def objective(X,
y,
clf,
meantype=meantype,
block_size=block_size,
percentage=0.01):
if binary:
sample_objectives = vec_logit(clf.decision_function(X) * y)
else:
scores = clf.decision_function(X)
sample_objectives = -scores[np.arange(X.shape[0]), y] + logsumexp(
scores, axis=1)
if meantype == "ordinary":
obj = sample_objectives.mean()
elif meantype == "ch":
obj = holland_catoni_estimator(sample_objectives)
elif meantype == "mom":
obj = median_of_means(sample_objectives,
int(block_size * len(sample_objectives)))
elif meantype == "tmean":
obj = fast_trimmed_mean(sample_objectives, len(sample_objectives),
percentage)
else:
raise ValueError("unknown mean")
if penalty != "none":
obj += lamda * penalties[penalty](clf.coef_)
return obj
def accuracy(X,
y,
clf,
meantype=meantype,
block_size=block_size,
percentage=0.01):
if binary:
scores = clf.decision_function(X) # clf.predict(X)
decisions = ((y * scores) > 0).astype(int).astype(float)
else:
predictions = clf.predict(X)
decisions = (y == predictions).astype(int).astype(float)
if meantype == "ordinary":
acc = decisions.mean()
elif meantype == "ch":
acc = holland_catoni_estimator(decisions)
elif meantype == "mom":
acc = median_of_means(decisions, int(block_size * len(decisions)))
elif meantype == "tmean":
acc = fast_trimmed_mean(decisions,
len(decisions),
percentage=percentage)
else:
raise ValueError("unknown mean")
return acc
def risk(X, y, clf, meantype=meantype, block_size=block_size, percentage=percentage):
if loss == "leastsquares":
objectives = 0.5 * ((clf.decision_function(X) - y) ** 2)
else:
ValueError("unimplemented loss %s"%loss)
if meantype == "ordinary":
obj = objectives.mean()
elif meantype == "catoni":
obj = holland_catoni_estimator(objectives)
elif meantype == "mom":
obj = median_of_means(objectives, int(block_size * len(objectives)))
elif meantype == "tmean":
obj = fast_trimmed_mean(objectives, len(objectives), percentage)
else:
raise ValueError("unknown mean")
if penalty:
obj += lamda * penalties[penalty](clf.coef_)
return obj