def class_loss(d):
r'''
Calculate discrete class loss for every instance in d.
The resulting array contains 1 for instance incorrectly classified,
0 otherwise:
>>> d = DataSet(X=helpers.to_one_of_n([0, 0, 0, 1, 1, 2]),
... Y=helpers.to_one_of_n([0, 0, 1, 0, 1, 2]))
>>> class_loss(d)
array([ 0., 0., 1., 1., 0., 0.])
'''
assert d.nfeatures == d.nclasses
return np.any(
helpers.hard_max(d.X) != helpers.hard_max(d.Y), axis=0).astype(float)
def auc(d):
'''
Calculate area under curve of the ROC for a dataset d. Expects the
predictions for a two-class problem.
'''
assert d.nclasses == 2 and d.nfeatures == 2
return stat.auc(np.diff(d.data, axis=0)[0], helpers.hard_max(d.labels)[1])
def ninstances_per_class(self):
if self.labels.shape[0] == 1 and self.labels.dtype == np.int:
return [len(np.flatnonzero(self.labels[0,:] == l)) for l in self.possible_labels]
elif self.labels.dtype == np.bool:
return np.sum(self.labels, axis=1).astype(int).tolist()
else:
return np.sum(helpers.hard_max(self.labels), axis=1).astype(int).tolist()
def class_loss(d):
r'''
Calculate discrete class loss for every instance in d.
The resulting array contains 1 for instance incorrectly classified,
0 otherwise:
>>> from psychic import *
>>> d = DataSet(data=helpers.to_one_of_n([0, 0, 0, 1, 1, 2]),
... labels=helpers.to_one_of_n([0, 0, 1, 0, 1, 2]))
>>> perf.class_loss(d)
array([ 0., 0., 1., 1., 0., 0.])
'''
assert d.nfeatures == d.nclasses
if d.labels.shape[0] == 1 and d.labels.dtype == np.int:
return (np.argmax(d.data, axis=0) != d.labels[0,:]).astype(np.float)
else:
return np.any(
helpers.hard_max(d.data) != helpers.hard_max(d.labels), axis=0).astype(float)
def ninstances_per_class(self): return np.sum(helpers.hard_max(self.Y), axis=1).astype(int).tolist()
def conf_mat(d):
'''
Make a confusion matrix. Rows contain the label, columns the predictions.
'''
return np.dot(helpers.hard_max(d.labels), helpers.hard_max(d.data).T)
