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
|
package ai.vespa.embedding.huggingface;
import ai.vespa.embedding.PoolingStrategy;
import ai.vespa.modelintegration.evaluator.OnnxEvaluator;
import ai.vespa.modelintegration.evaluator.OnnxEvaluatorOptions;
import ai.vespa.modelintegration.evaluator.OnnxRuntime;
import com.yahoo.api.annotations.Beta;
import com.yahoo.component.AbstractComponent;
import com.yahoo.component.annotation.Inject;
import com.yahoo.embedding.huggingface.HuggingFaceEmbedderConfig;
import com.yahoo.language.huggingface.HuggingFaceTokenizer;
import com.yahoo.language.process.Embedder;
import com.yahoo.tensor.IndexedTensor;
import com.yahoo.tensor.Tensor;
import com.yahoo.tensor.TensorAddress;
import com.yahoo.tensor.TensorType;
import java.nio.file.Paths;
import java.util.List;
import java.util.Map;
import java.util.logging.Logger;
import static com.yahoo.language.huggingface.ModelInfo.TruncationStrategy.LONGEST_FIRST;
@Beta
public class HuggingFaceEmbedder extends AbstractComponent implements Embedder {
private static final Logger log = Logger.getLogger(HuggingFaceEmbedder.class.getName());
private final String inputIdsName;
private final String attentionMaskName;
private final String tokenTypeIdsName;
private final String outputName;
private final boolean normalize;
private final HuggingFaceTokenizer tokenizer;
private final OnnxEvaluator evaluator;
private final PoolingStrategy poolingStrategy;
@Inject
public HuggingFaceEmbedder(OnnxRuntime onnx, HuggingFaceEmbedderConfig config) {
inputIdsName = config.transformerInputIds();
attentionMaskName = config.transformerAttentionMask();
tokenTypeIdsName = config.transformerTokenTypeIds();
outputName = config.transformerOutput();
normalize = config.normalize();
var tokenizerPath = Paths.get(config.tokenizerPath().toString());
var builder = new HuggingFaceTokenizer.Builder()
.addSpecialTokens(true)
.addDefaultModel(tokenizerPath)
.setPadding(false);
var info = HuggingFaceTokenizer.getModelInfo(tokenizerPath);
log.fine(() -> "'%s' has info '%s'".formatted(tokenizerPath, info));
if (info.maxLength() == -1 || info.truncation() != LONGEST_FIRST) {
// Force truncation to max token vector length accepted by model if tokenizer.json contains no valid truncation configuration
int maxLength = info.maxLength() > 0 && info.maxLength() <= config.transformerMaxTokens()
? info.maxLength()
: config.transformerMaxTokens();
builder.setTruncation(true).setMaxLength(maxLength);
}
this.tokenizer = builder.build();
poolingStrategy = PoolingStrategy.fromString(config.poolingStrategy().toString());
var onnxOpts = new OnnxEvaluatorOptions();
if (config.transformerGpuDevice() >= 0)
onnxOpts.setGpuDevice(config.transformerGpuDevice());
onnxOpts.setExecutionMode(config.transformerExecutionMode().toString());
onnxOpts.setThreads(config.transformerInterOpThreads(), config.transformerIntraOpThreads());
evaluator = onnx.evaluatorOf(config.transformerModel().toString(), onnxOpts);
validateModel();
}
public void validateModel() {
Map<String, TensorType> inputs = evaluator.getInputInfo();
validateName(inputs, inputIdsName, "input");
validateName(inputs, attentionMaskName, "input");
if (!tokenTypeIdsName.isEmpty()) validateName(inputs, tokenTypeIdsName, "input");
Map<String, TensorType> outputs = evaluator.getOutputInfo();
validateName(outputs, outputName, "output");
}
private void validateName(Map<String, TensorType> types, String name, String type) {
if ( ! types.containsKey(name)) {
throw new IllegalArgumentException("Model does not contain required " + type + ": '" + name + "'. " +
"Model contains: " + String.join(",", types.keySet()));
}
}
@Override
public List<Integer> embed(String s, Context context) {
return tokenizer.embed(s, context);
}
@Override
public void deconstruct() {
evaluator.close();
tokenizer.close();
}
@Override
public Tensor embed(String s, Context context, TensorType tensorType) {
var encoding = tokenizer.encode(s, context.getLanguage());
Tensor inputSequence = createTensorRepresentation(encoding.ids(), "d1");
Tensor attentionMask = createTensorRepresentation(encoding.attentionMask(), "d1");
Tensor tokenTypeIds = tokenTypeIdsName.isEmpty() ? null : createTensorRepresentation(encoding.typeIds(), "d1");
Map<String, Tensor> inputs;
if (tokenTypeIdsName.isEmpty() || tokenTypeIds.isEmpty()) {
inputs = Map.of(inputIdsName, inputSequence.expand("d0"),
attentionMaskName, attentionMask.expand("d0"));
} else {
inputs = Map.of(inputIdsName, inputSequence.expand("d0"),
attentionMaskName, attentionMask.expand("d0"),
tokenTypeIdsName, tokenTypeIds.expand("d0"));
}
Map<String, Tensor> outputs = evaluator.evaluate(inputs);
Tensor tokenEmbeddings = outputs.get(outputName);
var result = poolingStrategy.toSentenceEmbedding(tensorType, tokenEmbeddings, attentionMask);
return normalize ? normalize(result, tensorType) : result;
}
Tensor normalize(Tensor embedding, TensorType tensorType) {
double sumOfSquares = 0.0;
Tensor.Builder builder = Tensor.Builder.of(tensorType);
for (int i = 0; i < tensorType.dimensions().get(0).size().get(); i++) {
double item = embedding.get(TensorAddress.of(i));
sumOfSquares += item * item;
}
double magnitude = Math.sqrt(sumOfSquares);
for (int i = 0; i < tensorType.dimensions().get(0).size().get(); i++) {
double value = embedding.get(TensorAddress.of(i));
builder.cell(value / magnitude, i);
}
return builder.build();
}
private IndexedTensor createTensorRepresentation(List<Long> input, String dimension) {
int size = input.size();
TensorType type = new TensorType.Builder(TensorType.Value.FLOAT).indexed(dimension, size).build();
IndexedTensor.Builder builder = IndexedTensor.Builder.of(type);
for (int i = 0; i < size; ++i) {
builder.cell(input.get(i), i);
}
return builder.build();
}
}
|
