def matvecFWD(self, arg):
'''Callback function for performing the matrix vector product of the
workflow's full Jacobian with an incoming vector arg.'''
comps = self._comp_edge_list()
if '@fake' in comps:
del comps['@fake']
result = zeros(len(arg))
# We can call applyJ on each component one-at-a-time, and poke the
# results into the result vector.
for compname, data in comps.iteritems():
comp_inputs = data['inputs']
comp_outputs = data['outputs']
comp_residuals = data['residuals']
inputs = {}
outputs = {}
out_bounds = []
for varname in comp_inputs:
node = '%s.%s' % (compname, varname)
i1, i2 = self.get_bounds(node)
if isinstance(i1, list):
inputs[varname] = arg[i1].copy()
else:
inputs[varname] = arg[i1:i2].copy()
for varname in comp_outputs:
node = '%s.%s' % (compname, varname)
i1, i2 = self.get_bounds(node)
out_bounds.append((varname, i1, i2))
if isinstance(i1, list):
if varname in comp_residuals:
outputs[varname] = zeros((1, 1))
else:
inputs[varname] = arg[i1].copy()
outputs[varname] = arg[i1].copy()
else:
if varname in comp_residuals:
outputs[varname] = zeros((i2 - i1))
else:
inputs[varname] = arg[i1:i2].copy()
outputs[varname] = arg[i1:i2].copy()
if '~' in compname:
comp = self._derivative_graph.node[compname]['pa_object']
else:
comp = self.scope.get(compname)
# Preconditioning
# Currently not implemented in forward mode, mostly because this
# mode requires post multiplication of the result by the M after
# you have the final gradient.
#if hasattr(comp, 'applyMinv'):
#inputs = applyMinv(comp, inputs)
applyJ(comp, inputs, outputs, comp_residuals,
self._shape_cache.get(compname),
self._J_cache.get(compname))
#print inputs, outputs
for varname, i1, i2 in out_bounds:
if isinstance(i1, list):
result[i1] = outputs[varname].copy()
else:
result[i1:i2] = outputs[varname].copy()
# Each parameter adds an equation
for src, targets in self._edges.iteritems():
if '@in' in src or '@fake' in src:
if not isinstance(targets, list):
targets = [targets]
for target in targets:
i1, i2 = self.get_bounds(target)
#if isinstance(i1, list):
# result[i1] = arg[i1]
#else:
result[i1:i2] = arg[i1:i2]
#print arg, result
return result
def _matvecFWD(self, arg):
'''Callback function for performing the matrix vector product of the
state-to-residual Jacobian with an incoming vector arg.'''
result = np.zeros(len(arg))
comp_inputs = self.list_states()
comp_outputs = self.list_residuals()
inputs = {}
outputs = {}
idx = 0
for varname in comp_inputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
inputs[varname] = arg[i1:i2].copy()
idx += size
for varname in comp_outputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
inputs[varname] = arg[i1:i2].copy()
outputs[varname] = arg[i1:i2].copy()
idx += size
applyJ(self, inputs, outputs, [], self._shape_cache, J=self._cache_J)
#print inputs, outputs
idx = 0
# Each state input adds an equation
for varname in comp_inputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
result[i1:i2] = arg[i1:i2].copy()
idx += size
for varname in comp_outputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
result[i1:i2] = outputs[varname]
idx += size
#print arg, result
return result
def applyJ(self, system, variables):
""" Wrapper for component derivative specification methods.
Forward Mode.
"""
applyJ(system, variables)
def _matvecFWD(self, arg):
'''Callback function for performing the matrix vector product of the
state-to-residual Jacobian with an incoming vector arg.'''
result = np.zeros(len(arg))
comp_inputs = self.list_states()
comp_outputs = self.list_residuals()
inputs = {}
outputs = {}
idx = 0
for varname in comp_inputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
inputs[varname] = arg[i1:i2].copy()
idx += size
for varname in comp_outputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
inputs[varname] = arg[i1:i2].copy()
outputs[varname] = arg[i1:i2].copy()
idx += size
applyJ(self, inputs, outputs, [], self._shape_cache, J=self._cache_J)
#print inputs, outputs
idx = 0
# Each state input adds an equation
for varname in comp_inputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
result[i1:i2] = arg[i1:i2].copy()
idx += size
for varname in comp_outputs:
val = getattr(self, varname)
flatval = flattened_value(varname, val)
size = len(flatval)
i1, i2 = idx, idx + size
result[i1:i2] = outputs[varname]
idx += size
#print arg, result
return result
def matvecFWD(self, arg):
'''Callback function for performing the matrix vector product of the
workflow's full Jacobian with an incoming vector arg.'''
comps = self._comp_edge_list()
if '@fake' in comps:
del comps['@fake']
result = zeros(len(arg))
# We can call applyJ on each component one-at-a-time, and poke the
# results into the result vector.
for compname, data in comps.iteritems():
comp_inputs = data['inputs']
comp_outputs = data['outputs']
comp_residuals = data['residuals']
inputs = {}
outputs = {}
out_bounds = []
for varname in comp_inputs:
node = '%s.%s' % (compname, varname)
i1, i2 = self.get_bounds(node)
if isinstance(i1, list):
inputs[varname] = arg[i1].copy()
else:
inputs[varname] = arg[i1:i2].copy()
for varname in comp_outputs:
node = '%s.%s' % (compname, varname)
i1, i2 = self.get_bounds(node)
out_bounds.append((varname, i1, i2))
if isinstance(i1, list):
if varname in comp_residuals:
outputs[varname] = zeros((1, 1))
else:
inputs[varname] = arg[i1].copy()
outputs[varname] = arg[i1].copy()
else:
if varname in comp_residuals:
outputs[varname] = zeros((i2-i1))
else:
inputs[varname] = arg[i1:i2].copy()
outputs[varname] = arg[i1:i2].copy()
if '~' in compname:
comp = self._derivative_graph.node[compname]['pa_object']
else:
comp = self.scope.get(compname)
# Preconditioning
# Currently not implemented in forward mode, mostly because this
# mode requires post multiplication of the result by the M after
# you have the final gradient.
#if hasattr(comp, 'applyMinv'):
#inputs = applyMinv(comp, inputs)
applyJ(comp, inputs, outputs, comp_residuals,
self._shape_cache.get(compname), self._J_cache.get(compname))
#print inputs, outputs
for varname, i1, i2 in out_bounds:
if isinstance(i1, list):
result[i1] = outputs[varname].copy()
else:
result[i1:i2] = outputs[varname].copy()
# Each parameter adds an equation
for src, targets in self._edges.iteritems():
if '@in' in src or '@fake' in src:
if not isinstance(targets, list):
targets = [targets]
for target in targets:
i1, i2 = self.get_bounds(target)
result[i1:i2] = arg[i1:i2]
#print arg, result
return result
def applyJ(self, system, variables):
""" Wrapper for component derivative specification methods.
Forward Mode.
"""
applyJ(system, variables)
