Function create_open_split_byclass

lib/utils/split.py:75–109

split whole dataset into closed set and open set (for testing only), class by class Args: frac (float): fraction of number of open classes Returns:

(df, open_frac, label_name='Y', label_weight_name='Y_Weight', lt_label_name='LT_Class', num_class=10, label_type='classification', scale=None)

	return cls_label_lst.index(lbl)

def create_open_split_byclass(df, open_frac, label_name='Y', label_weight_name='Y_Weight', lt_label_name='LT_Class', num_class=10, label_type='classification', scale=None):
	"""split whole dataset into closed set and open set (for testing only), class by class

	Args:
		frac (float): fraction of number of open classes

	Returns:
	"""
	cls_labels, bins, hist = get_cls_labels(df=df, label_name=label_name, num_class=num_class, label_type=label_type, scale=scale)
	num_open_classes = int(np.ceil(len(cls_labels) * open_frac))
	# open_cls_indices = np.argpartition(hist, num_open_classes)[:num_open_classes]
	open_cls_indices = cls_labels[-num_open_classes:]
	print('Label distribution: ', hist)
	print('Open classes: ', open_cls_indices)

	df_cpy = copy.deepcopy(df)
	for i, open_cls_idx in enumerate(open_cls_indices):
		df_cpy[lt_label_name] = 'open{}'.format(i)

	if label_type == "regression":
		label_name = "reg_cls"
	open_df = df_cpy[df_cpy[label_name].isin(open_cls_indices)]

	# regard all open classes as one unknown class, as in https://arxiv.org/pdf/1904.05160.pdf
	open_df = get_label_weight(open_df, label_weight_name)

	closed_df = df[~df.index.isin(open_df.index)]

	# relabel indices of the closed set for training/validation, to avoid index mismatch when calculating losses such as cross entropy
	closed_df[label_name] = closed_df[label_name].apply(lambda x: relabel(x, cls_labels))

	# relabel indices of the open set as one outlier class, thus formulating testing as anomalu detection 
	open_df[label_name] = len(cls_labels) - num_open_classes
	open_df[lt_label_name] = 'open'
	return closed_df, open_df


def create_fold_byclass(df, seed, frac, lt_frac, label_name='Y', label_weight_name='Y_Weight', lt_label_name='LT_Class', num_class=10, label_type='classification', scale=None):