def imperative_grad(
tape,
target,
sources,
output_gradients=None):
"""Computes gradients from the imperatively defined tape on top of the stack.
Works by filtering the tape, computing how many downstream usages are of each
tensor and entry, and repeatedly applying backward functions until we have
gradients for all sources.
Args:
tape: the gradient tape which stores the trace.
target: either a Tensor or list of Tensors to be differentiated.
sources: list of Tensors for which we want gradients
output_gradients: if not None, a list of gradient provided for each Target,
or None if we are to use the target's computed downstream gradient.
Returns:
the gradient wrt each of the sources.
Raises:
RuntimeError: if something goes wrong.
"""
return pywrap_tensorflow.TFE_Py_TapeGradient(
tape._tape, # pylint: disable=protected-access
target,
sources,
output_gradients)
def imperative_grad(vspace, tape, target, sources, output_gradients=None):
"""Computes gradients from the imperatively defined tape on top of the stack.
Works by filtering the tape, computing how many downstream usages are of each
tensor and entry, and repeatedly applying backward functions until we have
gradients for all sources.
Args:
vspace: the vector space in which to differentiate.
tape: the gradient tape which stores the trace.
target: either a Tensor or list of Tensors to be differentiated.
sources: list of Tensors for which we want gradients
output_gradients: if not None, a list of gradient provided for each Target,
or None if we are to use the target's computed downstream gradient.
Returns:
the gradient wrt each of the sources.
Raises:
RuntimeError: if something goes wrong.
ValueError: if there is no sequence of differentiable operations connecting
a source and any target Tensor. This can happen either if the target is
not computed based on the source, if the tracing was set up incorrectly,
or if only non-differentiable functions of the source were used in the
computation of target.
"""
with errors.raise_exception_on_not_ok_status() as status:
return pywrap_tensorflow.TFE_Py_TapeGradient(tape._tape, vspace,
target, sources,
output_gradients, status) # pylint: disable=protected-access
def imperative_grad(tape,
target,
sources,
output_gradients=None,
sources_raw=None,
unconnected_gradients=UnconnectedGradients.NONE):
"""Computes gradients from the imperatively defined tape on top of the stack.
Works by filtering the tape, computing how many downstream usages are of each
tensor and entry, and repeatedly applying backward functions until we have
gradients for all sources.
Args:
tape: the gradient tape which stores the trace.
target: either a Tensor or list of Tensors to be differentiated.
sources: list of Tensors for which we want gradients
output_gradients: if not None, a list of gradient provided for each Target,
or None if we are to use the target's computed downstream gradient.
sources_raw: if not None, a list of the source python objects from which the
sources were generated. Should have the same length as sources. Only needs
to be populated if unconnected_gradients is 'zero'.
unconnected_gradients: determines the value returned if the target and
sources are unconnected. When 'none' the value returned is None wheras when
'zero' a zero tensor in the same shape as the sources is returned.
Returns:
the gradient wrt each of the sources.
Raises:
ValueError: if the arguments are invalid.
RuntimeError: if something goes wrong.
"""
try:
unconnected_gradients = UnconnectedGradients(unconnected_gradients)
except ValueError:
raise ValueError(
"Unknown value for unconnected_gradients: %r" % unconnected_gradients)
return pywrap_tensorflow.TFE_Py_TapeGradient(
tape._tape, # pylint: disable=protected-access
target,
sources,
output_gradients,
sources_raw,
compat.as_str(unconnected_gradients.value))
