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
|
package com.yahoo.tensor.serialization;
import com.google.common.annotations.Beta;
import com.yahoo.io.GrowableByteBuffer;
import com.yahoo.tensor.DimensionSizes;
import com.yahoo.tensor.IndexedTensor;
import com.yahoo.tensor.Tensor;
import com.yahoo.tensor.TensorType;
import com.yahoo.text.Utf8;
import java.util.Iterator;
/**
* Implementation of a dense binary format for a tensor on the form:
*
* Sorted dimensions = num_dimensions [dimension_str_len dimension_str_bytes dimension_size_int]*
* Cell_values = [double, double, double, ...]*
* where values are encoded in order of increasing indexes in each dimension, increasing
* indexes of later dimensions in the dimension type before earlier.
*
* @author bratseth
*/
@Beta
public class DenseBinaryFormat implements BinaryFormat {
@Override
public void encode(GrowableByteBuffer buffer, Tensor tensor) {
if ( ! ( tensor instanceof IndexedTensor))
throw new RuntimeException("The dense format is only supported for indexed tensors");
encodeDimensions(buffer, (IndexedTensor)tensor);
encodeCells(buffer, tensor);
}
private void encodeDimensions(GrowableByteBuffer buffer, IndexedTensor tensor) {
buffer.putInt1_4Bytes(tensor.type().dimensions().size());
for (int i = 0; i < tensor.type().dimensions().size(); i++) {
encodeString(buffer, tensor.type().dimensions().get(i).name());
buffer.putInt1_4Bytes(tensor.dimensionSizes().size(i));
}
}
private void encodeCells(GrowableByteBuffer buffer, Tensor tensor) {
Iterator<Double> i = tensor.valueIterator();
if ( ! i.hasNext()) { // no values: Encode as NaN, as 0 dimensions may also mean 1 value
buffer.putDouble(Double.NaN);
}
else {
while (i.hasNext())
buffer.putDouble(i.next());
}
}
private void encodeString(GrowableByteBuffer buffer, String value) {
byte[] stringBytes = Utf8.toBytes(value);
buffer.putInt1_4Bytes(stringBytes.length);
buffer.put(stringBytes);
}
@Override
public Tensor decode(TensorType type, GrowableByteBuffer buffer) {
DimensionSizes sizes = decodeDimensionSizes(type, buffer);
Tensor.Builder builder = Tensor.Builder.of(type, sizes);
decodeCells(sizes, buffer, (IndexedTensor.BoundBuilder)builder);
return builder.build();
}
private DimensionSizes decodeDimensionSizes(TensorType type, GrowableByteBuffer buffer) {
int dimensionCount = buffer.getInt1_4Bytes();
if (type.dimensions().size() != dimensionCount)
throw new IllegalArgumentException("Type/instance mismatch: Instance has " + dimensionCount +
" dimensions but type is " + type);
DimensionSizes.Builder builder = new DimensionSizes.Builder(dimensionCount);
for (int i = 0; i < dimensionCount; i++) {
TensorType.Dimension expectedDimension = type.dimensions().get(i);
String encodedName = decodeString(buffer);
int encodedSize = buffer.getInt1_4Bytes();
if ( ! expectedDimension.name().equals(encodedName))
throw new IllegalArgumentException("Type/instance mismatch: Instance has '" + encodedName +
"' as dimension " + i + " but type is " + type);
if (expectedDimension.size().isPresent() && expectedDimension.size().get() < encodedSize)
throw new IllegalArgumentException("Type/instance mismatch: Instance has size " + encodedSize +
" in " + expectedDimension + " in type " + type);
builder.set(i, encodedSize);
}
return builder.build();
}
private void decodeCells(DimensionSizes sizes, GrowableByteBuffer buffer, IndexedTensor.BoundBuilder builder) {
for (int i = 0; i < sizes.totalSize(); i++)
builder.cellByDirectIndex(i, buffer.getDouble());
}
private String decodeString(GrowableByteBuffer buffer) {
int stringLength = buffer.getInt1_4Bytes();
byte[] stringBytes = new byte[stringLength];
buffer.get(stringBytes);
return Utf8.toString(stringBytes);
}
}
|
