)) { output = Body(output) }.
Args:
input_: A list of `Tensor` objects.
A list of input tensors whose types are T.
cond: . A function takes 'input' and returns a tensor. If the tensor is
a scalar of non-boolean, the scalar is converted to a boolean
according to the following rule: if the scalar is a numerical
value, non-zero means True and zero means False; if the scalar is
a string, non-empty means True and empty means False. If the
tensor is not a scalar, non-emptiness means True and False
otherwise.
body: . A funcion takes a list of tensors and returns another
list tensors. Both lists have the same types as specified
by T.
name: A name for the operation (optional).
hostmem: A list of integer. If i is in the list, input[i] is a
host memory tensor.
Returns:
A list of `Tensor` objects. Has the same type as `input`.
A list of output tensors whose types are T.
"""
ret = gen_functional_ops._while(input_, cond, body, name=name)
if hostmem:
input_attr = attr_value_pb2.AttrValue()
input_attr.list.i.extend(hostmem)
ret[0].op._set_attr("_input_hostmem", input_attr) # pylint: disable=protected-access
output_attr = attr_value_pb2.AttrValue()
output_attr.list.i.extend(hostmem)
ret[0].op._set_attr("_output_hostmem", output_attr) # pylint: disable=protected-access
return ret
def While(input_, cond, body, name=None, hostmem=None):
r"""output = input; While (Cond(output)) { output = Body(output) }.
Args:
input_: A list of `Tensor` objects.
A list of input tensors whose types are T.
cond: . A function takes 'input' and returns a tensor. If the tensor is
a scalar of non-boolean, the scalar is converted to a boolean
according to the following rule: if the scalar is a numerical
value, non-zero means True and zero means False; if the scalar is
a string, non-empty means True and empty means False. If the
tensor is not a scalar, non-emptiness means True and False
otherwise.
body: . A function takes a list of tensors and returns another
list tensors. Both lists have the same types as specified
by T.
name: A name for the operation (optional).
hostmem: A list of integer. If i is in the list, input[i] is a
host memory tensor.
Raises:
ValueError: if `cond` has implicitly captured inputs or if `cond` and `body`
have different signatures.
Returns:
A list of `Tensor` objects. Has the same type as `input`.
A list of output tensors whose types are T.
"""
if cond.captured_inputs:
raise ValueError("While op 'cond' argument must be a function "
"without implicitly captured inputs.")
if cond.declared_input_types != body.declared_input_types:
raise ValueError(
"While op 'cond' and 'body' signatures do not match. %r vs %r" %
(cond.declared_input_types, body.declared_input_types))
if body.captured_inputs:
cond_dtypes = list(
body.declared_input_types) + [t.dtype for t in body.captured_inputs]
@function.Defun(*cond_dtypes, func_name="%s_Wrapper" % cond.name)
def CondWrapper(*args):
"""A wrapper that handles loop-carried captured inputs."""
return cond(*args[:len(body.declared_input_types)])
ret = gen_functional_ops._while(
input_ + body.captured_inputs,
CondWrapper,
_LoopBodyCaptureWrapper(body),
name=name)
# Slice off the loop-carried captured inputs.
ret = ret[:-len(body.captured_inputs)]
else:
ret = gen_functional_ops._while(input_, cond, body, name=name)
if hostmem:
input_attr = attr_value_pb2.AttrValue()
input_attr.list.i.extend(hostmem)
ret[0].op._set_attr("_input_hostmem", input_attr) # pylint: disable=protected-access
output_attr = attr_value_pb2.AttrValue()
output_attr.list.i.extend(hostmem)
ret[0].op._set_attr("_output_hostmem", output_attr) # pylint: disable=protected-access
return ret
def while_loop(cond, body, loop_vars, shape_invariants=None, name=None):
"""Like tf.while_loop, except emits a single While op."""
flattened_loop_vars = nest.flatten(loop_vars)
if shape_invariants is not None:
nest.assert_same_structure(loop_vars, shape_invariants)
flattened_shapes = nest.flatten(shape_invariants)
else:
flattened_shapes = [t.shape for t in flattened_loop_vars]
del shape_invariants
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = _get_unique_name(("%scond" % scope).replace("/", "_"))
body_name = _get_unique_name(("%sbody" % scope).replace("/", "_"))
num_outputs = len(flattened_loop_vars)
# Add loop counter needed for computing gradients.
flattened_loop_vars = [constant_op.constant(0., name="loop_counter")
] + flattened_loop_vars
flattened_shapes = [tensor_shape.scalar()] + flattened_shapes
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(unused_loop_counter, *loop_vars):
return cond(*loop_vars)
signature = [
tensor_spec.TensorSpec(shape, t.dtype)
for shape, t in zip(flattened_shapes, flattened_loop_vars)
]
cond_graph = function.func_graph_from_py_func(
cond_name, wrapped_cond, flattened_loop_vars, {}, signature=signature)
# Add external_captures of cond to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
flattened_loop_vars = flattened_loop_vars + cond_graph.external_captures
flattened_shapes = flattened_shapes + [
t.shape for t in cond_graph.external_captures
]
def wrapped_body(loop_counter, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
*args: List of args
args[:num_outputs] - Args for the original loop body.
args[num_outputs:] - External captures of cond. These get passed
through as is.
Returns:
A list of tensors the same length as args.
"""
outputs = body(*args[:num_outputs])
if not isinstance(outputs, collections.Sequence):
outputs = [outputs]
# Return the external_captures of cond_graph as is, i.e., treat them as
# loop invariants.
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1] + list(outputs) + list(args[num_outputs:])
signature = [
tensor_spec.TensorSpec(shape, t.dtype)
for shape, t in zip(flattened_shapes, flattened_loop_vars)
]
body_graph = function.func_graph_from_py_func(
body_name, wrapped_body, flattened_loop_vars, {}, signature=signature)
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
flattened_loop_vars = flattened_loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture `external_captures` of `body_graph` in `cond_graph` so that it
# expects to receive those as arguments.
# TODO(srbs): Dedup tensors that are captured in both the cond and body.
# This logic already exists in cond_v2.
with cond_graph.as_default():
for external_capture in body_graph.external_captures:
cond_graph.capture(external_capture)
# Export all tensors in the loop body that may be needed for gradient
# computation. We do this by accumulating the intermediate values in
# TensorLists.
intermediate_tensors = _get_intermediates(body_graph)
for intermediate_tensor in intermediate_tensors:
# TODO(srbs): Cache and re-use empty tensor lists.
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape))
flattened_loop_vars.append(tensor_list)
with cond_graph.as_default():
# Add a placeholder to cond_graph's inputs corresponding to the
# tensor_list.
cond_graph.capture(tensor_list)
with body_graph.as_default():
# Push the intermediate tensor to the tensor list. This captures the
# `tensor_list` as well.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list,
intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_graph.outputs.append(appended_tensor_list)
# Make sure that the shapes of the loop outputs are compatible with the
# shape invariants, or the shapes of the loop vars if the invariants are not
# specified.
_check_shapes_compat(body_graph.outputs[1:1 + num_outputs],
flattened_shapes[1:1 + num_outputs],
flattened_loop_vars[1:1 + num_outputs])
outputs = gen_functional_ops._while(
flattened_loop_vars,
cond_v2._create_new_tf_function(cond_graph),
cond_v2._create_new_tf_function(body_graph),
output_shapes=[t.shape for t in body_graph.outputs],
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
# First var is loop counter.
if num_outputs == 1:
return outputs[1]
else:
return nest.pack_sequence_as(loop_vars, outputs[1:1 + num_outputs])
def while_loop(cond,
body,
loop_vars,
shape_invariants=None,
maximum_iterations=None,
name=None,
return_same_structure=True):
"""Like tf.while_loop, except emits a single While op."""
maximum_iterations = _validate_and_convert_to_tensor(maximum_iterations)
# Keep the original loop_vars around to know which args were TensorArrays.
orig_loop_vars = loop_vars
# Cache its length since we use it at multiple places below.
len_orig_loop_vars = len(orig_loop_vars)
# Convert TensorArrays to their flow variables. These get converted back to
# TensorArrays before calling `cond` and `body`. See `wrapped_cond` and
# `wrapped_body` below.
loop_vars = list(_tensor_array_to_flow(orig_loop_vars))
loop_vars = nest.map_structure(
ops.internal_convert_to_tensor_or_indexed_slices, loop_vars)
if shape_invariants is not None:
nest.assert_same_structure(orig_loop_vars, shape_invariants)
else:
shape_invariants = nest.map_structure(lambda t: t.shape, loop_vars)
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = util.unique_fn_name(scope, "cond")
body_name = util.unique_fn_name(scope, "body")
loop_counter = constant_op.constant(
0,
dtype=maximum_iterations.dtype
if maximum_iterations is not None else None,
name="loop_counter")
# Add loop counter needed for computing gradients.
loop_vars = [loop_counter] + loop_vars
shape_invariants = type(shape_invariants)([tensor_shape.scalar()
]) + shape_invariants
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(loop_counter, *args):
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
if maximum_iterations is None:
return cond(*_pack_sequence_as(orig_loop_vars, args))
else:
return math_ops.logical_and(
loop_counter < maximum_iterations,
cond(*_pack_sequence_as(orig_loop_vars, args)))
cond_graph = func_graph_module.func_graph_from_py_func(
cond_name,
wrapped_cond,
loop_vars, {},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileCondFuncGraph(cond_name),
add_control_dependencies=add_control_dependencies)
# Add external_captures of cond to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
loop_vars = loop_vars + cond_graph.external_captures
shape_invariants = shape_invariants + type(shape_invariants)(
[t.shape for t in cond_graph.external_captures])
def wrapped_body(loop_counter, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
*args: List of args
args[:len_orig_loop_vars] - Args for the original loop body.
args[len_orig_loop_vars:] - External captures of cond. These get
passed through as is.
Returns:
A list of tensors the same length as args.
"""
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
outputs = body(
*_pack_sequence_as(orig_loop_vars, args[:len_orig_loop_vars]))
if not nest.is_sequence(outputs):
outputs = [outputs]
# Compare the structure of input and output of body converting the
# top-level tuples to list to be compatible with legacy while_loop.
nest.assert_same_structure(list(outputs), list(orig_loop_vars))
outputs = _tensor_array_to_flow(outputs)
# Return the external_captures of cond_graph as is, i.e., treat them as
# loop invariants.
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1] + list(outputs) + list(
args[len_orig_loop_vars:])
body_graph = func_graph_module.func_graph_from_py_func(
body_name,
wrapped_body,
loop_vars, {},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileBodyFuncGraph(body_name),
add_control_dependencies=add_control_dependencies)
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
loop_vars = loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture `external_captures` of `body_graph` in `cond_graph` so that it
# expects to receive those as arguments.
# TODO(b/118457764): Dedup tensors that are captured in both the cond and
# body. This logic already exists in cond_v2.
with cond_graph.as_default():
for external_capture in body_graph.external_captures:
assert external_capture not in cond_graph.captures, (
"Looks like both cond and body are capturing the same tensor %s. "
"This is not supported yet. For now consider passing,"
" this as a loop variable." % str(external_capture))
cond_graph.capture(external_capture)
# Export all tensors in the loop body that may be needed for gradient
# computation. We do this by accumulating the intermediate values in
# TensorLists.
intermediate_tensors = _get_intermediates(body_graph)
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=intermediate_tensor.shape,
max_num_elements=maximum_iterations)
loop_vars.append(tensor_list)
with cond_graph.as_default():
# Add a placeholder to cond_graph's inputs corresponding to the
# tensor_list.
cond_graph.capture(tensor_list)
with body_graph.as_default():
# Push the intermediate tensor to the tensor list. This captures the
# `tensor_list` as well.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list,
intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_graph.outputs.append(appended_tensor_list)
# Make sure that the shapes of the loop outputs are compatible with the
# shape invariants, or the shapes of the loop vars if the invariants are not
# specified.
num_flattened_outputs = len(nest.flatten(orig_loop_vars))
_check_shapes_compat(
body_graph.outputs[1:1 + num_flattened_outputs],
nest.flatten(shape_invariants[1:1 + len_orig_loop_vars]),
nest.flatten(loop_vars[1:1 + len_orig_loop_vars]))
flattened_loop_vars = nest.flatten(loop_vars)
_check_num_inputs_outputs(cond_graph, body_graph,
len(flattened_loop_vars))
outputs = gen_functional_ops._while(
flattened_loop_vars,
util.create_new_tf_function(cond_graph),
util.create_new_tf_function(body_graph),
output_shapes=[t.shape for t in body_graph.outputs],
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
util.maybe_set_lowering_attr(outputs[0].op)
_maybe_set_maximum_iterations_attr(outputs[0].op, maximum_iterations)
# Return identities for each output of the While op, rather than the output
# of the While op directly. This makes pruning work if the output of
# while_loop() is fetched: the lowering pass converts the While outputs into
# IdentityN outputs, which if fetched will cause all ops in the body to be
# run (since it takes all exit ops as input). After lowering, each output
# identity op will end up with only the appropriate exit op as input.
outputs = tuple(array_ops.identity(t) for t in outputs)
# First var is loop counter.
outputs = _pack_sequence_as(orig_loop_vars,
outputs[1:1 + num_flattened_outputs])
if return_same_structure:
return outputs
flattened_outputs = nest.flatten(outputs)
if len(flattened_outputs) == 1:
return flattened_outputs[0]
else:
return outputs
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
body_graph = _get_body_graph(op)
# Set the incoming gradient of TensorArray handles to None. The gradient
# implementation currently assumes all resource tensors correspond to float32
# ResourceVariables, which can lead to runtime shape errors when used with a
# TensorArray. This is a workaround until TensorArrays are reimplemented with
# TensorLists instead of resources.
# Also set the incoming gradient of non-trainable inputs to None. It is
# possible that we receive non-None gradients for non-trainable types in
# nested while loops because we accumulate outputs of the inner while as
# variant tensors which are trainable and hence receive zeros_like tensors in
# the gradient pass. The non-trainable tensors then receive the popped zeros
# tensor from this zeros variant. The gradient for the loop vars corresponding
# to these tensors is None or zeros (this happens only if the loop var is
# accumulated as well) in _grad_fn so we reset these.
# TODO(b/118712257): Remove the IsTrainable filter once we can handle None
# output grads in _grad_fn.
grads = [
None if _is_tensor_array_handle(output) or
not gradients_impl.IsTrainable(output) else grad
for grad, output in zip(grads, op.outputs)
]
# Ensure that all non-resource trainable outputs have incoming gradients.
assert all(g is not None or o.dtype == dtypes.resource or
not gradients_impl.IsTrainable(o)
for o, g in zip(op.outputs, grads)
), "All trainable loop vars must receive incoming gradients."
# We compute the gradient for the sub-graph between trainable ys and xs
# with non-None incoming gradients. We later pad the None's to the list of
# outputs.
ys, xs, non_none_grads = zip(*[(y, x, grad) for (y, x, grad) in zip(
body_graph.outputs, body_graph.inputs, grads) if grad is not None])
body_grad_graph, args = _create_grad_func(
ys, xs, non_none_grads, body_graph,
util.unique_grad_fn_name(body_graph.name), op)
intermediate_tensors = _get_intermediates(body_grad_graph)
maximum_iterations = op.get_attr(
"_maximum_iterations") if _is_in_xla_context() else None
assert not _is_in_xla_context() or maximum_iterations is not None
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=intermediate_tensor.shape,
max_num_elements=maximum_iterations)
with body_grad_graph.as_default():
tensor_list_ph = body_grad_graph.capture(tensor_list, whitelisted=True)
# Push the intermediate tensor to the tensor list.
appended_tensor_list = list_ops.tensor_list_push_back(tensor_list_ph,
intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_grad_graph.outputs.append(appended_tensor_list)
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
loop_vars = args + body_grad_graph.external_captures
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name, grad_cond, loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
_check_num_inputs_outputs(cond_grad_graph, body_grad_graph, len(loop_vars))
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
name="%s_grad" % op.name)
_copy_handle_data(body_grad_graph.outputs, outputs)
util.maybe_set_lowering_attr(outputs[0].op)
_maybe_set_maximum_iterations_attr(outputs[0].op, maximum_iterations)
# See comment in while_loop.
outputs = [array_ops.identity(t) for t in outputs]
# Set None as the output gradient for tensors with None input gradient
# e.g. TensorArray handles.
# outputs[0] is the loop counter.
# outputs[1] is the total number of loop iterations.
index = 2
none_padded_outputs = []
for g in grads:
if g is None:
none_padded_outputs.append(None)
else:
none_padded_outputs.append(outputs[index])
index += 1
return none_padded_outputs
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
body_graph = _get_body_graph(op)
# Set the incoming gradient of TensorArray handle to None.
# TODO(b/118164915): We need a way of distinguising b/w TensorArray resource
# handles and ResourceVariables and set the default gradient of only the
# TensorArray handle to None.
grads = [
None if output.dtype == dtypes.resource else g
for g, output in zip(grads, op.outputs)
]
# Ensure that all non-resource trainable outputs have incoming gradients.
assert all(g is not None or o.dtype == dtypes.resource
or not gradients_impl.IsTrainable(o)
for o, g in zip(op.outputs, grads)
), "All trainable loop vars must receive incoming gradients."
# We compute the gradient for the sub-graph between trainable ys and xs
# with non-None incoming gradients. We later pad the None's to the list of
# outputs.
ys, xs, non_none_grads = zip(
*[(y, x, grad)
for (y, x, grad) in zip(body_graph.outputs, body_graph.inputs, grads)
if grad is not None])
body_grad_graph, args = _create_grad_func(
ys, xs, non_none_grads, body_graph,
util.unique_grad_fn_name(body_graph.name), op)
intermediate_tensors = _get_intermediates(body_grad_graph)
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape))
with body_grad_graph.as_default():
tensor_list_ph = body_grad_graph.capture(tensor_list,
whitelisted=True)
# Push the intermediate tensor to the tensor list.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list_ph, intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_grad_graph.outputs.append(appended_tensor_list)
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
loop_vars = args + body_grad_graph.external_captures
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name,
grad_cond,
loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
assert len(loop_vars) == len(body_grad_graph.inputs)
assert len(loop_vars) == len(body_grad_graph.outputs)
assert len(loop_vars) == len(cond_grad_graph.inputs)
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
name="%s_grad" % op.name)
_copy_handle_data(body_grad_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
# Set None as the output gradient for tensors with None input gradient
# e.g. TensorArray handles.
# outputs[0] is the loop counter.
# outputs[1] is the total number of loop iterations.
index = 2
none_padded_outputs = []
for g in grads:
if g is None:
none_padded_outputs.append(None)
else:
none_padded_outputs.append(outputs[index])
index += 1
return none_padded_outputs
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
# Note that op is not always the same as while_op because the gradient tape,
# for eager mode compatibility, forgets information about the proper op. Since
# the loop cannot run in eager mode, however, we can safely introspect into
# the graph here.
while_op = op.outputs[0].op
cond_graph = _get_graph(while_op, "cond")
body_graph = _get_graph(while_op, "body")
orig_num_params = len(body_graph.outputs)
maximum_iterations = op.inputs[1]
parallel_iterations = op.get_attr("parallel_iterations")
grads = [
_preprocess_grad(grad, body_out,
while_out) for grad, body_out, while_out in zip(
grads, body_graph.outputs, while_op.outputs)
]
# We compute the gradient for the sub-graph between trainable ys and xs
# with non-None incoming gradients. We later pad the None's to the list of
# outputs.
ys, xs, non_none_grads = zip(
*[(y, x, grad)
for (y, x, grad) in zip(body_graph.outputs, body_graph.inputs, grads)
if grad is not None])
body_grad_graph, args = _create_grad_func(
ys, xs, non_none_grads, cond_graph, body_graph,
util.unique_grad_fn_name(body_graph.name), op, maximum_iterations)
if body_grad_graph.while_op_needs_rewrite:
# Modify 'op' to output the intermediate accumulators needed by the grad
# function.
# NOTE(skyewm): if there are any active sessions, this modification to `op`
# may make them unrunnable!
cond_graph.name += "_rewritten"
body_graph.name += "_rewritten"
new_inputs = body_grad_graph.empty_tensor_lists
new_outputs = body_graph.outputs[orig_num_params:]
while_op._set_func_attr("cond",
util.create_new_tf_function(cond_graph))
while_op._set_func_attr("body",
util.create_new_tf_function(body_graph))
while_op._set_type_list_attr("T", body_graph.output_types)
while_op._set_shape_list_attr("output_shapes",
body_graph.output_shapes)
while_op._add_while_inputs(new_inputs)
while_op._add_outputs([t.dtype for t in new_outputs],
[t.shape for t in new_outputs])
_copy_handle_data(new_outputs, op.outputs[orig_num_params:])
captured_inputs = _resolve_grad_captures(body_graph, body_grad_graph,
while_op)
loop_vars = args + captured_inputs
# This modifies body_grad_graph.
loop_vars = while_v2_indexed_slices_rewriter.rewrite_grad_indexed_slices(
grads, body_grad_graph, loop_vars, while_op.inputs)
def grad_cond(counter, unused_maximum_iterations_arg, forward_loop_iters,
*unused_args):
return counter < forward_loop_iters
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name,
grad_cond,
loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
_check_num_inputs_outputs(cond_grad_graph, body_grad_graph, len(loop_vars))
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
parallel_iterations=parallel_iterations,
name="%s_grad" % while_op.name)
grad_op = outputs[0].op
_copy_handle_data(body_grad_graph.outputs, outputs)
util.maybe_set_lowering_attr(grad_op)
util.maybe_propagate_compile_time_consts_in_xla(grad_op)
# See comment in while_loop.
outputs = [array_ops.identity(t) for t in outputs]
return _get_structured_grad_output(outputs, grads, body_grad_graph)
def while_loop(cond,
body,
loop_vars,
shape_invariants=None,
parallel_iterations=10,
maximum_iterations=None,
name=None,
return_same_structure=True):
"""Like tf.while_loop, except emits a single While op."""
# Keep the original loop_vars around to know which args were TensorArrays.
orig_loop_vars = loop_vars
# Cache its length since we use it at multiple places below.
len_orig_loop_vars = len(orig_loop_vars)
# Convert TensorArrays to their flow variables. These get converted back to
# TensorArrays before calling `cond` and `body`. See `wrapped_cond` and
# `wrapped_body` below.
loop_vars = list(_tensor_array_to_flow(orig_loop_vars))
loop_vars = nest.map_structure(
ops.internal_convert_to_tensor_or_indexed_slices, loop_vars)
if shape_invariants is not None:
nest.assert_same_structure(orig_loop_vars, shape_invariants)
else:
shape_invariants = nest.map_structure(lambda t: t.shape, loop_vars)
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = util.unique_fn_name(scope, "cond")
body_name = util.unique_fn_name(scope, "body")
maximum_iterations_loop_var = _build_maximum_iterations_loop_var(
maximum_iterations)
loop_counter = constant_op.constant(
0,
dtype=maximum_iterations_loop_var.dtype
if maximum_iterations is not None else None,
name="loop_counter")
# Add loop counter needed for computing gradients.
loop_vars = [loop_counter, maximum_iterations_loop_var] + loop_vars
shape_invariants = type(shape_invariants)(
[tensor_shape.scalar(),
tensor_shape.scalar()]) + shape_invariants
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = ops.get_default_graph(
)._add_control_dependencies
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(loop_counter, maximum_iterations_arg, *args):
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
if maximum_iterations is None:
return cond(*_pack_sequence_as(orig_loop_vars, args))
else:
return math_ops.logical_and(
loop_counter < maximum_iterations_arg,
cond(*_pack_sequence_as(orig_loop_vars, args)))
# NOTE(skyewm): we set collections to the outer graph's collections for
# compatibility with TPUEstimator.
cond_graph = func_graph_module.func_graph_from_py_func(
cond_name,
wrapped_cond,
[], # We provide signature instead of args.
{},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileCondFuncGraph(
cond_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies)
def wrapped_body(loop_counter, maximum_iterations_arg, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
maximum_iterations_arg: Maximum iterations of the loop.
*args: List of args
Returns:
A list of tensors the same length as args.
"""
# Capture the tensors already captured in cond_graph so that they appear
# in the same order in body_graph.external_captures.
for t in cond_graph.external_captures:
ops.get_default_graph().capture(t)
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
outputs = body(*_pack_sequence_as(orig_loop_vars, args))
if not nest.is_sequence(outputs):
outputs = [outputs]
# Compare the structure of input and output of body converting the
# top-level tuples to list to be compatible with legacy while_loop.
nest.assert_same_structure(list(outputs), list(orig_loop_vars))
outputs = _tensor_array_to_flow(outputs)
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1, maximum_iterations_arg] + list(outputs)
body_graph = func_graph_module.func_graph_from_py_func(
body_name,
wrapped_body,
[], # We provide signature instead of args.
{},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileBodyFuncGraph(
body_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies)
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
loop_vars = loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture the extra `external_captures` of `body_graph` in `cond_graph` so
# that it expects to receive those as arguments.
with cond_graph.as_default():
num_cond_captures = len(cond_graph.external_captures)
assert (cond_graph.external_captures ==
body_graph.external_captures[:num_cond_captures])
for body_capture in body_graph.external_captures[
num_cond_captures:]:
assert body_capture not in cond_graph.captures
cond_graph.capture(body_capture)
# Make sure that the shapes of the loop outputs are compatible with the
# shape invariants, or the shapes of the loop vars if the invariants are not
# specified.
num_flattened_outputs = len(nest.flatten(orig_loop_vars))
# First var is loop counter and second var is maximum_iterations.
first_loop_var_index = 2
_check_shapes_compat(
body_graph.outputs[first_loop_var_index:first_loop_var_index +
num_flattened_outputs],
nest.flatten(
shape_invariants[first_loop_var_index:first_loop_var_index +
len_orig_loop_vars]),
nest.flatten(loop_vars[first_loop_var_index:first_loop_var_index +
len_orig_loop_vars]))
flattened_loop_vars = nest.flatten(loop_vars)
_check_num_inputs_outputs(cond_graph, body_graph,
len(flattened_loop_vars))
with ops.control_dependencies(
list(cond_graph.control_captures) +
list(body_graph.control_captures)):
outputs = gen_functional_ops._while(
flattened_loop_vars,
util.create_new_tf_function(cond_graph),
util.create_new_tf_function(body_graph),
output_shapes=[t.shape for t in body_graph.outputs],
parallel_iterations=parallel_iterations,
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
util.maybe_set_lowering_attr(outputs[0].op)
util.maybe_propagate_compile_time_consts_in_xla(outputs[0].op)
# Return identities for each output of the While op, rather than the output
# of the While op directly. This makes pruning work if the output of
# while_loop() is fetched: the lowering pass converts the While outputs into
# IdentityN outputs, which if fetched will cause all ops in the body to be
# run (since it takes all exit ops as input). After lowering, each output
# identity op will end up with only the appropriate exit op as input.
outputs = tuple(array_ops.identity(t) for t in outputs)
outputs = _pack_sequence_as(
orig_loop_vars, outputs[first_loop_var_index:first_loop_var_index +
num_flattened_outputs])
if return_same_structure:
return outputs
flattened_outputs = nest.flatten(outputs)
if len(flattened_outputs) == 1:
return flattened_outputs[0]
else:
return outputs
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
body_graph = _get_body_graph(op)
# Set the incoming gradient of TensorArray handles to None. The gradient
# implementation currently assumes all resource tensors correspond to float32
# ResourceVariables, which can lead to runtime shape errors when used with a
# TensorArray. This is a workaround until TensorArrays are reimplemented with
# TensorLists instead of resources.
# Also set the incoming gradient of non-trainable inputs to None. It is
# possible that we receive non-None gradients for non-trainable types in
# nested while loops because we accumulate outputs of the inner while as
# variant tensors which are trainable and hence receive zeros_like tensors in
# the gradient pass. The non-trainable tensors then receive the popped zeros
# tensor from this zeros variant. The gradient for the loop vars corresponding
# to these tensors is None or zeros (this happens only if the loop var is
# accumulated as well) in _grad_fn so we reset these.
# TODO(b/118712257): Remove the IsTrainable filter once we can handle None
# output grads in _grad_fn.
grads = [
None if _is_tensor_array_handle(output)
or not gradients_impl.IsTrainable(output) else grad
for grad, output in zip(grads, op.outputs)
]
# Ensure that all non-resource trainable outputs have incoming gradients.
assert all(g is not None or o.dtype == dtypes.resource
or not gradients_impl.IsTrainable(o)
for o, g in zip(op.outputs, grads)
), "All trainable loop vars must receive incoming gradients."
# We compute the gradient for the sub-graph between trainable ys and xs
# with non-None incoming gradients. We later pad the None's to the list of
# outputs.
ys, xs, non_none_grads = zip(
*[(y, x, grad)
for (y, x, grad) in zip(body_graph.outputs, body_graph.inputs, grads)
if grad is not None])
body_grad_graph, args = _create_grad_func(
ys, xs, non_none_grads, body_graph,
util.unique_grad_fn_name(body_graph.name), op)
intermediate_tensors = _get_intermediates(body_grad_graph)
maximum_iterations = op.get_attr(
"_maximum_iterations") if _is_in_xla_context() else None
assert not _is_in_xla_context() or maximum_iterations is not None
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape),
max_num_elements=maximum_iterations)
with body_grad_graph.as_default():
tensor_list_ph = body_grad_graph.capture(tensor_list,
whitelisted=True)
# Push the intermediate tensor to the tensor list.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list_ph, intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_grad_graph.outputs.append(appended_tensor_list)
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
loop_vars = args + body_grad_graph.external_captures
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name,
grad_cond,
loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
_check_num_inputs_outputs(cond_grad_graph, body_grad_graph, len(loop_vars))
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
name="%s_grad" % op.name)
_copy_handle_data(body_grad_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
_maybe_set_maximum_iterations_attr(outputs[0].op, maximum_iterations)
# Set None as the output gradient for tensors with None input gradient
# e.g. TensorArray handles.
# outputs[0] is the loop counter.
# outputs[1] is the total number of loop iterations.
index = 2
none_padded_outputs = []
for g in grads:
if g is None:
none_padded_outputs.append(None)
else:
none_padded_outputs.append(outputs[index])
index += 1
return none_padded_outputs
def while_loop(cond,
body,
loop_vars,
shape_invariants=None,
maximum_iterations=None,
name=None):
"""Like tf.while_loop, except emits a single While op."""
maximum_iterations = _validate_and_convert_to_tensor(maximum_iterations)
# Keep the original loop_vars around to know which args were TensorArrays.
orig_loop_vars = loop_vars
# Cache its length since we use it at multiple places below.
len_orig_loop_vars = len(orig_loop_vars)
# Convert TensorArrays to their flow variables. These get converted back to
# TensorArrays before calling `cond` and `body`. See `wrapped_cond` and
# `wrapped_body` below.
loop_vars = list(_tensor_array_to_flow(orig_loop_vars))
loop_vars = nest.map_structure(
ops.internal_convert_to_tensor_or_indexed_slices, loop_vars)
if shape_invariants is not None:
nest.assert_same_structure(orig_loop_vars, shape_invariants)
else:
shape_invariants = nest.map_structure(lambda t: t.shape, loop_vars)
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = util.unique_fn_name(scope, "cond")
body_name = util.unique_fn_name(scope, "body")
loop_counter = constant_op.constant(
0,
dtype=maximum_iterations.dtype
if maximum_iterations is not None else None,
name="loop_counter")
# Add loop counter needed for computing gradients.
loop_vars = [loop_counter] + loop_vars
shape_invariants = [tensor_shape.scalar()] + shape_invariants
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(loop_counter, *args):
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
if maximum_iterations is None:
return cond(*_pack_sequence_as(orig_loop_vars, args))
else:
return math_ops.logical_and(
loop_counter < maximum_iterations,
cond(*_pack_sequence_as(orig_loop_vars, args)))
cond_graph = func_graph_module.func_graph_from_py_func(
cond_name,
wrapped_cond,
loop_vars, {},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileCondFuncGraph(cond_name),
add_control_dependencies=add_control_dependencies)
# Add external_captures of cond to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
loop_vars = loop_vars + cond_graph.external_captures
shape_invariants = shape_invariants + [
t.shape for t in cond_graph.external_captures
]
def wrapped_body(loop_counter, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
*args: List of args
args[:len_orig_loop_vars] - Args for the original loop body.
args[len_orig_loop_vars:] - External captures of cond. These get
passed through as is.
Returns:
A list of tensors the same length as args.
"""
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
outputs = body(
*_pack_sequence_as(orig_loop_vars, args[:len_orig_loop_vars]))
if not nest.is_sequence(outputs):
outputs = [outputs]
# Compare the structure of input and output of body converting the
# top-level tuples to list to be compatible with legacy while_loop.
nest.assert_same_structure(list(outputs), list(orig_loop_vars))
outputs = _tensor_array_to_flow(outputs)
# Return the external_captures of cond_graph as is, i.e., treat them as
# loop invariants.
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1] + list(outputs) + list(
args[len_orig_loop_vars:])
body_graph = func_graph_module.func_graph_from_py_func(
body_name,
wrapped_body,
loop_vars, {},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileBodyFuncGraph(body_name),
add_control_dependencies=add_control_dependencies)
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
loop_vars = loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture `external_captures` of `body_graph` in `cond_graph` so that it
# expects to receive those as arguments.
# TODO(b/118457764): Dedup tensors that are captured in both the cond and
# body. This logic already exists in cond_v2.
with cond_graph.as_default():
for external_capture in body_graph.external_captures:
assert external_capture not in cond_graph.captures, (
"Looks like both cond and body are capturing the same tensor %s. "
"This is not supported yet. For now consider passing,"
" this as a loop variable." % str(external_capture))
cond_graph.capture(external_capture)
# Export all tensors in the loop body that may be needed for gradient
# computation. We do this by accumulating the intermediate values in
# TensorLists.
intermediate_tensors = _get_intermediates(body_graph)
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape),
max_num_elements=maximum_iterations)
loop_vars.append(tensor_list)
with cond_graph.as_default():
# Add a placeholder to cond_graph's inputs corresponding to the
# tensor_list.
cond_graph.capture(tensor_list)
with body_graph.as_default():
# Push the intermediate tensor to the tensor list. This captures the
# `tensor_list` as well.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list, intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_graph.outputs.append(appended_tensor_list)
# Make sure that the shapes of the loop outputs are compatible with the
# shape invariants, or the shapes of the loop vars if the invariants are not
# specified.
num_flattened_outputs = len(nest.flatten(orig_loop_vars))
_check_shapes_compat(
body_graph.outputs[1:1 + num_flattened_outputs],
nest.flatten(shape_invariants[1:1 + len_orig_loop_vars]),
nest.flatten(loop_vars[1:1 + len_orig_loop_vars]))
flattened_loop_vars = nest.flatten(loop_vars)
_check_num_inputs_outputs(cond_graph, body_graph,
len(flattened_loop_vars))
outputs = gen_functional_ops._while(
flattened_loop_vars,
util.create_new_tf_function(cond_graph),
util.create_new_tf_function(body_graph),
output_shapes=[t.shape for t in body_graph.outputs],
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
_maybe_set_maximum_iterations_attr(outputs[0].op, maximum_iterations)
# Return identities for each output of the While op, rather than the output
# of the While op directly. This makes pruning work if the output of
# while_loop() is fetched: the lowering pass converts the While outputs into
# IdentityN outputs, which if fetched will cause all ops in the body to be
# run (since it takes all exit ops as input). After lowering, each output
# identity op will end up with only the appropriate exit op as input.
outputs = tuple(array_ops.identity(t) for t in outputs)
# First var is loop counter.
if num_flattened_outputs == 1:
return outputs[1]
else:
return _pack_sequence_as(orig_loop_vars,
outputs[1:1 + num_flattened_outputs])
def while_loop(cond, body, loop_vars, shape_invariants=None, name=None):
"""Like tf.while_loop, except emits a single While op."""
flattened_loop_vars = nest.flatten(loop_vars)
if shape_invariants is not None:
nest.assert_same_structure(loop_vars, shape_invariants)
flattened_shapes = nest.flatten(shape_invariants)
else:
flattened_shapes = [t.shape for t in flattened_loop_vars]
del shape_invariants
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = util.unique_fn_name(scope, "cond")
body_name = util.unique_fn_name(scope, "body")
num_outputs = len(flattened_loop_vars)
# Add loop counter needed for computing gradients.
flattened_loop_vars = [constant_op.constant(0., name="loop_counter")
] + flattened_loop_vars
flattened_shapes = [tensor_shape.scalar()] + flattened_shapes
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(unused_loop_counter, *loop_vars):
return cond(*loop_vars)
signature = [
tensor_spec.TensorSpec(shape, t.dtype)
for shape, t in zip(flattened_shapes, flattened_loop_vars)
]
cond_graph = function.func_graph_from_py_func(
cond_name,
wrapped_cond,
flattened_loop_vars, {},
signature=signature,
func_graph=util.WhileCondFuncGraph(cond_name),
add_control_dependencies=add_control_dependencies)
# Add external_captures of cond to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
flattened_loop_vars = flattened_loop_vars + cond_graph.external_captures
flattened_shapes = flattened_shapes + [
t.shape for t in cond_graph.external_captures
]
def wrapped_body(loop_counter, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
*args: List of args
args[:num_outputs] - Args for the original loop body.
args[num_outputs:] - External captures of cond. These get passed
through as is.
Returns:
A list of tensors the same length as args.
"""
outputs = body(*args[:num_outputs])
if not isinstance(outputs, collections.Sequence):
outputs = [outputs]
# Return the external_captures of cond_graph as is, i.e., treat them as
# loop invariants.
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1] + list(outputs) + list(
args[num_outputs:])
signature = [
tensor_spec.TensorSpec(shape, t.dtype)
for shape, t in zip(flattened_shapes, flattened_loop_vars)
]
body_graph = function.func_graph_from_py_func(
body_name,
wrapped_body,
flattened_loop_vars, {},
signature=signature,
func_graph=util.WhileBodyFuncGraph(body_name),
add_control_dependencies=add_control_dependencies)
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
flattened_loop_vars = flattened_loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture `external_captures` of `body_graph` in `cond_graph` so that it
# expects to receive those as arguments.
# TODO(srbs): Dedup tensors that are captured in both the cond and body.
# This logic already exists in cond_v2.
with cond_graph.as_default():
for external_capture in body_graph.external_captures:
cond_graph.capture(external_capture)
# Export all tensors in the loop body that may be needed for gradient
# computation. We do this by accumulating the intermediate values in
# TensorLists.
intermediate_tensors = _get_intermediates(body_graph)
for intermediate_tensor in intermediate_tensors:
# TODO(srbs): Cache and re-use empty tensor lists.
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape))
flattened_loop_vars.append(tensor_list)
with cond_graph.as_default():
# Add a placeholder to cond_graph's inputs corresponding to the
# tensor_list.
cond_graph.capture(tensor_list)
with body_graph.as_default():
# Push the intermediate tensor to the tensor list. This captures the
# `tensor_list` as well.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list, intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_graph.outputs.append(appended_tensor_list)
# Make sure that the shapes of the loop outputs are compatible with the
# shape invariants, or the shapes of the loop vars if the invariants are not
# specified.
_check_shapes_compat(body_graph.outputs[1:1 + num_outputs],
flattened_shapes[1:1 + num_outputs],
flattened_loop_vars[1:1 + num_outputs])
outputs = gen_functional_ops._while(
flattened_loop_vars,
util.create_new_tf_function(cond_graph),
util.create_new_tf_function(body_graph),
output_shapes=[t.shape for t in body_graph.outputs],
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
# Return identities for each output of the While op, rather than the output
# of the While op directly. This makes pruning work if the output of
# while_loop() is fetched: the lowering pass converts the While outputs into
# IdentityN outputs, which if fetched will cause all ops in the body to be
# run (since it takes all exit ops as input). After lowering, each output
# identity op will end up with only the appropriate exit op as input.
outputs = tuple(array_ops.identity(t) for t in outputs)
# First var is loop counter.
if num_outputs == 1:
return outputs[1]
else:
return nest.pack_sequence_as(loop_vars, outputs[1:1 + num_outputs])
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
body_graph = _get_body_graph(op)
# Replace None gradients with zeros. This is needed because `grads` could have
# None incoming gradients for the TensorLists. If we pass None's through, the
# custom gradient of TensorListPopBack will create an EmptyTensorList inside
# the FuncGraph which is undesirable.
# TODO(b/80444525): There might be an issue with treating no gradient as zero
# gradient in certain cases. Consider replacing None gradients with Zeros
# for accumulators only.
grads = [
g if g is not None else array_ops.zeros_like(output)
for g, output in zip(grads, op.outputs)
]
body_grad_graph, args = _create_grad_func(
body_graph, grads, util.unique_grad_fn_name(body_graph.name), op)
intermediate_tensors = _get_intermediates(body_grad_graph)
for intermediate_tensor in intermediate_tensors:
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape))
with body_grad_graph.as_default():
tensor_list_ph = body_grad_graph.capture(tensor_list,
whitelisted=True)
# Push the intermediate tensor to the tensor list.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list_ph, intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_grad_graph.outputs.append(appended_tensor_list)
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
loop_vars = args + body_grad_graph.external_captures
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = function.func_graph_from_py_func(
grad_cond_name,
grad_cond,
loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
assert len(loop_vars) == len(body_grad_graph.inputs)
assert len(loop_vars) == len(body_grad_graph.outputs)
assert len(loop_vars) == len(cond_grad_graph.inputs)
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
name="%s_grad" % op.name)
_copy_handle_data(body_grad_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
# outputs[0] is the loop counter.
# outputs[1] is the total number of loop iterations.
return outputs[2:2 + len(op.inputs)]
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
cond_graph = _get_graph(op, "cond")
body_graph = _get_graph(op, "body")
orig_num_params = len(body_graph.outputs)
maximum_iterations = op.get_attr(
"_maximum_iterations") if _is_in_xla_context() else None
assert not _is_in_xla_context() or maximum_iterations is not None
# Set the incoming gradient of non-trainable inputs to None. It is possible
# that we receive non-None gradients for non-trainable types in nested while
# loops because we accumulate outputs of the inner while as variant tensors
# which are trainable and hence receive zeros_like tensors in the gradient
# pass. The non-trainable tensors then receive the popped zeros tensor from
# this zeros variant. The gradient for the loop vars corresponding to these
# tensors is None or zeros (this happens only if the loop var is accumulated
# as well) in _grad_fn so we reset these.
# TODO(b/118712257): Remove the IsTrainable filter once we can handle None
# output grads in _grad_fn.
grads = [
None if not _is_trainable(output) else grad
for grad, output in zip(grads, body_graph.outputs)
]
# Ensure that all non-resource trainable outputs have incoming gradients.
assert all(g is not None or o.dtype == dtypes.resource or not _is_trainable(o)
for o, g in zip(body_graph.outputs, grads)
), "All trainable loop vars must receive incoming gradients."
# We compute the gradient for the sub-graph between trainable ys and xs
# with non-None incoming gradients. We later pad the None's to the list of
# outputs.
ys, xs, non_none_grads = zip(*[(y, x, grad) for (y, x, grad) in zip(
body_graph.outputs, body_graph.inputs, grads) if grad is not None])
body_grad_graph, args = _create_grad_func(
ys, xs, non_none_grads, cond_graph, body_graph,
util.unique_grad_fn_name(body_graph.name), op, maximum_iterations)
if body_grad_graph.while_op_needs_rewrite:
# Modify 'op' to output the intermediate accumulators needed by the grad
# function.
# NOTE(skyewm): if there are any active sessions, this modification to `op`
# may make them unrunnable!
cond_graph.name += "_rewritten"
body_graph.name += "_rewritten"
new_inputs = body_grad_graph.empty_tensor_lists
new_outputs = body_graph.outputs[orig_num_params:]
op._set_func_attr("cond", util.create_new_tf_function(cond_graph))
op._set_func_attr("body", util.create_new_tf_function(body_graph))
op._set_type_list_attr("T", body_graph.output_types)
op._set_shape_list_attr("output_shapes", body_graph.output_shapes)
op._add_while_inputs(new_inputs)
op._add_outputs([t.dtype for t in new_outputs],
[t.shape for t in new_outputs])
_copy_handle_data(new_outputs, op.outputs[orig_num_params:])
captured_inputs = _resolve_grad_captures(body_graph, body_grad_graph, op)
loop_vars = args + captured_inputs
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name, grad_cond, loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
_check_num_inputs_outputs(cond_grad_graph, body_grad_graph, len(loop_vars))
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
name="%s_grad" % op.name)
_copy_handle_data(body_grad_graph.outputs, outputs)
util.maybe_set_lowering_attr(outputs[0].op)
_maybe_set_maximum_iterations_attr(outputs[0].op, maximum_iterations)
# See comment in while_loop.
outputs = [array_ops.identity(t) for t in outputs]
# Set None as the output gradient for tensors with None input gradient.
# outputs[0] is the loop counter.
# outputs[1] is the total number of loop iterations.
index = 2
none_padded_outputs = []
for g in grads:
if g is None:
none_padded_outputs.append(None)
else:
none_padded_outputs.append(outputs[index])
index += 1
return none_padded_outputs
def _WhileGrad(op, *grads): # pylint: disable=invalid-name
"""The gradient of a While op produced by while_loop."""
# Note that op is not always the same as while_op because the gradient tape,
# for eager mode compatibility, forgets information about the proper op. Since
# the loop cannot run in eager mode, however, we can safely introspect into
# the graph here.
while_op = op.outputs[0].op
cond_graph = _get_graph(while_op, "cond")
body_graph = _get_graph(while_op, "body")
orig_num_params = len(body_graph.outputs)
maximum_iterations = op.get_attr(
"_maximum_iterations") if _is_in_xla_context() else None
parallel_iterations = op.get_attr("parallel_iterations")
assert not _is_in_xla_context() or maximum_iterations is not None
maximum_iterations = _validate_and_convert_to_tensor(maximum_iterations)
grads = [_preprocess_grad(grad, body_out, while_out)
for grad, body_out, while_out
in zip(grads, body_graph.outputs, while_op.outputs)]
# We compute the gradient for the sub-graph between trainable ys and xs
# with non-None incoming gradients. We later pad the None's to the list of
# outputs.
ys, xs, non_none_grads = zip(*[(y, x, grad) for (y, x, grad) in zip(
body_graph.outputs, body_graph.inputs, grads) if grad is not None])
body_grad_graph, args = _create_grad_func(
ys, xs, non_none_grads, cond_graph, body_graph,
util.unique_grad_fn_name(body_graph.name), op, maximum_iterations)
if body_grad_graph.while_op_needs_rewrite:
# Modify 'op' to output the intermediate accumulators needed by the grad
# function.
# NOTE(skyewm): if there are any active sessions, this modification to `op`
# may make them unrunnable!
cond_graph.name += "_rewritten"
body_graph.name += "_rewritten"
new_inputs = body_grad_graph.empty_tensor_lists
new_outputs = body_graph.outputs[orig_num_params:]
while_op._set_func_attr("cond", util.create_new_tf_function(cond_graph))
while_op._set_func_attr("body", util.create_new_tf_function(body_graph))
while_op._set_type_list_attr("T", body_graph.output_types)
while_op._set_shape_list_attr("output_shapes", body_graph.output_shapes)
while_op._add_while_inputs(new_inputs)
while_op._add_outputs([t.dtype for t in new_outputs],
[t.shape for t in new_outputs])
_copy_handle_data(new_outputs, op.outputs[orig_num_params:])
captured_inputs = _resolve_grad_captures(body_graph, body_grad_graph,
while_op)
loop_vars = args + captured_inputs
# This modifies body_grad_graph.
loop_vars = while_v2_indexed_slices_rewriter.rewrite_grad_indexed_slices(
grads, body_grad_graph, loop_vars, while_op.inputs)
def grad_cond(counter, max_iters, *unused_args):
return counter < max_iters
grad_cond_name = util.unique_grad_fn_name(op.get_attr("cond").name)
cond_grad_graph = func_graph_module.func_graph_from_py_func(
grad_cond_name, grad_cond, loop_vars, {},
func_graph=util.WhileCondFuncGraph(grad_cond_name))
_check_num_inputs_outputs(cond_grad_graph, body_grad_graph, len(loop_vars))
outputs = gen_functional_ops._while(
loop_vars,
util.create_new_tf_function(cond_grad_graph),
util.create_new_tf_function(body_grad_graph),
output_shapes=[t.shape for t in body_grad_graph.outputs],
parallel_iterations=parallel_iterations,
name="%s_grad" % while_op.name)
grad_op = outputs[0].op
_copy_handle_data(body_grad_graph.outputs, outputs)
util.maybe_set_lowering_attr(grad_op)
_maybe_set_maximum_iterations_attr(grad_op, maximum_iterations)
# See comment in while_loop.
outputs = [array_ops.identity(t) for t in outputs]
return _get_structured_grad_output(outputs, grads, body_grad_graph)
def While(input_, cond, body, name=None, hostmem=None):
r"""output = input; While (Cond(output)) { output = Body(output) }.
Args:
input_: A list of `Tensor` objects.
A list of input tensors whose types are T.
cond: . A function takes 'input' and returns a tensor. If the tensor is
a scalar of non-boolean, the scalar is converted to a boolean
according to the following rule: if the scalar is a numerical
value, non-zero means True and zero means False; if the scalar is
a string, non-empty means True and empty means False. If the
tensor is not a scalar, non-emptiness means True and False
otherwise.
body: . A function takes a list of tensors and returns another
list tensors. Both lists have the same types as specified
by T.
name: A name for the operation (optional).
hostmem: A list of integer. If i is in the list, input[i] is a
host memory tensor.
Raises:
ValueError: if `cond` has implicitly captured inputs or if `cond` and `body`
have different signatures.
Returns:
A list of `Tensor` objects. Has the same type as `input`.
A list of output tensors whose types are T.
"""
if cond.captured_inputs:
raise ValueError("While op 'cond' argument must be a function "
"without implicitly captured inputs.")
if cond.declared_input_types != body.declared_input_types:
raise ValueError(
"While op 'cond' and 'body' signatures do not match. %r vs %r" %
(cond.declared_input_types, body.declared_input_types))
if body.captured_inputs:
cond_dtypes = list(body.declared_input_types) + [
t.dtype for t in body.captured_inputs
]
@function.Defun(*cond_dtypes, func_name="%s_Wrapper" % cond.name)
def CondWrapper(*args):
"""A wrapper that handles loop-carried captured inputs."""
return cond(*args[:len(body.declared_input_types)])
ret = gen_functional_ops._while(input_ + body.captured_inputs,
CondWrapper,
_LoopBodyCaptureWrapper(body),
name=name)
# Slice off the loop-carried captured inputs.
ret = ret[:-len(body.captured_inputs)]
else:
ret = gen_functional_ops._while(input_, cond, body, name=name)
if hostmem:
input_attr = attr_value_pb2.AttrValue()
input_attr.list.i.extend(hostmem)
ret[0].op._set_attr("_input_hostmem", input_attr) # pylint: disable=protected-access
output_attr = attr_value_pb2.AttrValue()
output_attr.list.i.extend(hostmem)
ret[0].op._set_attr("_output_hostmem", output_attr) # pylint: disable=protected-access
return ret
def while_loop(cond,
body,
loop_vars,
shape_invariants=None,
parallel_iterations=10,
maximum_iterations=None,
name=None,
return_same_structure=True):
"""Like tf.while_loop, except emits a single While op."""
maximum_iterations = _validate_and_convert_to_tensor(maximum_iterations)
# Keep the original loop_vars around to know which args were TensorArrays.
orig_loop_vars = loop_vars
# Cache its length since we use it at multiple places below.
len_orig_loop_vars = len(orig_loop_vars)
# Convert TensorArrays to their flow variables. These get converted back to
# TensorArrays before calling `cond` and `body`. See `wrapped_cond` and
# `wrapped_body` below.
loop_vars = list(_tensor_array_to_flow(orig_loop_vars))
loop_vars = nest.map_structure(
ops.internal_convert_to_tensor_or_indexed_slices, loop_vars)
if shape_invariants is not None:
nest.assert_same_structure(orig_loop_vars, shape_invariants)
else:
shape_invariants = nest.map_structure(lambda t: t.shape, loop_vars)
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = util.unique_fn_name(scope, "cond")
body_name = util.unique_fn_name(scope, "body")
loop_counter = constant_op.constant(
0,
dtype=maximum_iterations.dtype
if maximum_iterations is not None else None,
name="loop_counter")
# Add loop counter needed for computing gradients.
loop_vars = [loop_counter] + loop_vars
shape_invariants = type(shape_invariants)([tensor_shape.scalar()
]) + shape_invariants
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = ops.get_default_graph()._add_control_dependencies
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(loop_counter, *args):
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
if maximum_iterations is None:
return cond(*_pack_sequence_as(orig_loop_vars, args))
else:
return math_ops.logical_and(
loop_counter < maximum_iterations,
cond(*_pack_sequence_as(orig_loop_vars, args)))
# NOTE(skyewm): we set collections to the outer graph's collections for
# compatibility with TPUEstimator.
cond_graph = func_graph_module.func_graph_from_py_func(
cond_name,
wrapped_cond,
[], # We provide signature instead of args.
{},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileCondFuncGraph(
cond_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies)
def wrapped_body(loop_counter, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
*args: List of args
Returns:
A list of tensors the same length as args.
"""
# Capture the tensors already captured in cond_graph so that they appear
# in the same order in body_graph.external_captures.
for t in cond_graph.external_captures:
ops.get_default_graph().capture(t)
# Convert the flow variables in `args` to TensorArrays. `args` should
# already have the same structure as `orig_loop_vars` but currently there
# is no nest.zip so we call `_pack_sequence_as` which flattens both
# `orig_loop_vars` and `args`, converts flows in `args` to TensorArrays
# and packs it into the structure of `orig_loop_vars`.
outputs = body(*_pack_sequence_as(orig_loop_vars, args))
if not nest.is_sequence(outputs):
outputs = [outputs]
# Compare the structure of input and output of body converting the
# top-level tuples to list to be compatible with legacy while_loop.
nest.assert_same_structure(list(outputs), list(orig_loop_vars))
outputs = _tensor_array_to_flow(outputs)
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1] + list(outputs)
body_graph = func_graph_module.func_graph_from_py_func(
body_name,
wrapped_body,
[], # We provide signature instead of args.
{},
signature=_build_signature(loop_vars, shape_invariants),
func_graph=util.WhileBodyFuncGraph(
body_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies)
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
loop_vars = loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture the extra `external_captures` of `body_graph` in `cond_graph` so
# that it expects to receive those as arguments.
with cond_graph.as_default():
num_cond_captures = len(cond_graph.external_captures)
assert (cond_graph.external_captures ==
body_graph.external_captures[:num_cond_captures])
for body_capture in body_graph.external_captures[num_cond_captures:]:
assert body_capture not in cond_graph.captures
cond_graph.capture(body_capture)
# Make sure that the shapes of the loop outputs are compatible with the
# shape invariants, or the shapes of the loop vars if the invariants are not
# specified.
num_flattened_outputs = len(nest.flatten(orig_loop_vars))
_check_shapes_compat(
body_graph.outputs[1:1 + num_flattened_outputs],
nest.flatten(shape_invariants[1:1 + len_orig_loop_vars]),
nest.flatten(loop_vars[1:1 + len_orig_loop_vars]))
flattened_loop_vars = nest.flatten(loop_vars)
_check_num_inputs_outputs(cond_graph, body_graph,
len(flattened_loop_vars))
outputs = gen_functional_ops._while(
flattened_loop_vars,
util.create_new_tf_function(cond_graph),
util.create_new_tf_function(body_graph),
output_shapes=[t.shape for t in body_graph.outputs],
parallel_iterations=parallel_iterations,
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
util.maybe_set_lowering_attr(outputs[0].op)
_maybe_set_maximum_iterations_attr(outputs[0].op, maximum_iterations)
# Return identities for each output of the While op, rather than the output
# of the While op directly. This makes pruning work if the output of
# while_loop() is fetched: the lowering pass converts the While outputs into
# IdentityN outputs, which if fetched will cause all ops in the body to be
# run (since it takes all exit ops as input). After lowering, each output
# identity op will end up with only the appropriate exit op as input.
outputs = tuple(array_ops.identity(t) for t in outputs)
# First var is loop counter.
outputs = _pack_sequence_as(orig_loop_vars,
outputs[1:1 + num_flattened_outputs])
if return_same_structure:
return outputs
flattened_outputs = nest.flatten(outputs)
if len(flattened_outputs) == 1:
return flattened_outputs[0]
else:
return outputs
def while_loop(cond, body, loop_vars, name=None):
"""Like tf.while_loop, except emits a single While op."""
if not name:
name = "while"
with ops.name_scope(name) as scope:
with ops.name_scope(None):
cond_name = _get_unique_name(("%scond" % scope).replace("/", "_"))
body_name = _get_unique_name(("%sbody" % scope).replace("/", "_"))
flattened_loop_vars = nest.flatten(loop_vars)
num_outputs = len(flattened_loop_vars)
# Add loop counter needed for computing gradients.
flattened_loop_vars = [constant_op.constant(0., name="loop_counter")
] + flattened_loop_vars
# Build a `cond` wrapper that can handle the extra counter loop_var.
def wrapped_cond(unused_loop_counter, *loop_vars):
return cond(*loop_vars)
cond_graph = function.func_graph_from_py_func(cond_name, wrapped_cond,
flattened_loop_vars, {})
# Add external_captures of cond to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
flattened_loop_vars = flattened_loop_vars + cond_graph.external_captures
def wrapped_body(loop_counter, *args):
"""Loop body augmented with counter update.
Args:
loop_counter: Loop counter which needs to be incremented in the body.
*args: List of args
args[:num_outputs] - Args for the original loop body.
args[num_outputs:] - External captures of cond. These get passed
through as is.
Returns:
A list of tensors the same length as args.
"""
outputs = body(*args[:num_outputs])
if not isinstance(outputs, collections.Sequence):
outputs = [outputs]
# Return the external_captures of cond_graph as is, i.e., treat them as
# loop invariants.
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
return [loop_counter + 1] + list(outputs) + list(
args[num_outputs:])
body_graph = function.func_graph_from_py_func(body_name, wrapped_body,
flattened_loop_vars, {})
# Add external captures of body to the list of loop vars.
# Note that external tensors will be treated as loop invariants, i.e.,
# the value of that tensor in each iteration is the same as it was at the
# beginning of the loop execution.
flattened_loop_vars = flattened_loop_vars + body_graph.external_captures
# TODO(srbs): Update lowering code to create _Enter nodes with
# is_constant=True for inputs that are directly passed to outputs.
body_graph.outputs.extend(body_graph.internal_captures)
# Capture `external_captures` of `body_graph` in `cond_graph` so that it
# expects to receive those as arguments.
# TODO(srbs): Dedup tensors that are captured in both the cond and body.
# This logic already exists in cond_v2.
with cond_graph.as_default():
for external_capture in body_graph.external_captures:
cond_graph.capture(external_capture)
# Export all tensors in the loop body that may be needed for gradient
# computation. We do this by accumulating the intermediate values in
# TensorLists.
intermediate_tensors = _get_intermediates(body_graph)
for intermediate_tensor in intermediate_tensors:
# TODO(srbs): Cache and re-use empty tensor lists.
tensor_list = list_ops.empty_tensor_list(
element_dtype=intermediate_tensor.dtype,
element_shape=_get_tensor_convertible_shape(
intermediate_tensor.shape))
flattened_loop_vars.append(tensor_list)
with cond_graph.as_default():
# Add a placeholder to cond_graph's inputs corresponding to the
# tensor_list.
cond_graph.capture(tensor_list)
with body_graph.as_default():
# Push the intermediate tensor to the tensor list. This captures the
# `tensor_list` as well.
appended_tensor_list = list_ops.tensor_list_push_back(
tensor_list, intermediate_tensor)
# Add this modified tensor list to the list of outputs.
body_graph.outputs.append(appended_tensor_list)
outputs = gen_functional_ops._while(
flattened_loop_vars,
cond_v2._create_new_tf_function(cond_graph),
cond_v2._create_new_tf_function(body_graph),
name=scope)
_copy_handle_data(body_graph.outputs, outputs)
_maybe_set_lowering_attr(outputs[0].op)
# First var is loop counter.
if num_outputs == 1:
return outputs[1]
else:
return nest.pack_sequence_as(loop_vars, outputs[1:1 + num_outputs])
