pytorch

1
syntax = "proto2";
2

3
package caffe2;
4

5
// Hierarchical Softmax protobuffer convention:
6
// The HSM operator requires a hierarchy of vocabulary words in the form of a
7
// tree from the user. This tree is expressed using the proto format.
8
// TreeProto points to the root NodeProto which can recursively contain children
9
// NodeProtos (internal nodes) or word_ids (leaf nodes).
10

11
// The aforementioned TreeProto is internally translated into a list of word_ids
12
// tagged with a list of NodeProtos that lie in the path from the root to that
13
// word_id using hsm_util.create_hierarchy(tree_proto).
14
// Specifically, HierarchyProto contains a list of PathProtos. Each PathProto
15
// belongs to a word_id and contains a list of PathNodeProtos. Each
16
// PathNodeProto contains information about the number of children the node has
17
// (length), the index of the child node that lies in the path from root to
18
// word_id (target) and a cumulative sum of children nodes (index; this acts as
19
// the weight parameter matrix offset).
20

21
// Each node in the hierarchy contains links to either leaf nodes or more
22
// non-terminal nodes
23
message NodeProto {
24
  // Links to non-terminal children nodes
25
  repeated NodeProto children = 1;
26
  // Links to terminal (leaf) nodes
27
  repeated int32 word_ids = 2;
28
  optional int32 offset = 3;
29
  optional string name = 4;
30
  repeated float scores = 5;
31
}
32

33
// Protobuf format to accept hierarchy for hierarchical softmax operator.
34
// TreeProto points to the root node.
35
message TreeProto {
36
  optional NodeProto root_node = 1;
37
}
38

39
// Internal Protobuf format which represents the path in the tree hierarchy for
40
// each word in the vocabulary.
41
message HierarchyProto {
42
  optional int32 size = 1;
43
  repeated PathProto paths = 2;
44
}
45

46
// Each PathProto belongs to a word and is an array of nodes in the
47
// path from the root to the leaf (which is the word itself) in the tree.
48
message PathProto {
49
  optional int32 word_id = 1;
50
  repeated PathNodeProto path_nodes = 2;
51
}
52

53
// Represents a node in the path from the root node all the way down to the
54
// word (leaf).
55
message PathNodeProto {
56
  // Parameter matrix offset for this node
57
  optional int32 index = 1;
58
  // Number of children
59
  optional int32 length = 2;
60
  // Index of the next node in the path
61
  optional int32 target = 3;
62
}
63