def evaluate(self, opt, grads, infos, objVar):
"""
Approximates gradient based on evaluated points.
@ In, opt, dict, current opt point (normalized)
@ In, grads, list(dict), evaluated neighbor points
@ In, infos, list(dict), info about evaluated neighbor points
@ In, objVar, string, objective variable
@ Out, magnitude, float, magnitude of gradient
@ Out, direction, dict, versor (unit vector) for gradient direction
@ Out, foundInf, bool, if True then infinity calculations were used
"""
gradient = {}
for g, pt in enumerate(grads):
info = infos[g]
delta = info['delta']
activeVar = info['optVar']
lossDiff = np.atleast_1d(
mathUtils.diffWithInfinites(pt[objVar], opt[objVar]))
grad = lossDiff / delta
gradient[activeVar] = grad
# obtain the magnitude and versor of the gradient to return
magnitude, direction, foundInf = mathUtils.calculateMagnitudeAndVersor(
list(gradient.values()))
direction = dict((var, float(direction[v]))
for v, var in enumerate(gradient.keys()))
return magnitude, direction, foundInf
def evaluate(self, opt, grads, infos, objVar):
"""
Approximates gradient based on evaluated points.
@ In, opt, dict, current opt point (normalized)
@ In, grads, list(dict), evaluated neighbor points
@ In, infos, list(dict), info about evaluated neighbor points
@ In, objVar, string, objective variable
@ Out, magnitude, float, magnitude of gradient
@ Out, direction, dict, versor (unit vector) for gradient direction
"""
gradient = {}
lossDiff = np.atleast_1d(mathUtils.diffWithInfinites(grads[0][objVar], opt[objVar]))
for var in self._optVars:
# don't assume delta is unchanged; calculate it here
delta = grads[0][var] - opt[var]
gradient[var] = lossDiff / delta
magnitude, direction, foundInf = mathUtils.calculateMagnitudeAndVersor(list(gradient.values()))
direction = dict((var, float(direction[v])) for v, var in enumerate(gradient.keys()))
return magnitude, direction, foundInf