def cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# 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
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn,
[],
{},
func_graph=util.CondBranchFuncGraph(
true_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn,
[],
{},
func_graph=util.CondBranchFuncGraph(
false_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
return _build_cond(pred,
true_graph,
false_graph,
true_graph.external_captures,
false_graph.external_captures,
name=scope)
def cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn, [], {},
func_graph=util.CondBranchFuncGraph(true_name,
read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn, [], {},
func_graph=util.CondBranchFuncGraph(false_name,
read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
outputs = _build_cond(pred,
true_graph,
false_graph,
true_graph.external_captures,
false_graph.external_captures,
name=scope)
return func_graph_module.pack_sequence_as(
true_graph.structured_outputs, outputs)
def cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn, [], {},
func_graph=util.CondBranchFuncGraph(
true_name, read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn, [], {},
func_graph=util.CondBranchFuncGraph(
false_name, read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
outputs = _build_cond(pred, true_graph, false_graph,
true_graph.external_captures,
false_graph.external_captures,
name=scope)
return func_graph_module.pack_sequence_as(true_graph.structured_outputs,
outputs)
def cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# 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
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn, [], {},
func_graph=util.CondBranchFuncGraph(
true_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn, [], {},
func_graph=util.CondBranchFuncGraph(
false_name, collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
verify_captures(true_graph, false_graph)
return _build_cond(pred, true_graph, false_graph,
true_graph.external_captures,
false_graph.external_captures,
name=scope)
def indexed_case(branch_index,
branch_fns,
name="indexed_case",
lower_using_switch_merge=None):
"""Like conv_v2, except emits a Case op instead of an If."""
if isinstance(branch_index, int):
raise TypeError("branch_index must not be a Python int", branch_index)
with ops.name_scope(name) as scope:
branch_names = [
util.unique_fn_name(scope, "branch{}".format(b))
for b in range(len(branch_fns))
]
# 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
branch_index = ops.convert_to_tensor(branch_index, name="branch_index")
branch_graphs = []
for branch_name, branch_fn in zip(branch_names, branch_fns):
branch_graphs.append(
func_graph_module.func_graph_from_py_func(
branch_name,
branch_fn,
[],
{},
func_graph=util.CondBranchFuncGraph(
branch_name,
collections=ops.get_default_graph()._collections), # pylint: disable=protected-access
add_control_dependencies=add_control_dependencies,
op_return_value=branch_index))
verify_captures(_CASE, branch_graphs)
return _build_case(
branch_index,
branch_graphs, [g.external_captures for g in branch_graphs],
name=scope,
lower_using_switch_merge=lower_using_switch_merge)
def while_loop(cond,
body,
loop_vars,
shape_invariants=None,
parallel_iterations=10,
maximum_iterations=None,
name=None,
return_same_structure=True,
back_prop=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,
expand_composites=True)
if shape_invariants is not None:
nest.assert_same_structure(orig_loop_vars, shape_invariants,
expand_composites=False)
signature = nest.map_structure(
control_flow_ops._shape_invariant_to_type_spec, loop_vars,
list(shape_invariants), expand_composites=False)
shape_invariants = nest.map_structure(
control_flow_ops._get_shape_invariant, loop_vars,
list(shape_invariants), expand_composites=False)
else:
signature = nest.map_structure(
type_spec.type_spec_from_value, loop_vars, expand_composites=False)
shape_invariants = nest.map_structure(
control_flow_ops._get_shape_invariant, loop_vars,
expand_composites=False)
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 = [tensor_shape.TensorShape([])] * 2 + shape_invariants
signature = (
[tensor_spec.TensorSpec.from_tensor(loop_counter),
tensor_spec.TensorSpec.from_tensor(maximum_iterations_loop_var)] +
signature)
# 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
def wrapped_cond(loop_counter, maximum_iterations_arg, *args):
"""Extra `cond` wrapper that can handle the extra counter loop_var."""
# 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`.
pred = cond(*_pack_sequence_as(orig_loop_vars, args))
if (tensor_util.is_tensor(pred) and
(pred.shape.dims is None or pred.shape.dims)):
pred = array_ops.squeeze_v2(pred)
if maximum_iterations is None:
return pred
else:
return math_ops.logical_and(
loop_counter < maximum_iterations_arg, pred)
# 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=signature,
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_or_composite(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),
expand_composites=True)
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=signature,
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])
cond_graph_captures = object_identity.ObjectIdentitySet(
cond_graph.external_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,
expand_composites=True))
# 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], expand_composites=True),
nest.flatten(loop_vars[first_loop_var_index:first_loop_var_index +
len_orig_loop_vars], expand_composites=True))
num_original_outputs = len(body_graph.outputs)
if back_prop and util.output_all_intermediates():
# 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)
flattened_loop_vars = nest.flatten(loop_vars, expand_composites=True)
_check_num_inputs_outputs(cond_graph, body_graph,
len(flattened_loop_vars))
_check_inputs_outputs_types_match(body_graph, flattened_loop_vars)
with ops.control_dependencies(
list(cond_graph.control_captures) + list(body_graph.control_captures)):
output_shapes = [t.shape for t in body_graph.outputs]
orig_loop_vars_range = slice(first_loop_var_index,
first_loop_var_index + num_flattened_outputs)
output_shapes[orig_loop_vars_range] = nest.flatten(
shape_invariants, expand_composites=True)[orig_loop_vars_range]
cond_stateful_ops = [
op for op in cond_graph.get_operations() if op._is_stateful
]
body_stateful_ops = [
op for op in body_graph.get_operations() if op._is_stateful
]
if (cond_stateful_ops or body_stateful_ops):
op_fn = gen_functional_ops._while
else:
op_fn = gen_functional_ops.stateless_while
outputs = op_fn(
flattened_loop_vars,
util.create_new_tf_function(cond_graph),
util.create_new_tf_function(body_graph),
output_shapes=output_shapes,
parallel_iterations=parallel_iterations,
name=scope)
# This is needed so we do not compute derivative wrt these extra outputs.
outputs[0].op._set_attr("_num_original_outputs",
attr_value_pb2.AttrValue(i=num_original_outputs))
_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, expand_composites=True)
if len(flattened_outputs) == 1:
return flattened_outputs[0]
else:
return outputs
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,
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 = 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)))
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)
# 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 cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn, [], {},
func_graph=util.CondBranchFuncGraph(
true_name, read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn, [], {},
func_graph=util.CondBranchFuncGraph(
false_name, read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
_check_same_outputs(true_graph, false_graph)
# Add inputs to true_graph and false_graph to make them match. Note that
# this modifies true_graph and false_graph.
cond_inputs = _make_inputs_match(true_graph, false_graph,
true_graph.external_captures,
false_graph.external_captures)
# Add all intermediate tensors as function outputs so they're available for
# the gradient computation.
true_intermediates = _get_intermediates(true_graph)
false_intermediates = _get_intermediates(false_graph)
# Save the original number of outputs to return to the caller.
num_cond_outputs = len(true_graph.outputs)
# Make the number/type of new intermediate outputs match.
extra_true_outputs, extra_false_outputs = _pad_params(
true_graph, false_graph, true_intermediates, false_intermediates)
true_graph.outputs.extend(extra_true_outputs)
false_graph.outputs.extend(extra_false_outputs)
# Create the If op.
tensors = gen_functional_ops._if( # pylint: disable=protected-access
pred,
cond_inputs, [t.dtype for t in true_graph.outputs],
util.create_new_tf_function(true_graph),
util.create_new_tf_function(false_graph),
output_shapes=_get_output_shapes(true_graph.outputs,
false_graph.outputs),
name=scope)
# TODO(b/110167197) this approach requires cond_v2 to have at least 1 output
util.maybe_set_lowering_attr(tensors[0].op)
# Return identities for each output of the If op, rather than the output of
# the If op directly. This makes pruning work if the output of cond() is
# fetched: the lowering pass converts the If outputs into IdentityN outputs,
# which if fetched will cause all ops in the taken branch to be run (since
# it takes all merge ops as input). After lowering, each output identity op
# will end up with only the appropriate merge op as input.
# TODO(b/79984175): this doesn't have to be a tuple once we covert to the
# correct output structure
tensors = tuple(array_ops.identity(t) for t in tensors)
return func_graph_module.pack_sequence_as(true_graph.structured_outputs,
tensors[:num_cond_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
# 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 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 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)
# 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,
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 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 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 = func_graph_module.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 = func_graph_module.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 cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn, [], {},
func_graph=util.CondBranchFuncGraph(true_name,
read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn, [], {},
func_graph=util.CondBranchFuncGraph(false_name,
read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
_check_same_outputs(true_graph, false_graph)
# Add inputs to true_graph and false_graph to make them match. Note that
# this modifies true_graph and false_graph.
cond_inputs = _make_inputs_match(true_graph, false_graph,
true_graph.external_captures,
false_graph.external_captures)
# Add all intermediate tensors as function outputs so they're available for
# the gradient computation.
true_intermediates = _get_intermediates(true_graph)
false_intermediates = _get_intermediates(false_graph)
# Save the original number of outputs to return to the caller.
num_cond_outputs = len(true_graph.outputs)
# Make the number/type of new intermediate outputs match.
extra_true_outputs, extra_false_outputs = _pad_params(
true_graph, false_graph, true_intermediates, false_intermediates)
true_graph.outputs.extend(extra_true_outputs)
false_graph.outputs.extend(extra_false_outputs)
# Create the If op.
tensors = gen_functional_ops._if( # pylint: disable=protected-access
pred,
cond_inputs, [t.dtype for t in true_graph.outputs],
util.create_new_tf_function(true_graph),
util.create_new_tf_function(false_graph),
output_shapes=_get_output_shapes(true_graph.outputs,
false_graph.outputs),
name=scope)
# Set the flag to enable lowering on the `if` op if necessary
# Lowering allows cond_v2 to avoid some of the limitations of Functions,
# allowing users to specify devices & colocation inside of cond_v2 branches,
# and enabling non-strict evaluation & partial pruning of cond_v2 branches.
# This brings cond_v2 closer to feature parity with tf.cond.
#
# However, we do not lower `If` in the XLA context because it is easier for
# XLA to apply its own optimizations when dealing with un-lowered `If`
# operators than with lowered switch/merge control flow.
#
# TODO(b/110167197) this approach requires cond_v2 to have at least 1 output
if_op = tensors[0].op
if not control_flow_util.IsInXLAContext(if_op):
# pylint: disable=protected-access
if_op._set_attr("_lower_using_switch_merge",
attr_value_pb2.AttrValue(b=True))
# pylint: enable=protected-access
# Return identities for each output of the If op, rather than the output of
# the If op directly. This makes pruning work if the output of cond() is
# fetched: the lowering pass converts the If outputs into IdentityN outputs,
# which if fetched will cause all ops in the taken branch to be run (since
# it takes all merge ops as input). After lowering, each output identity op
# will end up with only the appropriate merge op as input.
# TODO(b/79984175): this doesn't have to be a tuple once we covert to the
# correct output structure
tensors = tuple(array_ops.identity(t) for t in tensors)
result = tuple(tensors[:num_cond_outputs])
if len(result) == 1:
return result[0]
else:
return result
def cond_v2(pred, true_fn, false_fn, name="cond"):
"""Like tf.cond, except emits a single If op."""
if isinstance(pred, bool):
raise TypeError("pred must not be a Python bool", pred)
if not name:
name = "cond"
with ops.name_scope(name) as scope:
true_name = util.unique_fn_name(scope, "true")
false_name = util.unique_fn_name(scope, "false")
# Automatic control dependencies are added in defuns, but not in v1
# graphs. Propagate that behavior here.
add_control_dependencies = util.in_defun()
pred = ops.convert_to_tensor(pred)
true_graph = func_graph_module.func_graph_from_py_func(
true_name,
true_fn, [], {},
func_graph=util.CondBranchFuncGraph(
true_name, read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
false_graph = func_graph_module.func_graph_from_py_func(
false_name,
false_fn, [], {},
func_graph=util.CondBranchFuncGraph(
false_name, read_only_collections=False),
add_control_dependencies=add_control_dependencies,
op_return_value=pred)
_check_same_outputs(true_graph, false_graph)
# Add inputs to true_graph and false_graph to make them match. Note that
# this modifies true_graph and false_graph.
cond_inputs = _make_inputs_match(true_graph, false_graph,
true_graph.external_captures,
false_graph.external_captures)
# Add all intermediate tensors as function outputs so they're available for
# the gradient computation.
true_intermediates = _get_intermediates(true_graph)
false_intermediates = _get_intermediates(false_graph)
# Save the original number of outputs to return to the caller.
num_cond_outputs = len(true_graph.outputs)
# Make the number/type of new intermediate outputs match.
extra_true_outputs, extra_false_outputs = _pad_params(
true_graph, false_graph, true_intermediates, false_intermediates)
true_graph.outputs.extend(extra_true_outputs)
false_graph.outputs.extend(extra_false_outputs)
# Create the If op.
tensors = gen_functional_ops._if( # pylint: disable=protected-access
pred,
cond_inputs, [t.dtype for t in true_graph.outputs],
util.create_new_tf_function(true_graph),
util.create_new_tf_function(false_graph),
output_shapes=_get_output_shapes(true_graph.outputs,
false_graph.outputs),
name=scope)
# Set the flag to enable lowering on the `if` op if necessary
# Lowering allows cond_v2 to avoid some of the limitations of Functions,
# allowing users to specify devices & colocation inside of cond_v2 branches,
# and enabling non-strict evaluation & partial pruning of cond_v2 branches.
# This brings cond_v2 closer to feature parity with tf.cond.
#
# However, we do not lower `If` in the XLA context because it is easier for
# XLA to apply its own optimizations when dealing with un-lowered `If`
# operators than with lowered switch/merge control flow.
#
# TODO(b/110167197) this approach requires cond_v2 to have at least 1 output
if_op = tensors[0].op
if not control_flow_util.IsInXLAContext(if_op):
# pylint: disable=protected-access
if_op._set_attr("_lower_using_switch_merge",
attr_value_pb2.AttrValue(b=True))
# pylint: enable=protected-access
# Return identities for each output of the If op, rather than the output of
# the If op directly. This makes pruning work if the output of cond() is
# fetched: the lowering pass converts the If outputs into IdentityN outputs,
# which if fetched will cause all ops in the taken branch to be run (since
# it takes all merge ops as input). After lowering, each output identity op
# will end up with only the appropriate merge op as input.
# TODO(b/79984175): this doesn't have to be a tuple once we covert to the
# correct output structure
tensors = tuple(array_ops.identity(t) for t in tensors)
result = tuple(tensors[:num_cond_outputs])
if len(result) == 1:
return result[0]
else:
return result
