def _grad_fn(func_graph, grads):
"""The gradient function for each conditional branch.
This function builds the gradient graph of the corresponding forward-pass
conditional branch in `func_graph`. This is done by differentiating
func_graph's outputs w.r.t. its inputs.
Args:
func_graph: function.FuncGraph. The corresponding forward-pass function.
grads: The list of input gradient Tensors.
Returns:
The output gradient Tensors.
"""
# Filter out untrainable function outputs.
# NOTE(skyewm): If we don't do this, the untrainable tensors can sometimes
# cause _GradientsHelper to raise an exception (e.g. the implementation
# doesn't expect 'ys' to contain boolean tensors).
assert len(func_graph.outputs) == len(grads)
ys = []
grad_ys = []
for y, grad_y in zip(func_graph.outputs, grads):
if not gradients_impl._IsTrainable(y):
continue
ys.append(y)
grad_ys.append(grad_y)
# Build the gradient graph. Note that this builds the gradient computation of
# func_graph in the current graph, which requires capturing tensors from
# func_graph. The captured func_graph tensors are resolved to external tensors
# in _resolve_grad_inputs.
result = gradients_impl._GradientsHelper(ys,
func_graph.inputs,
grad_ys=grad_ys,
src_graph=func_graph)
# Functions can't return None; replace Nones with zero tensors.
# TODO(b/80444525): don't return anything here and make _IfGrad return None if
# both branches have zero gradient.
for i in range(len(result)):
if result[i] is None:
result[i] = array_ops.zeros_like(func_graph.inputs[i])
return result
def _grad_fn(func_graph, grads):
"""The gradient function for each conditional branch.
This function builds the gradient graph of the corresponding forward-pass
conditional branch in `func_graph`. This is done by differentiating
func_graph's outputs w.r.t. its inputs.
Args:
func_graph: function.FuncGraph. The corresponding forward-pass function.
grads: The list of input gradient Tensors.
Returns:
The output gradient Tensors.
"""
# Filter out untrainable function outputs.
# NOTE(skyewm): If we don't do this, the untrainable tensors can sometimes
# cause _GradientsHelper to raise an exception (e.g. the implementation
# doesn't expect 'ys' to contain boolean tensors).
assert len(func_graph.outputs) == len(grads)
ys = []
grad_ys = []
for y, grad_y in zip(func_graph.outputs, grads):
if not gradients_impl._IsTrainable(y):
continue
ys.append(y)
grad_ys.append(grad_y)
# Build the gradient graph. Note that this builds the gradient computation of
# func_graph in the current graph, which requires capturing tensors from
# func_graph. The captured func_graph tensors are resolved to external tensors
# in _resolve_grad_inputs.
result = gradients_impl._GradientsHelper(
ys, func_graph.inputs, grad_ys=grad_ys,
src_graph=func_graph)
# Functions can't return None; replace Nones with zero tensors.
# TODO(b/80444525): don't return anything here and make _IfGrad return None if
# both branches have zero gradient.
for i in range(len(result)):
if result[i] is None:
result[i] = array_ops.zeros_like(func_graph.inputs[i])
return result
def _GradientsHelper(ys, xs, grad_ys, name, colocate_gradients_with_ops,
gate_gradients, aggregation_method, stop_gradients):
"""Implementation of gradients()."""
if context.executing_eagerly():
raise RuntimeError("tf.gradients not supported when eager execution "
"is enabled. Use tf.contrib.eager.GradientTape "
"instead.")
ys = _AsList(ys)
xs = _AsList(xs)
stop_gradients = [] if stop_gradients is None else _AsList(stop_gradients)
if grad_ys is None:
grad_ys = [None] * len(ys)
else:
grad_ys = _AsList(grad_ys)
with ops.name_scope(
name, "gradients",
list(ys) + list(xs) + list(stop_gradients) +
list(grad_ys)) as grad_scope:
ys = ops.convert_n_to_tensor_or_indexed_slices(ys, name="y")
xs = [
x.handle if resource_variable_ops.is_resource_variable(x) else x
for x in xs
]
xs = ops.internal_convert_n_to_tensor_or_indexed_slices(xs,
name="x",
as_ref=True)
grad_ys = _DefaultGradYs(grad_ys, ys, colocate_gradients_with_ops)
# The approach we take here is as follows: Create a list of all ops in the
# subgraph between the ys and xs. Visit these ops in reverse order of ids
# to ensure that when we visit an op the gradients w.r.t its outputs have
# been collected. Then aggregate these gradients if needed, call the op's
# gradient function, and add the generated gradients to the gradients for
# its input.
# Initialize the pending count for ops in the connected subgraph from ys
# to the xs.
if len(ys) > 1:
ys = [array_ops.identity(y) if y.consumers() else y for y in ys]
to_ops = [t.op for t in ys]
from_ops = [t.op for t in xs]
stop_gradient_ops = [t.op for t in stop_gradients]
pending_count, loop_state = _PendingCount(ys[0].graph, to_ops,
from_ops,
colocate_gradients_with_ops)
# Iterate over the collected ops.
#
# grads: op => list of gradients received on each output endpoint of the
# op. The gradients for each endpoint are initially collected as a list.
# When it is time to call the op's gradient function, for each endpoint we
# aggregate the list of received gradients into a Add() Operation if there
# is more than one.
grads = {}
# Add the initial gradients for the ys.
for y, grad_y in zip(ys, grad_ys):
_SetGrad(grads, y, grad_y)
# Initialize queue with to_ops.
queue = collections.deque()
# Add the ops in 'to_ops' into the queue.
to_ops_set = set()
for op in to_ops:
# 'ready' handles the case where one output gradient relies on
# another output's gradient.
# pylint: disable=protected-access
ready = (pending_count[op._id] == 0)
if ready and op._id not in to_ops_set:
to_ops_set.add(op._id)
queue.append(op)
# pylint: enable=protected-access
if loop_state:
loop_exits = loop_state.ProcessUnusedLoopExits(
pending_count, to_ops_set)
for y in loop_exits:
if _IsTrainable(y):
_SetGrad(grads, y, loop_state.ZerosLikeForExit(y))
queue.append(y.op)
stop_ops = _StopOps(from_ops, stop_gradient_ops, pending_count)
while queue:
# generate gradient subgraph for op.
op = queue.popleft()
with _maybe_colocate_with(op, colocate_gradients_with_ops):
if loop_state:
loop_state.EnterGradWhileContext(op, before=True)
out_grads = _AggregatedGrads(grads, op, loop_state,
aggregation_method)
if loop_state:
loop_state.ExitGradWhileContext(op, before=True)
grad_fn = None
# pylint: disable=protected-access
func_call = None
is_func_call = ops.get_default_graph()._is_function(op.type)
has_out_grads = any(
isinstance(g, ops.Tensor) or g for g in out_grads)
if has_out_grads and (op._id not in stop_ops):
if is_func_call:
func_call = ops.get_default_graph()._get_function(
op.type)
grad_fn = func_call.python_grad_func
# pylint: enable=protected-access
else:
# A grad_fn must be defined, either as a function or as None
# for ops that do not have gradients.
try:
grad_fn = ops.get_gradient_function(op)
except LookupError:
raise LookupError(
"No gradient defined for operation '%s' (op type: %s)"
% (op.name, op.type))
if loop_state:
loop_state.EnterGradWhileContext(op, before=False)
if (grad_fn or is_func_call) and has_out_grads:
# NOTE: If _AggregatedGrads didn't compute a value for the i'th
# output, it means that the cost does not depend on output[i],
# therefore dC/doutput[i] is 0.
for i, out_grad in enumerate(out_grads):
if (not isinstance(out_grad, ops.Tensor)
and not out_grad) and (
(not grad_fn and is_func_call)
or _IsTrainable(op.outputs[i])):
# Only trainable outputs or outputs for a function call that
# will use SymbolicGradient get a zero gradient. Gradient
# functions should ignore the gradient for other outputs.
# TODO(apassos) gradients of resource handles might be an
# issue here because of zeros.
if loop_state:
out_grads[i] = loop_state.ZerosLike(op, i)
else:
out_grads[
i] = control_flow_ops.ZerosLikeOutsideLoop(
op, i)
with ops.name_scope(op.name + "_grad"):
# pylint: disable=protected-access
with ops.get_default_graph()._original_op(op):
# pylint: enable=protected-access
if grad_fn:
# If grad_fn was found, do not use SymbolicGradient even for
# functions.
in_grads = _MaybeCompile(
grad_scope, op, func_call,
lambda: grad_fn(op, *out_grads))
else:
# For function call ops, we add a 'SymbolicGradient'
# node to the graph to compute gradients.
in_grads = _MaybeCompile(
grad_scope, op, func_call,
lambda: _SymGrad(op, out_grads))
in_grads = _AsList(in_grads)
_VerifyGeneratedGradients(in_grads, op)
if gate_gradients and len(
[x for x in in_grads if x is not None]) > 1:
with ops.device(None):
with ops.colocate_with(
None, ignore_existing=True):
in_grads = control_flow_ops.tuple(
in_grads)
_LogOpGradients(op, out_grads, in_grads)
else:
# If no grad_fn is defined or none of out_grads is available,
# just propagate a list of None backwards.
in_grads = [None] * len(op.inputs)
for i, (t_in, in_grad) in enumerate(zip(op.inputs, in_grads)):
if in_grad is not None:
if (isinstance(in_grad, ops.Tensor)
and t_in.dtype != dtypes.resource):
try:
in_grad.set_shape(t_in.get_shape())
except ValueError:
raise ValueError(
"Incompatible shapes between op input and calculated "
"input gradient. Forward operation: %s. Input index: %d. "
"Original input shape: %s. "
"Calculated input gradient shape: %s" %
(op.name, i, t_in.shape, in_grad.shape))
_SetGrad(grads, t_in, in_grad)
if loop_state:
loop_state.ExitGradWhileContext(op, before=False)
# Update pending count for the inputs of op and enqueue ready ops.
assert op.graph == ys[0].graph
_UpdatePendingAndEnqueueReady(grads, op, queue, pending_count,
loop_state)
if loop_state:
loop_state.PostProcessing()
return [_GetGrad(grads, x) for x in xs]