1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
|
#pragma once
#include "graph/node_operators_unary.h"
namespace marian {
// Base class for a node that has more than one forward value tensor.
// For now we only do one additional value.
class TupleNode {
protected:
Tensor tupleVal_; // the additional forward value tensor
friend struct TupleViewNodeOp;
// Force implementation of these functions. The node that inherits from this
// should use them during allocation/deallocation.
virtual void allocateTuple() = 0;
virtual void freeTuple() = 0;
public:
// The specific node needs to implement what the view is actually looking at
virtual Expr tupleView() = 0;
};
// This is a view similar to ReshapeNodeOp or SliceViewNodeOp
// that uses the additional value as its tensor without
// allocating or destroying anthing. This has the purpose to
// create an Expression that can be put on the tape of the graph and
// visited in correct topological order. It should also make sure that
// the node that actually holds the memory persists via the reference
// to tupleNode_.
struct TupleViewNodeOp : public UnaryNodeOp {
private:
Expr tupleNode_; // hold a reference to the actually node with the tuple data
public:
TupleViewNodeOp(Expr origin, Shape shape, Type type)
: UnaryNodeOp(origin, shape, type),
tupleNode_(origin) {
Node::destroy_ = false; // should not be detroyed or freed, the origin node is handling that
Node::trainable_ = false; // for now this is not trainable
}
// make sure these functions don't actually do anything, origin node handles all this
~TupleViewNodeOp() {}
void allocate() override {}
void free() override {}
void forward() override {}
void backward() override {}
void init_dependent() override {}
void set_zero_adjoint() override {}
// Return the additional tuple tensor as the val() tensor for the view
Tensor& val() override {
// we have to use a raw pointer cast here as we cannot add IntrusivePtr ref-counting to TupleNode
// otherwise we would have ambigous inheritence
auto tptr = dynamic_cast<TupleNode*>(tupleNode_.get());
ABORT_IF(!tptr, "Could not convert to tuple?");
return tptr->tupleVal_;
};
// Currently not trainable. We will see if that will be useful at some point
Tensor& grad() override {
ABORT("There should be no gradients for tuple values");
};
const std::string type() override { return "tupleView"; }
const std::string color() override { return "grey"; }
};
// This is an implementation of topk, similar to the PyTorch node.
// At the moment we only handle axis=-1 in here, but do transposes
// in the actual operator to handle other axes (inefficiently).
// The normal forward values here are the top-k values per axis,
// the additional value from the TupleNode contains the integer
// indices of the top-k values.
struct TopKNodeOp : public UnaryNodeOp,
public TupleNode {
private:
int k_; // how many top-k results?
int axis_; // on which axis
bool descending_; // sort-order, by default descending. PyTorch has a version without sorting, we always sort.
public:
TopKNodeOp(Expr a, int k, int axis, bool descending = true)
: UnaryNodeOp(a, newShape(a, k, axis)),
k_{k}, descending_{descending} {}
Shape newShape(Expr a, int k, int axis) {
Shape shape = a->shape();
axis_ = shape.axis(axis);
shape.set(axis_, k);
return shape;
}
// imlementation of TupleNode-specific pure-virtual functions for allocation
void allocateTuple() override final {
graph()->getTensorAllocator()->allocate(tupleVal_, shape(), Type::uint32);
}
// we override the normal allocation to include the TupleNode allocation
void allocate() override {
UnaryNodeOp::allocate();
allocateTuple();
}
// imlementation of TupleNode-specific pure-virtual functions for de-allocation
void freeTuple() override final {
if(graph()) {
if(tupleVal_) {
graph()->free(tupleVal_);
tupleVal_ = nullptr;
}
}
}
// we override the normal allocation to include the TupleNode de-allocation
void free() override {
UnaryNodeOp::free();
freeTuple();
}
// Create and return a TupleView to the additional forward value
virtual Expr tupleView() override final {
return Expression<TupleViewNodeOp>(this, shape(), Type::uint32);
}
void forward() override {
TopK(/*out*/val_, /*out: topkIndices=*/tupleVal_,
graph()->allocator(),
child(0)->val(), k_, axis_, descending_);
}
void backward() override {
Insert</*add=*/true>(/*out*/child(0)->grad(), adj_, val_, axis_);
}
const std::string type() override { return "topk"; }
virtual size_t hash() override {
if(!hash_) {
hash_ = NaryNodeOp::hash();
util::hash_combine(hash_, k_);
util::hash_combine(hash_, axis_);
util::hash_combine(hash_, descending_);
}
return hash_;
}
virtual bool equal(Expr node) override {
if(!NaryNodeOp::equal(node))
return false;
auto cnode = std::dynamic_pointer_cast<TopKNodeOp>(node);
if(!cnode)
return false;
if(k_ != cnode->k_)
return false;
if(axis_ != cnode->axis_)
return false;
if(descending_ != cnode->descending_)
return false;
return true;
}
};
}
|
