import caffe2.python.hsm_util as hsmu
# User inputs tree using protobuf file or, in this case, python utils
# The hierarchy in this test looks as shown below. Note that the final subtrees
# (with word_ids as leaves) have been collapsed for visualization
# *
# / \
# * 5,6,7,8
# / \
# 0,1,2 3,4
tree = hsm_pb2.TreeProto()
words = [[0, 1, 2], [3, 4], [5, 6, 7, 8]]
node1 = hsmu.create_node_with_words(words[0])
node2 = hsmu.create_node_with_words(words[1])
node3 = hsmu.create_node_with_words(words[2])
node4 = hsmu.create_node_with_nodes([node1, node2])
node = hsmu.create_node_with_nodes([node4, node3])
tree.root_node.MergeFrom(node)
# Internal util to translate input tree to list of (word_id,path). serialized
# hierarchy is passed into the operator_def as a string argument,
hierarchy_proto = hsmu.create_hierarchy(tree)
arg = caffe2_pb2.Argument()
arg.name = "hierarchy"
arg.s = hierarchy_proto.SerializeToString()
class TestHsm(hu.HypothesisTestCase):
def test_hsm_run_once(self):
workspace.GlobalInit(['caffe2'])
workspace.FeedBlob("data",
import caffe2.python.hsm_util as hsmu
# User inputs tree using protobuf file or, in this case, python utils
# The hierarchy in this test looks as shown below. Note that the final subtrees
# (with word_ids as leaves) have been collapsed for visualization
# *
# / \
# * 5,6,7,8
# / \
# 0,1,2 3,4
tree = hsm_pb2.TreeProto()
words = [[0, 1, 2], [3, 4], [5, 6, 7, 8]]
node1 = hsmu.create_node_with_words(words[0], "node1")
node2 = hsmu.create_node_with_words(words[1], "node2")
node3 = hsmu.create_node_with_words(words[2], "node3")
node4 = hsmu.create_node_with_nodes([node1, node2], "node4")
node = hsmu.create_node_with_nodes([node4, node3], "node5")
tree.root_node.MergeFrom(node)
# structure:
# node5: [0, 2, ["node4", "node3"]] # offset, length, "node4, node3"
# node4: [2, 2, ["node1", "node2"]]
# node1: [4, 3, [0, 1 ,2]]
# node2: [7, 2, [3, 4]
# node3: [9, 4, [5, 6, 7, 8]
struct = [[0, 2, ["node4", "node3"], "node5"],
[2, 2, ["node1", "node2"], "node4"],
[4, 3, [0, 1, 2], "node1"],
[7, 2, [3, 4], "node2"],
[9, 4, [5, 6, 7, 8], "node3"]]
import caffe2.python.hsm_util as hsmu
# User inputs tree using protobuf file or, in this case, python utils
# The hierarchy in this test looks as shown below. Note that the final subtrees
# (with word_ids as leaves) have been collapsed for visualization
# *
# / \
# * 5,6,7,8
# / \
# 0,1,2 3,4
tree = hsm_pb2.TreeProto()
words = [[0, 1, 2], [3, 4], [5, 6, 7, 8]]
node1 = hsmu.create_node_with_words(words[0], "node1")
node2 = hsmu.create_node_with_words(words[1], "node2")
node3 = hsmu.create_node_with_words(words[2], "node3")
node4 = hsmu.create_node_with_nodes([node1, node2], "node4")
node = hsmu.create_node_with_nodes([node4, node3], "node5")
tree.root_node.MergeFrom(node)
# structure:
# node5: [0, 2, ["node4", "node3"]] # offset, length, "node4, node3"
# node4: [2, 2, ["node1", "node2"]]
# node1: [4, 3, [0, 1 ,2]]
# node2: [7, 2, [3, 4]
# node3: [9, 4, [5, 6, 7, 8]
struct = [[0, 2, ["node4", "node3"], "node5"],
[2, 2, ["node1", "node2"], "node4"],
[4, 3, [0, 1, 2], "node1"],
[7, 2, [3, 4], "node2"],
[9, 4, [5, 6, 7, 8], "node3"]]