def precision(y_true, y_score, k=None, return_bounds=False):
"""
If return_bounds is False then returns precision on the
labeled examples in the top k.
If return_bounds is True the returns a tuple containing:
- precision on the labeled examples in the top k
- number of labeled examples in the top k
- lower bound of precision in the top k, assuming all
unlabaled examples are False
- upper bound of precision in the top k, assuming all
unlabaled examples are True
"""
y_true, y_score = to_float(y_true, y_score)
top = _argtop(y_score, k)
n = np.nan_to_num(y_true[top]).sum() # fill missing labels with 0
d = (~np.isnan(y_true[top])).sum() # count number of labels
p = n / d
if return_bounds:
k = len(y_true) if k is None else k
bounds = (n / k, (n + k - d) / k) if k != 0 else (np.nan, np.nan)
return p, d, bounds[0], bounds[1]
else:
return p
def precision(y_true, y_score, k=None, return_bounds=False):
"""
If return_bounds is False then returns precision on the
labeled examples in the top k.
If return_bounds is True the returns a tuple containing:
- precision on the labeled examples in the top k
- number of labeled examples in the top k
- lower bound of precision in the top k, assuming all
unlabaled examples are False
- upper bound of precision in the top k, assuming all
unlabaled examples are True
"""
y_true, y_score = to_float(y_true, y_score)
top = _argtop(y_score, k)
n = np.nan_to_num(y_true[top]).sum() # fill missing labels with 0
d = (~np.isnan(y_true[top])).sum() # count number of labels
p = n/d
if return_bounds:
k = len(y_true) if k is None else k
bounds = (n/k, (n+k-d)/k) if k != 0 else (np.nan, np.nan)
return p, d, bounds[0], bounds[1]
else:
return p
def count(y_true, y_score=None, countna=False):
"""
Counts the number of examples. If countna is False then only count labeled examples, i.e. those with y_true not NaN
"""
if not countna:
return (~np.isnan(to_float(y_true))).sum()
else:
return len(y_true)
def count(y_true, y_score=None, countna=False):
"""
Counts the number of examples. If countna is False then only count labeled examples,
i.e. those with y_true not NaN
"""
if not countna:
return (~np.isnan(to_float(y_true))).sum()
else:
return len(y_true)
def precision_series(y_true, y_score, k=None):
"""
Returns series of length k whose i-th entry is the precision in the top i
TODO: extrapolate here
"""
y_true, y_score = to_float(y_true, y_score)
top = _argsort(y_score, k)
n = np.nan_to_num(y_true[top]).cumsum() # fill missing labels with 0
d = (~np.isnan(y_true[top])).cumsum() # count number of labels
return pd.Series(n / d, index=np.arange(1, len(n) + 1))
def precision_series(y_true, y_score, k=None):
"""
Returns series of length k whose i-th entry is the precision in the top i
TODO: extrapolate here
"""
y_true, y_score = to_float(y_true, y_score)
top = _argsort(y_score, k)
n = np.nan_to_num(y_true[top]).cumsum() # fill missing labels with 0
d = (~np.isnan(y_true[top])).cumsum() # count number of labels
return pd.Series(n/d, index=np.arange(1,len(n)+1))
def recall_series(y_true, y_score, k=None, value=True):
"""
Returns series of length k whose i-th entry is the recall in the top i
"""
y_true, y_score = to_float(y_true, y_score)
top = _argsort(y_score, k)
if not value:
y_true = 1 - y_true
a = np.nan_to_num(y_true[top]).cumsum()
return pd.Series(a, index=np.arange(1, len(a) + 1))
def recall_series(y_true, y_score, k=None, value=True):
"""
Returns series of length k whose i-th entry is the recall in the top i
"""
y_true, y_score = to_float(y_true, y_score)
top = _argsort(y_score, k)
if not value:
y_true = 1-y_true
a = np.nan_to_num(y_true[top]).cumsum()
return pd.Series(a, index=np.arange(1,len(a)+1))
def count_series(y_true, y_score, countna=False):
"""
Returns series whose i-th entry is the number of examples in the top i
"""
y_true, y_score = to_float(y_true, y_score)
top = _argsort(y_score)
if not countna:
a = (~np.isnan(y_true[top])).cumsum()
else:
a = range(1, len(y_true) + 1)
return pd.Series(a, index=range(1, len(a) + 1))
def count_series(y_true, y_score, countna=False):
"""
Returns series whose i-th entry is the number of examples in the top i
"""
y_true, y_score = to_float(y_true, y_score)
top = _argsort(y_score)
if not countna:
a = (~np.isnan(y_true[top])).cumsum()
else:
a = range(1, len(y)+1)
return pd.Series(a, index=range(1, len(a)+1))
def recall(y_true, y_score, k=None, value=True):
"""
Returns recall (number of positive examples) in the top k
If value is False then counts number of negative examples
TODO: add prop argument to return recall proportion instead of count
"""
y_true, y_score = to_float(y_true, y_score)
top = _argtop(y_score, k)
if not value:
y_true = 1 - y_true
r = np.nan_to_num(y_true[top]).sum()
return r
def recall(y_true, y_score, k=None, value=True):
"""
Returns recall (number of positive examples) in the top k
If value is False then counts number of negative examples
TODO: add prop argument to return recall proportion instead of count
"""
y_true, y_score = to_float(y_true, y_score)
top = _argtop(y_score, k)
if not value:
y_true = 1-y_true
r = np.nan_to_num(y_true[top]).sum()
return r
def true_score(y, outcome='true', score='score', **subset_args):
y = y_subset(y, outcome=outcome, score=score, **subset_args)
return util.to_float(y[outcome], y[score])
def true_score(y, outcome='true', score='score', **subset_args):
y = y_subset(y, outcome=outcome, score=score, **subset_args)
return util.to_float(y[outcome], y[score])
