def key_abstype(self):
if hasattr(self.result, 'store'):
# have a store, can look stuff up
feat_name = 'isstructured'
ds = self.result.metadata['dataset']
cs = self.result.metadata['class_space']
fm = self.result.store.get_FeatureMap(ds, feat_name)
feat_space = self.result.store.get_Space(feat_name)
# Extract goldstandard and overall classification matrix
# TODO: Check that CL has full coverage w/o overlap
gs = self.result.store.get_ClassMap(ds, cs).raw
cl = np.zeros(gs.shape, dtype=self.result.results[0].classifications.dtype)
for r in self.result.results:
cl[r.instance_indices] = r.classifications
retval = {}
for i, cat in enumerate(feat_space):
cat_mask = fm[:,i].nonzero()[0]
cat_res = Result(gs[cat_mask], cl[cat_mask], cat_mask)
cm = cat_res.confusion_matrix(self.interpreter)
prf = CombinedMicroAverage(cm[None,:])
retval[cat] = {'precision':prf.precision, 'recall':prf.recall, 'fscore':prf.fscore}
return retval
else:
return {}
def key_breakdown(self):
if hasattr(self.result, 'store'):
store = self.result.store
# have a store, can look stuff up
ds = self.result.metadata['dataset']
cs = self.result.metadata['class_space']
try:
class_space = store.get_Space(cs)
space = store.get_Space(self.space_name)
space_md = store.get_SpaceMetadata(self.space_name)
except NoData:
# TODO: handle NoData here with an error message
raise
# get the actual space and check it is suitable
if space_md['type'] == 'feature':
fm = store.get_FeatureMap(ds, self.space_name).raw.toarray()
if fm.min() < 0 or fm.max() > 1:
raise TypeError('feature space has non 0/1 elements')
mat = fm.astype(bool)
elif space_md['type'] == 'class':
mat = store.get_ClassMap(ds, self.space_name).raw
else:
raise TypeError('space {0} has unsuitable type: {1}'.format(self.space_name, space_md['type']))
# Extract goldstandard and overall classification matrix
# TODO: Check that CL has full coverage w/o overlap
gs = self.result.store.get_ClassMap(ds, cs).raw
cl = np.zeros(gs.shape, dtype=self.result.results[0].classifications.dtype)
for r in self.result.results:
cl[r.instance_indices] = r.classifications
retval = {}
for i, cat in enumerate(space):
cat_mask = mat[:,i].nonzero()[0]
cat_res = Result(gs[cat_mask], cl[cat_mask], cat_mask)
cm = cat_res.confusion_matrix(self.interpreter)
prf = self.cmm(cm.sum(0))
breakdown = dict(zip(class_space, map(self.cmm,cm) ))
retval[cat] = (prf, breakdown)
return retval
else:
return {}
