def _create_sub_keys(
spec: config.MetricsSpec) -> Optional[List[metric_types.SubKey]]:
"""Creates subkeys associated with spec."""
sub_keys = None
if spec.HasField('binarize'):
sub_keys = []
if spec.binarize.class_ids.values:
for v in spec.binarize.class_ids.values:
sub_keys.append(metric_types.SubKey(class_id=v))
if spec.binarize.k_list.values:
for v in spec.binarize.k_list.values:
sub_keys.append(metric_types.SubKey(k=v))
if spec.binarize.top_k_list.values:
for v in spec.binarize.top_k_list.values:
sub_keys.append(metric_types.SubKey(top_k=v))
if spec.aggregate.micro_average:
# Micro averaging is performed by flattening the labels and predictions
# and treating them as independent pairs. This is done by default by most
# metrics whenever binarization is not used. If micro-averaging and
# binarization are used, then we need to create an empty subkey to ensure
# the overall aggregate key is still computed. Note that the class_weights
# should always be passed to all metric calculations to ensure they are
# taken into account when flattening is required.
sub_keys.append(None)
return sub_keys
def _create_sub_keys(
spec: config.MetricsSpec) -> Optional[List[metric_types.SubKey]]:
"""Creates subkeys associated with spec."""
sub_keys = None
if spec.HasField('binarize'):
sub_keys = []
if spec.binarize.class_ids:
for v in spec.binarize.class_ids:
sub_keys.append(metric_types.SubKey(class_id=v))
if spec.binarize.k_list:
for v in spec.binarize.k_list:
sub_keys.append(metric_types.SubKey(k=v))
if spec.binarize.top_k_list:
for v in spec.binarize.top_k_list:
sub_keys.append(metric_types.SubKey(top_k=v))
return sub_keys