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
168
169
170
171
172
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package ai.vespa.embedding;
import ai.vespa.modelintegration.evaluator.OnnxEvaluator;
import ai.vespa.modelintegration.evaluator.OnnxEvaluatorOptions;
import ai.vespa.modelintegration.evaluator.OnnxRuntime;
import com.yahoo.component.AbstractComponent;
import com.yahoo.component.annotation.Inject;
import com.yahoo.embedding.BertBaseEmbedderConfig;
import com.yahoo.language.process.Embedder;
import com.yahoo.language.wordpiece.WordPieceEmbedder;
import com.yahoo.tensor.IndexedTensor;
import com.yahoo.tensor.Tensor;
import com.yahoo.tensor.TensorType;
import java.util.ArrayList;
import java.util.List;
import java.util.Map;
/**
* A BERT Base compatible embedder. This embedder uses a WordPiece embedder to
* produce a token sequence that is then input to a transformer model. A BERT base
* compatible transformer model must have three inputs:
*
* - A token sequence (input_ids)
* - An attention mask (attention_mask)
* - Token types for cross encoding (token_type_ids)
*
* See bert-base-embedder.def for configurable parameters.
*
* @author lesters
*/
public class BertBaseEmbedder extends AbstractComponent implements Embedder {
private final int maxTokens;
private final int startSequenceToken;
private final int endSequenceToken;
private final String inputIdsName;
private final String attentionMaskName;
private final String tokenTypeIdsName;
private final String outputName;
private final PoolingStrategy poolingStrategy;
private final Embedder.Runtime runtime;
private final WordPieceEmbedder tokenizer;
private final OnnxEvaluator evaluator;
@Inject
public BertBaseEmbedder(OnnxRuntime onnx, Embedder.Runtime runtime, BertBaseEmbedderConfig config) {
this.runtime = runtime;
maxTokens = config.transformerMaxTokens();
startSequenceToken = config.transformerStartSequenceToken();
endSequenceToken = config.transformerEndSequenceToken();
inputIdsName = config.transformerInputIds();
attentionMaskName = config.transformerAttentionMask();
tokenTypeIdsName = config.transformerTokenTypeIds();
outputName = config.transformerOutput();
poolingStrategy = PoolingStrategy.fromString(config.poolingStrategy().toString());
OnnxEvaluatorOptions options = new OnnxEvaluatorOptions();
options.setExecutionMode(config.onnxExecutionMode().toString());
options.setThreads(config.onnxInterOpThreads(), config.onnxIntraOpThreads());
if (config.onnxGpuDevice() >= 0) options.setGpuDevice(config.onnxGpuDevice());
tokenizer = new WordPieceEmbedder.Builder(config.tokenizerVocab().toString()).build();
this.evaluator = onnx.evaluatorOf(config.transformerModel().toString(), options);
validateModel();
}
private void validateModel() {
Map<String, TensorType> inputs = evaluator.getInputInfo();
validateName(inputs, inputIdsName, "input");
validateName(inputs, attentionMaskName, "input");
// some BERT inspired models such as DistilBERT do not have token_type_ids input
// one can explicitly declare this is such model by setting that config to empty string
if (!"".equals(tokenTypeIdsName)) {
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 text, Context context) {
var start = System.nanoTime();
var tokens = tokenize(text, context);
runtime.sampleSequenceLength(tokens.size(), context);
runtime.sampleEmbeddingLatency((System.nanoTime() - start)/1_000_000d, context);
return tokens;
}
@Override
public Tensor embed(String text, Context context, TensorType type) {
var start = System.nanoTime();
if (type.dimensions().size() != 1) {
throw new IllegalArgumentException("Error in embedding to type '" + type + "': should only have one dimension.");
}
if (!type.dimensions().get(0).isIndexed()) {
throw new IllegalArgumentException("Error in embedding to type '" + type + "': dimension should be indexed.");
}
List<Integer> tokens = embedWithSeparatorTokens(text, context, maxTokens);
runtime.sampleSequenceLength(tokens.size(), context);
var embedding = embedTokens(tokens, type);
runtime.sampleEmbeddingLatency((System.nanoTime() - start)/1_000_000d, context);
return embedding;
}
@Override public void deconstruct() { evaluator.close(); }
private List<Integer> tokenize(String text, Context ctx) { return tokenizer.embed(text, ctx); }
Tensor embedTokens(List<Integer> tokens, TensorType type) {
Tensor inputSequence = createTensorRepresentation(tokens, "d1");
Tensor attentionMask = createAttentionMask(inputSequence);
Tensor tokenTypeIds = createTokenTypeIds(inputSequence);
Map<String, Tensor> inputs;
if (!"".equals(tokenTypeIdsName)) {
inputs = Map.of(inputIdsName, inputSequence.expand("d0"),
attentionMaskName, attentionMask.expand("d0"),
tokenTypeIdsName, tokenTypeIds.expand("d0"));
} else {
inputs = Map.of(inputIdsName, inputSequence.expand("d0"),
attentionMaskName, attentionMask.expand("d0"));
}
Map<String, Tensor> outputs = evaluator.evaluate(inputs);
Tensor tokenEmbeddings = outputs.get(outputName);
return poolingStrategy.toSentenceEmbedding(type, tokenEmbeddings, attentionMask);
}
private List<Integer> embedWithSeparatorTokens(String text, Context context, int maxLength) {
List<Integer> tokens = new ArrayList<>();
tokens.add(startSequenceToken);
tokens.addAll(tokenize(text, context));
tokens.add(endSequenceToken);
if (tokens.size() > maxLength) {
tokens = tokens.subList(0, maxLength-1);
tokens.add(endSequenceToken);
}
return tokens;
}
private IndexedTensor createTensorRepresentation(List<Integer> 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();
}
private static Tensor createAttentionMask(Tensor d) {
return d.map((x) -> x > 0 ? 1:0);
}
private static Tensor createTokenTypeIds(Tensor d) {
return d.map((x) -> 0); // Assume only one token type
}
}
|
