def __init__(self, **kwargs):
self.TILTED_Q = retrieve_param_opt("TILTED_Q", kwargs)
self.LINEX_C = retrieve_param_opt("LINEX_C", kwargs)
print("[LossFactory] Configuration: %s" %
" ".join("%s=%s" % (k, format_value(v))
for k, v in vars(self).items() if k.upper() == k))
def __init__(self, **kwargs):
self.H_NSAMPLES_UTILITY_TERM_THETA = retrieve_param(
"H_NSAMPLES_UTILITY_TERM_THETA", kwargs)
self.H_NSAMPLES_UTILITY_TERM_Y = retrieve_param(
"H_NSAMPLES_UTILITY_TERM_Y", kwargs)
self.H_NUMERICAL_MAX_NITER = retrieve_param_opt(
"H_NUMERICAL_MAX_NITER", kwargs, 10000)
self.H_NUMERICAL_MAX_NITER_TOL = retrieve_param_opt(
"H_NUMERICAL_MAX_NITER_TOL", kwargs, 0.0001)
self.H_NUMERICAL_MAX_NITER_TOL_GOAL = retrieve_param_opt(
"H_NUMERICAL_MAX_NITER_TOL_GOAL", kwargs, 1e-10)
self.H_NUMERICAL_LR = retrieve_param_opt("H_NUMERICAL_LR", kwargs, 0.1)
self.H_NUMERICAL_START_FROM_PREVIOUS = retrieve_param_opt(
"H_NUMERICAL_START_FROM_PREVIOUS", kwargs, False)
self.sample_predictive_y0 = retrieve_param(
"sample_predictive_y0", kwargs) #default sampler of y
print("[HOptimizerFactory] Configuration: %s" %
" ".join("%s=%s" % (k, format_value(v))
for k, v in vars(self).items()
if k.upper() == k or k.startswith("sample")))
self.optimal_h_bayes_estimator = kwargs.get(
"optimal_h_bayes_estimator", None)
if kwargs.get("H_NUMERICAL_OPT_SCIPY", False):
print("[HOptimizerFactory] Choosing SciPy numerical optimization.")
self.optimal_h_numerically = optimal_h_numerically_ty_scipy
else:
print(
"[HOptimizerFactory] Choosing PyTorch numerical optimization.")
self.optimal_h_numerically = optimal_h_numerically_ty
def __init__(self, **kwargs):
self.M = retrieve_param("M", kwargs)
self.GAMMA = retrieve_param_opt("GAMMA", kwargs, 1.0)
print("[UtilityFactory] Configuration: %s" %
" ".join("%s=%s" % (k, format_value(v)) for k, v in vars(self).items() if k.upper()==k) )
def __init__(self,
y,
loss,
u,
sample_predictive_y,
optimal_h_bayes_estimator=None,
y_mask=None,
GAIN_OPTIMAL_H_NUMERICALLY=True,
RISK_OPTIMAL_H_NUMERICALLY=False,
EVAL_NSAMPLES_UTILITY_TERM_THETA=1000,
EVAL_NSAMPLES_UTILITY_TERM_Y=1,
EVAL_MAX_NITER=10000,
EVAL_SGD_PREC=0.0001,
EVAL_LR=0.01,
EVAL_RESAMPLE_EVERY_TIME=False):
"""
Args:
y Evaluation data.
y_mask A mask selecting data points for evaluation (default: all).
loss A function y x h -> loss used to calculate risks.
u A function y x h -> utility used to calculate gains.
sample_predictive_y A function that for each data point from y, generates samples from predictive posterior.
EVAL_RESAMPLE_EVERY_TIME Can results of sample_predictive_y, optimal_h_for_gain and optimal_h_for_risk be cached?
"""
self.y = y
self.y_mask = y_mask
if self.y_mask is None:
print2(
"[Evaluation] WARNING: using default all data points in evaluation."
)
env = torch if "cpu" in str(self.y.device).lower() else torch.cuda
self.y_mask = torch.ones_like(self.y).type(env.ByteTensor)
self.loss = loss
self.utility = u
self.sample_predictive_y = sample_predictive_y
self.optimal_h_bayes_estimator = optimal_h_bayes_estimator
if (self.optimal_h_bayes_estimator is None) and \
(not GAIN_OPTIMAL_H_NUMERICALLY or not RISK_OPTIMAL_H_NUMERICALLY):
print2(
"[Evaluation] WARNING: Optimal decisions h for both Risk and Gain will be obtained numerically."
)
self.optimal_h_bayes_estimator = lambda ys: None
GAIN_OPTIMAL_H_NUMERICALLY, RISK_OPTIMAL_H_NUMERICALLY = True, True
self.GAIN_OPTIMAL_H_NUMERICALLY = GAIN_OPTIMAL_H_NUMERICALLY
self.RISK_OPTIMAL_H_NUMERICALLY = RISK_OPTIMAL_H_NUMERICALLY
self.EVAL_NSAMPLES_UTILITY_TERM_THETA = EVAL_NSAMPLES_UTILITY_TERM_THETA
self.EVAL_NSAMPLES_UTILITY_TERM_Y = EVAL_NSAMPLES_UTILITY_TERM_Y
self.EVAL_MAX_NITER = EVAL_MAX_NITER
self.EVAL_SGD_PREC = EVAL_SGD_PREC
self.EVAL_LR = EVAL_LR
print("[Evaluation] Configuration: %s" %
" ".join("%s=%s" % (k, format_value(v))
for k, v in vars(self).items()))
if not EVAL_RESAMPLE_EVERY_TIME:
self.optimal_h_for_gain = cache_function_last_result(
self.optimal_h_for_gain)
self.optimal_h_for_risk = cache_function_last_result(
self.optimal_h_for_risk)
self.sample_predictive_posterior = cache_function_last_result(
self.sample_predictive_posterior)