def init(self,
dataset,
criterion,
sample_nums,
path_graph,
path_wnids,
classes,
Rules,
tree_supervision_weight=1.):
"""
Extra init method makes clear which arguments are finally necessary for
this class to function. The constructor for this class may generate
some of these required arguments if initially missing.
"""
self.dataset = dataset
self.num_classes = len(classes)
self.nodes = Node.get_nodes(path_graph, path_wnids, classes)
self.rules = Rules(dataset, path_graph, path_wnids, classes)
self.tree_supervision_weight = tree_supervision_weight
self.criterion = criterion
self.sample_nums = np.array(sample_nums)
self.node_depths = defaultdict(lambda: [])
self.node_weights = defaultdict(lambda: [])
effective_num = 1.0 - np.power(0.999, self.sample_nums)
weights = (1.0 - 0.999) / np.array(effective_num)
self.weights = weights
for node in self.nodes:
key = node.num_classes
depth = node.get_depth()
self.node_depths[key].append(depth)
node_weight = []
for new_label in range(node.num_classes):
node_weight.append(weights[node.new_to_old_classes[new_label]])
self.node_weights[key].append(node_weight)
def __init__(self, dataset, path_graph=None, path_wnids=None, classes=()):
if not path_graph:
path_graph = dataset_to_default_path_graph(dataset)
if not path_wnids:
path_wnids = dataset_to_default_path_wnids(dataset)
if not classes:
classes = dataset_to_dummy_classes(dataset)
super().__init__()
assert all([dataset, path_graph, path_wnids, classes])
self.classes = classes
self.nodes = Node.get_nodes(path_graph, path_wnids, classes)
self.G = self.nodes[0].G
self.wnid_to_node = {node.wnid: node for node in self.nodes}
self.wnids = get_wnids(path_wnids)
self.wnid_to_class = {
wnid: cls
for wnid, cls in zip(self.wnids, self.classes)
}
self.correct = 0
self.total = 0
self.I = torch.eye(len(classes))
def init(self,
dataset,
criterion,
path_graph,
path_wnids,
classes,
Rules,
tree_supervision_weight=1.):
"""
Extra init method makes clear which arguments are finally necessary for
this class to function. The constructor for this class may generate
some of these required arguments if initially missing.
"""
self.dataset = dataset
self.num_classes = len(classes)
self.nodes = Node.get_nodes(path_graph, path_wnids, classes)
self.rules = Rules(dataset, path_graph, path_wnids, classes)
self.tree_supervision_weight = tree_supervision_weight
self.criterion = criterion