Lines Matching refs:flow
46 gradient nodes may accidentally flow throuth the same accumulator TensorArray.
47 This double counting breaks the TensorArray gradient flow.
89 A flow `Tensor`, which can be used in control dependencies to
92 # Note: the forward flow dependency in the call to grad() is necessary for
96 # the input flow of the original op.
99 flow = op.inputs[2]
102 g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
104 .grad(source=grad_source, flow=flow))
106 return [None, None, w_g.flow]
112 def _TensorArrayWriteGrad(op, flow):
117 flow: Gradient `Tensor` flow to TensorArrayWrite.
127 grad_source = _GetGradSource(flow)
128 g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
130 .grad(source=grad_source, flow=flow))
132 return [None, None, grad, flow]
146 A flow `Tensor`, which can be used in control dependencies to
149 # Note: the forward flow dependency in the call to grad() is necessary for
153 # the input flow of the original op.
156 flow = op.inputs[2]
159 g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
161 .grad(source=grad_source, flow=flow))
163 return [None, None, u_g.flow]
169 def _TensorArrayScatterGrad(op, flow):
174 flow: Gradient `Tensor` flow to TensorArrayScatter.
182 grad_source = _GetGradSource(flow)
183 g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
185 .grad(source=grad_source, flow=flow))
187 return [None, None, grad, flow]
201 A flow `Tensor`, which can be used in control dependencies to
204 # Note: the forward flow dependency in the call to grad() is necessary for
208 # the input flow of the original op.
210 flow = op.inputs[1]
214 g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
216 .grad(source=grad_source, flow=flow))
219 return [None, u_g.flow]
225 def _TensorArraySplitGrad(op, flow):
230 flow: Gradient `Tensor` flow to TensorArraySplit.
237 grad_source = _GetGradSource(flow)
238 g = (tensor_array_ops.TensorArray(dtype=dtype, handle=handle, flow=flow,
240 .grad(source=grad_source, flow=flow))
243 return [None, grad, None, flow]
