def balanced_log_loss(y_true, y_pred, method = 'edges', random_state=None):
"""Balanced log-loss metric (multi-class).
Keeps only a subset of the data instances corresponding to the rest class.
The size of the subset is equal to the median group size of the other
classes."""
# y_true = np.asarray(y_true)
# y_pred = np.asarray(y_pred)
classes, n_instances = np.unique(y_true, return_counts=True)
median_instances = np.median(n_instances[1:])
n_classes = classes.size
idx_rest = np.where(y_true == 0)[0] # Find rest instances
idx_else = np.where(y_true != 0)[0] # Find all other instances
if method == 'random':
if random_state is not None:
np.random.seed(random_state)
idx_keep = np.random.choice(idx_rest,median_instances, replace=False) # Keep a random subset
idx_final = np.sort(np.hstack((idx_keep, idx_else)))
if method == 'edges':
samples_per_rest_repetition = np.fix(median_instances / (2*n_classes - 1)).astype('int'); # How many we want to keep for each rest repetition
if samples_per_rest_repetition < 1:
samples_per_rest_repetition = 1;
changes = np.diff(y_true) # Stimulus change
idx_changes = np.nonzero(changes)[0] # Stimulus change
idx_from_rest = idx_changes[np.arange(start=0,stop=idx_changes.size,step=2)] # Changing from rest to movement
idx_to_rest = idx_changes[np.arange(start=1,stop=idx_changes.size,step=2)] # Changing from rest to movement
idx_to_rest = np.hstack(([0], idx_to_rest))
idx_keep = []
for ii,jj in zip(idx_to_rest,idx_from_rest):
center = np.fix(ii + (jj-ii)/2)
idx_keep.extend(np.arange(center,center+samples_per_rest_repetition))
idx_keep = np.asarray(idx_keep, dtype='int')
idx_final = np.sort(np.hstack((idx_keep, idx_else)))
true_new = y_true[idx_final]
pred_new = y_pred[idx_final]
return log_loss(true_new, pred_new)
评论列表
文章目录
