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
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package com.yahoo.tensor.functions;
import com.yahoo.tensor.Tensor;
import com.yahoo.tensor.TensorAddress;
import com.yahoo.tensor.TensorType;
import com.yahoo.tensor.TypeResolver;
import com.yahoo.tensor.evaluation.EvaluationContext;
import com.yahoo.tensor.evaluation.Name;
import com.yahoo.tensor.evaluation.TypeContext;
import java.util.HashMap;
import java.util.Iterator;
import java.util.List;
import java.util.List;
import java.util.Map;
import java.util.Objects;
/**
* The <i>map_subspaces</i> tensor function transforms each dense subspace in a (mixed) tensor
*
* @author arnej
*/
public class MapSubspaces<NAMETYPE extends Name> extends PrimitiveTensorFunction<NAMETYPE> {
private final TensorFunction<NAMETYPE> argument;
private final DenseSubspaceFunction<NAMETYPE> function;
private MapSubspaces(TensorFunction<NAMETYPE> argument, DenseSubspaceFunction<NAMETYPE> function) {
this.argument = argument;
this.function = function;
}
public MapSubspaces(TensorFunction<NAMETYPE> argument, String functionArg, TensorFunction<NAMETYPE> function) {
this(argument, new DenseSubspaceFunction<>(functionArg, function));
Objects.requireNonNull(argument, "The argument cannot be null");
Objects.requireNonNull(functionArg, "The functionArg cannot be null");
Objects.requireNonNull(function, "The function cannot be null");
}
private TensorType outputType(TensorType inputType) {
var m = inputType.mappedSubtype();
var d = function.outputType(inputType.indexedSubtype());
if (m.rank() == 0) {
return d;
}
if (d.rank() == 0) {
return TypeResolver.map(m); // decay cell type
}
TensorType.Value cellType = d.valueType();
Map<String, TensorType.Dimension> dims = new HashMap<>();
for (var dim : m.dimensions()) {
dims.put(dim.name(), dim);
}
for (var dim : d.dimensions()) {
var old = dims.put(dim.name(), dim);
if (old != null) {
throw new IllegalArgumentException("dimension name collision in map_subspaces: " + m + " vs " + d);
}
}
return new TensorType(cellType, dims.values());
}
public TensorFunction<NAMETYPE> argument() { return argument; }
@Override
public List<TensorFunction<NAMETYPE>> arguments() { return List.of(argument); }
@Override
public TensorFunction<NAMETYPE> withArguments(List<TensorFunction<NAMETYPE>> arguments) {
if ( arguments.size() != 1)
throw new IllegalArgumentException("MapSubspaces must have 1 argument, got " + arguments.size());
return new MapSubspaces<NAMETYPE>(arguments.get(0), function);
}
@Override
public PrimitiveTensorFunction<NAMETYPE> toPrimitive() {
return new MapSubspaces<>(argument.toPrimitive(), function);
}
@Override
public TensorType type(TypeContext<NAMETYPE> context) {
return outputType(argument.type(context));
}
@Override
public Tensor evaluate(EvaluationContext<NAMETYPE> context) {
Tensor input = argument().evaluate(context);
TensorType inputType = input.type();
TensorType inputTypeMapped = inputType.mappedSubtype();
TensorType inputTypeDense = inputType.indexedSubtype();
Map<TensorAddress, Tensor.Builder> builders = new HashMap<>();
for (Iterator<Tensor.Cell> iter = input.cellIterator(); iter.hasNext(); ) {
var cell = iter.next();
var fullAddr = cell.getKey();
var mapAddrBuilder = new TensorAddress.Builder(inputTypeMapped);
var idxAddrBuilder = new TensorAddress.Builder(inputTypeDense);
for (int i = 0; i < inputType.dimensions().size(); i++) {
var dim = inputType.dimensions().get(i);
if (dim.isMapped()) {
mapAddrBuilder.add(dim.name(), fullAddr.numericLabel(i));
} else {
idxAddrBuilder.add(dim.name(), fullAddr.numericLabel(i));
}
}
var mapAddr = mapAddrBuilder.build();
var builder = builders.computeIfAbsent(mapAddr, k -> Tensor.Builder.of(inputTypeDense));
var idxAddr = idxAddrBuilder.build();
builder.cell(idxAddr, cell.getValue());
}
TensorType outputType = outputType(input.type());
TensorType denseOutputType = outputType.indexedSubtype();
var denseOutputDims = denseOutputType.dimensions();
Tensor.Builder builder = Tensor.Builder.of(outputType);
for (var entry : builders.entrySet()) {
TensorAddress mappedAddr = entry.getKey();
Tensor denseInput = entry.getValue().build();
Tensor denseOutput = function.map(denseInput);
// XXX check denseOutput.type().dimensions()
for (Iterator<Tensor.Cell> iter = denseOutput.cellIterator(); iter.hasNext(); ) {
var cell = iter.next();
var denseAddr = cell.getKey();
var addrBuilder = new TensorAddress.Builder(outputType);
for (int i = 0; i < inputTypeMapped.dimensions().size(); i++) {
var dim = inputTypeMapped.dimensions().get(i);
addrBuilder.add(dim.name(), mappedAddr.numericLabel(i));
}
for (int i = 0; i < denseOutputDims.size(); i++) {
var dim = denseOutputDims.get(i);
addrBuilder.add(dim.name(), denseAddr.numericLabel(i));
}
builder.cell(addrBuilder.build(), cell.getValue());
}
}
return builder.build();
}
@Override
public String toString(ToStringContext<NAMETYPE> context) {
return "map_subspaces(" + argument.toString(context) + ", " + function + ")";
}
@Override
public int hashCode() { return Objects.hash("map_subspaces", argument, function); }
}
|
