1 /*
2 * Copyright (C) 2018 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include <memory>
18 #include <vector>
19
20 #include "utils/base/logging.h"
21 #include "utils/sentencepiece/double_array_trie.h"
22 #include "utils/sentencepiece/encoder.h"
23 #include "utils/sentencepiece/normalizer.h"
24 #include "utils/sentencepiece/sorted_strings_table.h"
25 #include "utils/strings/stringpiece.h"
26 #include "utils/tflite/encoder_common.h"
27 #include "utils/tflite/text_encoder.h"
28 #include "utils/tflite/text_encoder_config_generated.h"
29 #include "flatbuffers/flatbuffers.h"
30 #include "flatbuffers/flexbuffers.h"
31 #include "tensorflow/lite/kernels/kernel_util.h"
32 #include "tensorflow/lite/model.h"
33 #include "tensorflow/lite/string_util.h"
34
35 namespace libtextclassifier3 {
36 namespace {
37
38 struct TextEncoderOp {
39 std::unique_ptr<SentencePieceNormalizer> normalizer;
40 std::unique_ptr<Encoder> encoder;
41 std::unique_ptr<SentencePieceMatcher> matcher;
42 };
43
44 // Input parameters for the op.
45 // The conversation message as a (1, conversation length) string tensor.
46 constexpr const int kInputTexts = 0;
47
48 // The number of messages, the conversation length, int scalar.
49 constexpr const int kInputNumInputs = 1;
50
51 // Maximum output length of the encoding, int scalar.
52 constexpr const int kInputMaxLength = 2;
53
54 // Additional attributes to align to the sentence pieces, e.g. user ids per
55 // message.
56 constexpr const int kInputAttr = 3;
57
58 // Output parameters for the op.
59 // The text sentence piece encodings as ids, (1, max output length) int tensor.
60 constexpr const int kOutputEncoded = 0;
61
62 // Relative position of each sentence piece in the input text,
63 // (1, max output length) int tensor.
64 constexpr const int kOutputPosition = 1;
65
66 // Output length after trimming to the maximum output length specified.
67 // int scalar.
68 constexpr const int kOutputLengths = 2;
69
70 // Padded and sentence piece aligned provided attributes, e.g. user id per
71 // sentence piece.
72 constexpr const int kOutputAttr = 3;
73
74 const char kTextEncoderConfigAttr[] = "text_encoder_config";
75
76 // Initializes text encoder object from serialized options:
77 // The options are a flexbuffers attribute map that contain the op config
78 // with the key `text_encoder_config` as `TextEncoderConfig`.
Initialize(TfLiteContext * context,const char * buffer,size_t length)79 void* Initialize(TfLiteContext* context, const char* buffer, size_t length) {
80 const flexbuffers::Map& attr_map =
81 flexbuffers::GetRoot(reinterpret_cast<const uint8_t*>(buffer), length)
82 .AsMap();
83 const flexbuffers::Blob serialized_config =
84 attr_map[kTextEncoderConfigAttr].AsBlob();
85 const TextEncoderConfig* config =
86 flatbuffers::GetRoot<TextEncoderConfig>(serialized_config.data());
87
88 std::unique_ptr<TextEncoderOp> encoder_op(new TextEncoderOp());
89
90 // Create normalizer from options.
91 const TrieNode* charsmap_trie_nodes = reinterpret_cast<const TrieNode*>(
92 config->normalization_charsmap()->Data());
93 const int charsmap_trie_nodes_length =
94 config->normalization_charsmap()->Length() / sizeof(TrieNode);
95 encoder_op->normalizer.reset(new SentencePieceNormalizer(
96 DoubleArrayTrie(charsmap_trie_nodes, charsmap_trie_nodes_length),
97 StringPiece(config->normalization_charsmap_values()->data(),
98 config->normalization_charsmap_values()->size()),
99 config->add_dummy_prefix(), config->remove_extra_whitespaces(),
100 config->escape_whitespaces()));
101
102 const int num_pieces = config->pieces_scores()->Length();
103
104 switch (config->matcher_type()) {
105 case SentencePieceMatcherType_MAPPED_TRIE: {
106 const TrieNode* pieces_trie_nodes =
107 reinterpret_cast<const TrieNode*>(config->pieces()->Data());
108 const int pieces_trie_nodes_length =
109 config->pieces()->Length() / sizeof(TrieNode);
110 encoder_op->matcher.reset(
111 new DoubleArrayTrie(pieces_trie_nodes, pieces_trie_nodes_length));
112 break;
113 }
114 case SentencePieceMatcherType_SORTED_STRING_TABLE: {
115 encoder_op->matcher.reset(new SortedStringsTable(
116 num_pieces, config->pieces_offsets()->data(),
117 StringPiece(config->pieces()->data(), config->pieces()->Length())));
118 break;
119 }
120 default: {
121 TC3_LOG(ERROR) << "Unknown sentence piece matcher type.";
122 return nullptr;
123 }
124 }
125 encoder_op->encoder.reset(new Encoder(
126 encoder_op->matcher.get(), num_pieces, config->pieces_scores()->data(),
127 config->start_code(), config->end_code(), config->encoding_offset(),
128 config->unknown_code(), config->unknown_score()));
129 return encoder_op.release();
130 }
131
Free(TfLiteContext * context,void * buffer)132 void Free(TfLiteContext* context, void* buffer) {
133 delete reinterpret_cast<TextEncoderOp*>(buffer);
134 }
135
ResizeOutputTensors(TfLiteContext * context,TfLiteNode * node,int max_output_length)136 TfLiteStatus ResizeOutputTensors(TfLiteContext* context, TfLiteNode* node,
137 int max_output_length) {
138 TF_LITE_ENSURE_OK(
139 context,
140 ResizeOutputTensor(max_output_length,
141 &context->tensors[node->outputs->data[kOutputEncoded]],
142 context));
143
144 TF_LITE_ENSURE_OK(
145 context,
146 ResizeOutputTensor(
147 max_output_length,
148 &context->tensors[node->outputs->data[kOutputPosition]], context));
149
150 const int num_output_attrs = node->outputs->size - kOutputAttr;
151 for (int i = 0; i < num_output_attrs; ++i) {
152 TF_LITE_ENSURE_OK(
153 context,
154 ResizeOutputTensor(
155 max_output_length,
156 &context->tensors[node->outputs->data[kOutputAttr + i]], context));
157 }
158 return kTfLiteOk;
159 }
160
Prepare(TfLiteContext * context,TfLiteNode * node)161 TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
162 // Check that the batch dimension is kBatchSize.
163 const TfLiteTensor& input_text =
164 context->tensors[node->inputs->data[kInputTexts]];
165 TF_LITE_ENSURE_EQ(context, input_text.dims->size, kEncoderInputRank);
166 TF_LITE_ENSURE_EQ(context, input_text.dims->data[0], kEncoderBatchSize);
167
168 TfLiteTensor& output_lengths =
169 context->tensors[node->outputs->data[kOutputLengths]];
170 TfLiteTensor& output_encoded =
171 context->tensors[node->outputs->data[kOutputEncoded]];
172 TfLiteTensor& output_positions =
173 context->tensors[node->outputs->data[kOutputPosition]];
174
175 TF_LITE_ENSURE_OK(context,
176 context->ResizeTensor(context, &output_lengths,
177 CreateIntArray({kEncoderBatchSize})));
178
179 // Check that there are enough outputs for attributes.
180 const int num_output_attrs = node->outputs->size - kOutputAttr;
181 TF_LITE_ENSURE_EQ(context, node->inputs->size - kInputAttr, num_output_attrs);
182
183 // Copy attribute types from input to output tensors.
184 for (int i = 0; i < num_output_attrs; ++i) {
185 TfLiteTensor& input = context->tensors[node->inputs->data[kInputAttr + i]];
186 TfLiteTensor& output =
187 context->tensors[node->outputs->data[kOutputAttr + i]];
188 output.type = input.type;
189 }
190
191 const TfLiteTensor& output_length =
192 context->tensors[node->inputs->data[kInputMaxLength]];
193
194 if (tflite::IsConstantTensor(&output_length)) {
195 return ResizeOutputTensors(context, node, output_length.data.i64[0]);
196 } else {
197 tflite::SetTensorToDynamic(&output_encoded);
198 tflite::SetTensorToDynamic(&output_positions);
199 for (int i = 0; i < num_output_attrs; ++i) {
200 TfLiteTensor& output_attr =
201 context->tensors[node->outputs->data[kOutputAttr + i]];
202 tflite::SetTensorToDynamic(&output_attr);
203 }
204 }
205
206 return kTfLiteOk;
207 }
208
Eval(TfLiteContext * context,TfLiteNode * node)209 TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
210 if (node->user_data == nullptr) {
211 return kTfLiteError;
212 }
213 const TextEncoderOp* encoder_op =
214 reinterpret_cast<TextEncoderOp*>(node->user_data);
215 const TfLiteTensor& input_text =
216 context->tensors[node->inputs->data[kInputTexts]];
217 const int num_strings = tflite::GetStringCount(&input_text);
218 // Check that the number of strings matches the length parameter.
219 const int num_strings_param =
220 context->tensors[node->inputs->data[kInputNumInputs]].data.i32[0];
221 TF_LITE_ENSURE_EQ(context, num_strings, num_strings_param);
222
223 TfLiteTensor& output_encoded =
224 context->tensors[node->outputs->data[kOutputEncoded]];
225 if (tflite::IsDynamicTensor(&output_encoded)) {
226 const TfLiteTensor& output_length =
227 context->tensors[node->inputs->data[kInputMaxLength]];
228 TF_LITE_ENSURE_OK(
229 context, ResizeOutputTensors(context, node, output_length.data.i64[0]));
230 }
231 TfLiteTensor& output_positions =
232 context->tensors[node->outputs->data[kOutputPosition]];
233
234 std::vector<int> encoded_total;
235 std::vector<int> encoded_offsets;
236 std::vector<int> encoded_positions;
237 encoded_offsets.reserve(num_strings);
238 const int max_output_length = output_encoded.dims->data[1];
239 const int max_encoded_position = max_output_length;
240
241 for (int i = 0; i < num_strings; ++i) {
242 const auto& strref = tflite::GetString(&input_text, i);
243 std::string normalized;
244 TF_LITE_ENSURE(context,
245 encoder_op->normalizer->Normalize(
246 StringPiece(strref.str, strref.len), &normalized));
247 std::vector<int> encoded;
248 TF_LITE_ENSURE(context, encoder_op->encoder->Encode(normalized, &encoded));
249 encoded_total.insert(encoded_total.end(), encoded.begin(), encoded.end());
250 encoded_offsets.push_back(encoded_total.size());
251 for (int i = 0; i < encoded.size(); i++) {
252 encoded_positions.push_back(std::min(i, max_encoded_position - 1));
253 }
254 }
255
256 const int num_skip = CopyDataToTensorAndPadOrTruncate(
257 max_output_length, encoded_total,
258 /*padding_value=*/encoded_total.back(), &output_encoded);
259 TfLiteTensor& output_lengths =
260 context->tensors[node->outputs->data[kOutputLengths]];
261 output_lengths.data.i32[0] = encoded_total.size() - num_skip;
262 CopyDataToTensorAndPadOrTruncate(max_output_length, encoded_positions,
263 /*padding_value=*/max_encoded_position,
264 &output_positions);
265
266 // Process attributes, all checks of sizes and types are done in Prepare.
267 const int num_output_attrs = node->outputs->size - kOutputAttr;
268 TF_LITE_ENSURE_EQ(context, node->inputs->size - kInputAttr, num_output_attrs);
269 for (int i = 0; i < num_output_attrs; ++i) {
270 TfLiteStatus attr_status = CopyValuesToTensorAndPadOrTruncate(
271 context->tensors[node->inputs->data[kInputAttr + i]], encoded_offsets,
272 num_skip, context,
273 &context->tensors[node->outputs->data[kOutputAttr + i]]);
274 if (attr_status != kTfLiteOk) {
275 return attr_status;
276 }
277 }
278
279 return kTfLiteOk;
280 }
281
282 } // namespace
283 } // namespace libtextclassifier3
284
285 namespace tflite {
286 namespace ops {
287 namespace custom {
288
Register_TEXT_ENCODER()289 TfLiteRegistration* Register_TEXT_ENCODER() {
290 static TfLiteRegistration registration = {
291 libtextclassifier3::Initialize, libtextclassifier3::Free,
292 libtextclassifier3::Prepare, libtextclassifier3::Eval};
293 return ®istration;
294 }
295
296 } // namespace custom
297 } // namespace ops
298 } // namespace tflite
299