def dfinfo(dataframe, asstring=True):
'''Returns a a dataframe with info about the given `dataframe` representing
a given dataset
'''
dinfo = dataset_info(dataframe)
classes = {c: dinfo.class_selector[c] for c in dinfo.classnames}
sum_dfs = odict()
empty_classes = set()
infocols = ['Min', 'Median', 'Max', '#NAs', '#<1Perc.', '#>99Perc.']
for classname, class_selector in classes.items():
sum_df = odict()
_dfr = dataframe[class_selector(dataframe)]
if _dfr.empty:
empty_classes.add(classname)
continue
# if _dfr.empty
for col in floatingcols(dataframe):
q01 = np.nanquantile(_dfr[col], 0.01)
q99 = np.nanquantile(_dfr[col], 0.99)
df1, df99 = _dfr[(_dfr[col] < q01)], _dfr[(_dfr[col] > q99)]
# segs1 = len(pd.unique(df1[ID_COL]))
# segs99 = len(pd.unique(df99[ID_COL]))
# stas1 = len(pd.unique(df1['station_id']))
# stas99 = len(pd.unique(df99['station_id']))
sum_df[col] = {
infocols[0]: np.nanmin(_dfr[col]),
infocols[1]: np.nanquantile(_dfr[col], 0.5),
infocols[2]: np.nanmax(_dfr[col]),
infocols[3]: (~np.isfinite(_dfr[col])).sum(),
infocols[4]: len(df1),
infocols[5]: len(df99)
# columns[5]: stas1 + stas99,
}
sum_dfs[classname + " (%d instances)" % len(_dfr)] = \
pd.DataFrame(data=list(sum_df.values()),
columns=infocols,
index=list(sum_df.keys()))
# return df2str(pd.DataFrame(data, columns=columns, index=index))
if not asstring:
return pd.concat(sum_dfs.values(), axis=0, keys=sum_dfs.keys(),
sort=False)
allstrs = []
for (key, val) in sum_dfs.items():
allstrs.extend(['', key, val.to_string()])
return '\n'.join(allstrs)
def _dfinfo(dataframe, asstring=True):
'''Returns a a dataframe with info about the given `dataframe`
'''
infocols = ['Min', 'Median', 'Max', '#NAs', '#<1Perc.', '#>99Perc.']
sum_df = odict()
# if _dfr.empty
for col in floatingcols(dataframe):
q01 = np.nanquantile(dataframe[col], 0.01)
q99 = np.nanquantile(dataframe[col], 0.99)
df1, df99 = dataframe[(dataframe[col] < q01)], dataframe[(dataframe[col] > q99)]
# segs1 = len(pd.unique(df1[ID_COL]))
# segs99 = len(pd.unique(df99[ID_COL]))
# stas1 = len(pd.unique(df1['station_id']))
# stas99 = len(pd.unique(df99['station_id']))
sum_df[col] = {
infocols[0]: np.nanmin(dataframe[col]),
infocols[1]: np.nanquantile(dataframe[col], 0.5),
infocols[2]: np.nanmax(dataframe[col]),
infocols[3]: (~np.isfinite(dataframe[col])).sum(),
infocols[4]: len(df1),
infocols[5]: len(df99)
# columns[5]: stas1 + stas99,
}
return pd.DataFrame(data=list(sum_df.values()),
columns=infocols,
index=list(sum_df.keys()))
def _dfinfo(dataframe):
'''Returns a dataframe with statistical info about the given `dataframe`
'''
infocols = ['Min', 'Median', 'Max', '#NAs', '#<1Perc.', '#>99Perc.']
defaultcolvalues = [np.nan, np.nan, np.nan, 0, 0, 0]
sum_df = odict()
# if _dfr.empty
for col in floatingcols(dataframe):
colvalues = defaultcolvalues
if not dataframe.empty:
q01 = np.nanquantile(dataframe[col], 0.01)
q99 = np.nanquantile(dataframe[col], 0.99)
df1, df99 = dataframe[(dataframe[col] <
q01)], dataframe[(dataframe[col] > q99)]
colvalues = [
np.nanmin(dataframe[col]), # Min
np.nanquantile(dataframe[col], 0.5), # Median
np.nanmax(dataframe[col]), # Max
(~np.isfinite(dataframe[col])).sum(), # #NAs
len(df1), # #<1Perc.
len(df99) # @>99Perc.
]
sum_df[col] = {i: v for i, v in zip(infocols, colvalues)}
return pd.DataFrame(data=list(sum_df.values()),
columns=infocols,
index=list(sum_df.keys()))
def dfinfo(dataframe, asstring=True):
'''Returns a a dataframe with info about the given `dataframe` representing
a given dataset (if asstring=False) or, if asstring=True, a string
representing the dataframe.
'''
classnames = CLASSNAMES
class_selectors = CLASS_SELECTORS
sum_dfs = odict()
for classname, class_selector in zip(classnames, class_selectors):
_dfr = dataframe[class_selector(dataframe)]
title = "Class '%s': %d of %d instances" % (classname, len(_dfr),
len(dataframe))
sum_dfs[title] = _dfinfo(_dfr)
# return a MultiIndex DataFrame:
if not asstring:
return pd.concat(sum_dfs.values(),
axis=0,
keys=sum_dfs.keys(),
sort=False)
allstrs = []
for (key, val) in sum_dfs.items():
allstrs.extend(['', key, '-' * len(key), val.to_string()])
return '\n'.join(allstrs)
def read_all_pred_dfs():
'''dict of pred_dataframe paths -> st_time'''
ret = []
for fle in listdir(tmpdir):
_ = join(tmpdir, fle, 'allset_test.hdf_')
if isfile(_):
ret.append(_)
ret2 = odict()
for _ in sorted(ret):
ret2[_] = stat(_).st_mtime
return ret2
def model_params(model_filename):
'''Converts the given model_filename (or absolute path) into a dict
of key -> tuple of **strings** (single values parameters will be mapped
to a 1-element tuple)
'''
pth = basename(model_filename)
pth_, ext = splitext(pth)
if ext == '.sklmodel':
pth = pth_
pars = pth.split('&')
ret = odict()
for par in pars:
key, val = par.split('=')
ret[uq(key)] = tuple(uq(_) for _ in val.split(','))
return ret
def set_classifiers_paths(self, classifiers_paths):
'''Builds and returns from this object parameters a list of N arguments
to be passed to `_create_save_classifier`.
Skips classifiers whose prediction dataframe already exists.
See `run_evaluation` for details
'''
ret = odict()
for clfpath in classifiers_paths:
outdir = splitext(clfpath)[0]
destpath = join(outdir, basename(self.testset_filepath))
if isfile(destpath):
continue
elif not isdir(dirname(destpath)):
makedirs(dirname(destpath))
if not isdir(dirname(destpath)):
continue
feats = self.model_params(clfpath)['feats']
ret[clfpath] = (destpath, feats)
self.classifiers_paths = ret
def model_filename(clf_class, tr_set, *features, **clf_params):
'''converts the given argument to a model filename, with extension
.sklmodel'''
pars = odict()
pars['clf'] = [str(clf_class.__name__)]
pars['tr_set'] = [
str(_)
for _ in (tr_set if isinstance(tr_set, (list,
tuple)) else [tr_set])
]
pars['feats'] = [str(_) for _ in sorted(features)]
for key in sorted(clf_params):
val = clf_params[key]
pars[q(key, safe=_safe)] = [
str(_)
for _ in (val if isinstance(val, (list, tuple)) else [val])
]
return '&'.join("%s=%s" % (k, ','.join(q(_, safe=_safe) for _ in v))
for k, v in pars.items()) + '.sklmodel'
def _get_summary_evaluationmetrics_mp(clfdir_and_testname):
clfdir, testname = clfdir_and_testname
filepath = join(clfdir, testname)
predicted_df = pd.read_hdf(filepath, columns=[OUTLIER_COL, PREDICT_COL])
cols = tuple(_METRIC_COLUMNS)
try:
# parse the clf file name (which is the directory name of the
# prediction dataframe we want to calculate metrics from), and
# try to guess if they can be floats or ints:
ret = odict()
for key, value in TestParam.model_params(clfdir).items():
# value is a tuple. First thing is to define how to store it
# in a pandas DataFrame. Use its string method without brackets
# (so that e.g., ['a', 'b'] will be stored as 'a,b' and
# ['abc'] will be stored as 'abc':
stored_value = ",".join(str(_) for _ in value)
if len(value) == 1:
try:
stored_value = int(value[0])
except (ValueError, TypeError):
try:
stored_value = float(value[0])
except (ValueError, TypeError):
pass
ret[key] = stored_value
ret[cols[0]] = log_loss(predicted_df)
ret[cols[1]] = roc_auc_score(predicted_df)
ret[cols[2]] = average_precision_score(predicted_df)
ret[cols[3]] = roc_curve(predicted_df)[-1]
ret[cols[4]] = precision_recall_curve(predicted_df)[-1]
return clfdir, testname, ret
except Exception as exc:
return clfdir, testname, exc