"""
Metrics about regressions.
:githublink:`%|py|5`
"""
import io
import numpy
import pandas
[docs]def l1_reg_max(exp, val, max_val=180, nomax=False, exc=True):
"""
Implements a :epkg:`L1` scoring function which does not consider
error above threshold *max_val*.
:param exp: list of values or :epkg:`numpy:array`
:param val: list of values or :epkg:`numpy:array`
:param max_val: every value above *max_val* is replaced by *max_val*
before computing the differences
:param nomax: removes every value equal or above *nomax* in expected set,
then compute the score
:param raises: an exception if not enough submitted items
:return: score
If ``max_val==180``, the function computes:
.. math::
E = \\frac{1}{n} \\sum_{i=1}^n \\frac{\\left| \\min (Y_i, 180) - \\min(f(X_i), 180) \\right|}{180}
The computation is faster if :epkg:`numpy:array` are used
(for *exp* and *val*). *exp and *val* can be filenames or streams.
In that case, the function expects to find two columns: id, value
in both files or streams.
:githublink:`%|py|34`
"""
if isinstance(exp, numpy.ndarray) and isinstance(val, numpy.ndarray):
if len(exp) != len(val):
raise ValueError(
"Dimension mismatch {0} != {1}".format(len(exp), len(val)))
an = numpy.zeros((len(exp),))
an[:] = max_val
mv = numpy.minimum(an, val) # pylint: disable=E1111
me = numpy.minimum(an, exp) # pylint: disable=E1111
if nomax:
mv = mv[me < max_val] # pylint: disable=W0143,E1136
me = me[me < max_val] # pylint: disable=W0143,E1136
df = numpy.abs(mv - me) / max_val
return df.mean()
elif isinstance(exp, dict) and isinstance(val, dict):
if exc and len(exp) != len(val):
number_common = len(set(exp) & set(val))
raise ValueError(
"Dimension mismatch {0} != {1} (#common={2})".format(len(exp), len(val), number_common))
r = 0.0
nb = 0
for k, e in exp.items():
if k in val:
v = val[k]
try:
mv = min(v, max_val)
except TypeError:
return numpy.nan
try:
ev = min(e, max_val)
except TypeError:
return numpy.nan
if nomax and ev >= max_val:
continue
d = abs(mv - ev)
r += 1. * d / max_val
nb += 1
elif exc:
raise ValueError("Missing key in prediction {0}".format(k))
else:
r += 1.
nb += 1
return r / nb if nb > 0 else 0.0
elif isinstance(exp, (str, io.StringIO)) and isinstance(val, (str, io.StringIO)):
# We expect filenames.
d1 = pandas.read_csv(exp, header=None, sep=";")
d2 = pandas.read_csv(val, header=None, sep=";")
dd1 = {}
for k, v in d1.itertuples(name=None, index=False):
if k in dd1:
raise KeyError("Key '{}' present at least twice.".format(k))
dd1[k] = v
dd2 = {}
for k, v in d2.itertuples(name=None, index=False):
if k in dd2:
raise KeyError("Key '{}' present at least twice.".format(k))
dd2[k] = v
return l1_reg_max(dd1, dd2, max_val=max_val, nomax=nomax, exc=exc)
elif isinstance(exp, list) and isinstance(val, list):
if len(exp) != len(val):
raise ValueError(
"Dimension mismatch {0} != {1}".format(len(exp), len(val)))
r = 0.0
nb = 0
for e, v in zip(exp, val):
ev = min(e, max_val)
if nomax and ev >= max_val:
continue
mv = min(v, max_val)
d = abs(mv - ev)
r += 1. * d / max_val
nb += 1
return r / nb if nb > 0 else 0.0
else:
raise TypeError(
"Inconsistent types {0} != {1}".format(type(exp), type(val)))
