e24331bf11
It was possible for a TensorArrayWrite/Scatter/Split to occur after its corresponding TensorArrayGrad op was executed. The TensorArrayGrad op must execute after the last write so that it is created with the correct size. Since TensorArrayGrad converts its source TensorArray to fixed-size, this causes an exception when the final TensorArrayWrite executes since it cannot grow the TensorArray. This fix introduces a control dependency on the TensorArrayWrite op, to ensure TensorArrayGrad runs after the last write. Also, fix a similar pattern in TensorArrayScatter and TensorArraySplit's gradients even though the race has not been observed for them. PiperOrigin-RevId: 303515506 Change-Id: Ia363b94922b8855f01faa801818839559d55b0da
268 lines
10 KiB
Python
268 lines
10 KiB
Python
# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
# ==============================================================================
|
|
"""Gradients for operators defined in tensor_array_ops.py."""
|
|
from __future__ import absolute_import
|
|
from __future__ import division
|
|
from __future__ import print_function
|
|
|
|
from tensorflow.python.framework import ops
|
|
from tensorflow.python.ops import array_ops
|
|
from tensorflow.python.ops import tensor_array_ops
|
|
|
|
# TODO(b/31222613): These ops may be differentiable, and there may be
|
|
# latent bugs here.
|
|
ops.NotDifferentiable("TensorArray")
|
|
ops.NotDifferentiable("TensorArrayGrad")
|
|
ops.NotDifferentiable("TensorArraySize")
|
|
ops.NotDifferentiable("TensorArrayClose")
|
|
|
|
ops.NotDifferentiable("TensorArrayV2")
|
|
ops.NotDifferentiable("TensorArrayGradV2")
|
|
ops.NotDifferentiable("TensorArraySizeV2")
|
|
ops.NotDifferentiable("TensorArrayCloseV2")
|
|
|
|
ops.NotDifferentiable("TensorArrayV3")
|
|
ops.NotDifferentiable("TensorArrayGradV3")
|
|
ops.NotDifferentiable("TensorArrayGradWithShape")
|
|
ops.NotDifferentiable("TensorArraySizeV3")
|
|
ops.NotDifferentiable("TensorArrayCloseV3")
|
|
|
|
|
|
def _GetGradSource(op_or_tensor):
|
|
"""Identify which call to tf.gradients created this gradient op or tensor.
|
|
|
|
TensorArray gradient calls use an accumulator TensorArray object. If
|
|
multiple gradients are calculated and run in the same session, the multiple
|
|
gradient nodes may accidentally flow through the same accumulator TensorArray.
|
|
This double counting breaks the TensorArray gradient flow.
|
|
|
|
The solution is to identify which gradient call this particular
|
|
TensorArray*Grad is being called in, by looking at the input gradient
|
|
tensor's name, and create or lookup an accumulator gradient TensorArray
|
|
associated with this specific call. This solves any confusion and ensures
|
|
different gradients from the same forward graph get their own accumulators.
|
|
|
|
This function creates the unique label associated with the tf.gradients call
|
|
that is used to create the gradient TensorArray.
|
|
|
|
Args:
|
|
op_or_tensor: `Tensor` or `Operation` which is an input to a
|
|
TensorArray*Grad call.
|
|
|
|
Returns:
|
|
A python string, the unique label associated with this particular
|
|
gradients calculation.
|
|
|
|
Raises:
|
|
ValueError: If not called within a gradients calculation.
|
|
"""
|
|
name_tokens = op_or_tensor.name.split("/")
|
|
grad_pos = [i for i, x in enumerate(name_tokens) if x.startswith("gradients")]
|
|
if not grad_pos:
|
|
raise ValueError(
|
|
"Expected op/tensor name to start with gradients (excluding scope)"
|
|
", got: {}. This means that a tf.gradients op with this op in its "
|
|
"dependency path has a custom name that does not start with "
|
|
"'gradients'. Please make sure all calls to tf.gradients that have "
|
|
"non-empty 'name' arguments use names that start with "
|
|
"'gradients'.".format(op_or_tensor.name))
|
|
return "/".join(name_tokens[:grad_pos[-1] + 1])
|
|
|
|
|
|
@ops.RegisterGradient("TensorArrayRead")
|
|
@ops.RegisterGradient("TensorArrayReadV2")
|
|
@ops.RegisterGradient("TensorArrayReadV3")
|
|
def _TensorArrayReadGrad(op, grad):
|
|
"""Gradient for TensorArrayRead.
|
|
|
|
Args:
|
|
op: Forward TensorArrayRead op.
|
|
grad: Gradient `Tensor` to TensorArrayRead.
|
|
|
|
Returns:
|
|
A flow `Tensor`, which can be used in control dependencies to
|
|
force the write of `grad` to the gradient `TensorArray`.
|
|
"""
|
|
# Note: the forward flow dependency in the call to grad() is necessary for
|
|
# the case of dynamic sized TensorArrays. When creating the gradient
|
|
# TensorArray, the final size of the forward array must be known.
|
|
# For this we need to wait until it has been created by depending on
|
|
# the input flow of the original op.
|
|
handle = op.inputs[0]
|
|
index = op.inputs[1]
|
|
flow = op.inputs[2]
|
|
dtype = op.get_attr("dtype")
|
|
grad_source = _GetGradSource(grad)
|
|
g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
|
|
colocate_with_first_write_call=False)
|
|
.grad(source=grad_source, flow=flow))
|
|
w_g = g.write(index, grad)
|
|
return [None, None, w_g.flow]
|
|
|
|
|
|
@ops.RegisterGradient("TensorArrayWrite")
|
|
@ops.RegisterGradient("TensorArrayWriteV2")
|
|
@ops.RegisterGradient("TensorArrayWriteV3")
|
|
def _TensorArrayWriteGrad(op, flow):
|
|
"""Gradient for TensorArrayWrite.
|
|
|
|
Args:
|
|
op: Forward TensorArrayWrite op.
|
|
flow: Gradient `Tensor` flow to TensorArrayWrite.
|
|
|
|
Returns:
|
|
A grad `Tensor`, the gradient created in an upstream ReadGrad or PackGrad.
|
|
"""
|
|
# handle is the output store_handle of TensorArrayReadGrad or
|
|
# the handle output of TensorArrayWriteGrad. we must use this one.
|
|
handle = op.inputs[0]
|
|
index = op.inputs[1]
|
|
dtype = op.get_attr("T")
|
|
grad_source = _GetGradSource(flow)
|
|
flow_out = array_ops.identity(op.outputs[0], "flow_out")
|
|
# Avoid a race condition where the TensorArrayGrad op is executed before the
|
|
# final TensorArrayWrite by adding a control dependency on the output flow of
|
|
# the write to the input flow to the TensorArrayGrad.
|
|
with ops.control_dependencies([flow_out]):
|
|
flow = array_ops.identity(flow, "write_barrier")
|
|
g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
|
|
colocate_with_first_write_call=False)
|
|
.grad(source=grad_source, flow=flow))
|
|
grad = g.read(index)
|
|
return [None, None, grad, flow]
|
|
|
|
|
|
@ops.RegisterGradient("TensorArrayGather")
|
|
@ops.RegisterGradient("TensorArrayGatherV2")
|
|
@ops.RegisterGradient("TensorArrayGatherV3")
|
|
def _TensorArrayGatherGrad(op, grad):
|
|
"""Gradient for TensorArrayGather.
|
|
|
|
Args:
|
|
op: Forward TensorArrayGather op.
|
|
grad: Gradient `Tensor` to TensorArrayGather.
|
|
|
|
Returns:
|
|
A flow `Tensor`, which can be used in control dependencies to
|
|
force the write of `grad` to the gradient `TensorArray`.
|
|
"""
|
|
# Note: the forward flow dependency in the call to grad() is necessary for
|
|
# the case of dynamic sized TensorArrays. When creating the gradient
|
|
# TensorArray, the final size of the forward array must be known.
|
|
# For this we need to wait until it has been created by depending on
|
|
# the input flow of the original op.
|
|
handle = op.inputs[0]
|
|
indices = op.inputs[1]
|
|
flow = op.inputs[2]
|
|
dtype = op.get_attr("dtype")
|
|
grad_source = _GetGradSource(grad)
|
|
g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
|
|
colocate_with_first_write_call=False)
|
|
.grad(source=grad_source, flow=flow))
|
|
u_g = g.scatter(indices, grad)
|
|
return [None, None, u_g.flow]
|
|
|
|
|
|
@ops.RegisterGradient("TensorArrayScatter")
|
|
@ops.RegisterGradient("TensorArrayScatterV2")
|
|
@ops.RegisterGradient("TensorArrayScatterV3")
|
|
def _TensorArrayScatterGrad(op, flow):
|
|
"""Gradient for TensorArrayScatter.
|
|
|
|
Args:
|
|
op: Forward TensorArrayScatter op.
|
|
flow: Gradient `Tensor` flow to TensorArrayScatter.
|
|
|
|
Returns:
|
|
A grad `Tensor`, the gradient created in upstream ReadGrads or PackGrad.
|
|
"""
|
|
handle = op.inputs[0]
|
|
indices = op.inputs[1]
|
|
dtype = op.get_attr("T")
|
|
grad_source = _GetGradSource(flow)
|
|
flow_out = array_ops.identity(op.outputs[0], "flow_out")
|
|
# Avoid a race condition where the TensorArrayGrad op is executed before the
|
|
# TensorArrayScatter by adding a control dependency on the output flow of
|
|
# the scatter to the input flow to the TensorArrayGrad.
|
|
with ops.control_dependencies([flow_out]):
|
|
flow = array_ops.identity(flow, "write_barrier")
|
|
g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
|
|
colocate_with_first_write_call=False)
|
|
.grad(source=grad_source, flow=flow))
|
|
grad = g.gather(indices)
|
|
return [None, None, grad, flow]
|
|
|
|
|
|
@ops.RegisterGradient("TensorArrayConcat")
|
|
@ops.RegisterGradient("TensorArrayConcatV2")
|
|
@ops.RegisterGradient("TensorArrayConcatV3")
|
|
def _TensorArrayConcatGrad(op, grad, unused_lengths_grad):
|
|
"""Gradient for TensorArrayConcat.
|
|
|
|
Args:
|
|
op: Forward TensorArrayConcat op.
|
|
grad: Gradient `Tensor` to TensorArrayConcat.
|
|
|
|
Returns:
|
|
A flow `Tensor`, which can be used in control dependencies to
|
|
force the write of `grad` to the gradient `TensorArray`.
|
|
"""
|
|
# Note: the forward flow dependency in the call to grad() is necessary for
|
|
# the case of dynamic sized TensorArrays. When creating the gradient
|
|
# TensorArray, the final size of the forward array must be known.
|
|
# For this we need to wait until it has been created by depending on
|
|
# the input flow of the original op.
|
|
handle = op.inputs[0]
|
|
flow = op.inputs[1]
|
|
lengths = op.outputs[1]
|
|
dtype = op.get_attr("dtype")
|
|
grad_source = _GetGradSource(grad)
|
|
g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
|
|
colocate_with_first_write_call=False)
|
|
.grad(source=grad_source, flow=flow))
|
|
u_g = g.split(grad, lengths=lengths)
|
|
# handle, flow_in
|
|
return [None, u_g.flow]
|
|
|
|
|
|
@ops.RegisterGradient("TensorArraySplit")
|
|
@ops.RegisterGradient("TensorArraySplitV2")
|
|
@ops.RegisterGradient("TensorArraySplitV3")
|
|
def _TensorArraySplitGrad(op, flow):
|
|
"""Gradient for TensorArraySplit.
|
|
|
|
Args:
|
|
op: Forward TensorArraySplit op.
|
|
flow: Gradient `Tensor` flow to TensorArraySplit.
|
|
|
|
Returns:
|
|
A grad `Tensor`, the gradient created in upstream ReadGrads or PackGrad.
|
|
"""
|
|
handle = op.inputs[0]
|
|
dtype = op.get_attr("T")
|
|
grad_source = _GetGradSource(flow)
|
|
flow_out = array_ops.identity(op.outputs[0], "flow_out")
|
|
# Avoid a race condition where the TensorArrayGrad op is executed before the
|
|
# TensorArraySplit by adding a control dependency on the output flow of
|
|
# the split to the input flow to the TensorArrayGrad.
|
|
with ops.control_dependencies([flow_out]):
|
|
flow = array_ops.identity(flow, "write_barrier")
|
|
g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
|
|
colocate_with_first_write_call=False)
|
|
.grad(source=grad_source, flow=flow))
|
|
grad = g.concat()
|
|
# handle, value, lengths, flow_in
|
|
return [None, grad, None, flow]
|