1 /*
2 * Copyright (C) 2017 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 "common/embedding-network.h"
18
19 #include <math.h>
20
21 #include "common/simple-adder.h"
22 #include "util/base/integral_types.h"
23 #include "util/base/logging.h"
24
25 namespace libtextclassifier {
26 namespace nlp_core {
27
28 namespace {
29
30 // Returns true if and only if matrix does not use any quantization.
CheckNoQuantization(const EmbeddingNetworkParams::Matrix & matrix)31 bool CheckNoQuantization(const EmbeddingNetworkParams::Matrix &matrix) {
32 if (matrix.quant_type != QuantizationType::NONE) {
33 TC_LOG(ERROR) << "Unsupported quantization";
34 TC_DCHECK(false); // Crash in debug mode.
35 return false;
36 }
37 return true;
38 }
39
40 // Initializes a Matrix object with the parameters from the MatrixParams
41 // source_matrix. source_matrix should not use quantization.
42 //
43 // Returns true on success, false on error.
InitNonQuantizedMatrix(const EmbeddingNetworkParams::Matrix & source_matrix,EmbeddingNetwork::Matrix * mat)44 bool InitNonQuantizedMatrix(const EmbeddingNetworkParams::Matrix &source_matrix,
45 EmbeddingNetwork::Matrix *mat) {
46 mat->resize(source_matrix.rows);
47
48 // Before we access the weights as floats, we need to check that they are
49 // really floats, i.e., no quantization is used.
50 if (!CheckNoQuantization(source_matrix)) return false;
51 const float *weights =
52 reinterpret_cast<const float *>(source_matrix.elements);
53 for (int r = 0; r < source_matrix.rows; ++r) {
54 (*mat)[r] = EmbeddingNetwork::VectorWrapper(weights, source_matrix.cols);
55 weights += source_matrix.cols;
56 }
57 return true;
58 }
59
60 // Initializes a VectorWrapper object with the parameters from the MatrixParams
61 // source_matrix. source_matrix should have exactly one column and should not
62 // use quantization.
63 //
64 // Returns true on success, false on error.
InitNonQuantizedVector(const EmbeddingNetworkParams::Matrix & source_matrix,EmbeddingNetwork::VectorWrapper * vector)65 bool InitNonQuantizedVector(const EmbeddingNetworkParams::Matrix &source_matrix,
66 EmbeddingNetwork::VectorWrapper *vector) {
67 if (source_matrix.cols != 1) {
68 TC_LOG(ERROR) << "wrong #cols " << source_matrix.cols;
69 return false;
70 }
71 if (!CheckNoQuantization(source_matrix)) {
72 TC_LOG(ERROR) << "unsupported quantization";
73 return false;
74 }
75 // Before we access the weights as floats, we need to check that they are
76 // really floats, i.e., no quantization is used.
77 if (!CheckNoQuantization(source_matrix)) return false;
78 const float *weights =
79 reinterpret_cast<const float *>(source_matrix.elements);
80 *vector = EmbeddingNetwork::VectorWrapper(weights, source_matrix.rows);
81 return true;
82 }
83
84 // Computes y = weights * Relu(x) + b where Relu is optionally applied.
85 template <typename ScaleAdderClass>
SparseReluProductPlusBias(bool apply_relu,const EmbeddingNetwork::Matrix & weights,const EmbeddingNetwork::VectorWrapper & b,const VectorSpan<float> & x,EmbeddingNetwork::Vector * y)86 bool SparseReluProductPlusBias(bool apply_relu,
87 const EmbeddingNetwork::Matrix &weights,
88 const EmbeddingNetwork::VectorWrapper &b,
89 const VectorSpan<float> &x,
90 EmbeddingNetwork::Vector *y) {
91 // Check that dimensions match.
92 if ((x.size() != weights.size()) || weights.empty()) {
93 TC_LOG(ERROR) << x.size() << " != " << weights.size();
94 return false;
95 }
96 if (weights[0].size() != b.size()) {
97 TC_LOG(ERROR) << weights[0].size() << " != " << b.size();
98 return false;
99 }
100
101 y->assign(b.data(), b.data() + b.size());
102 ScaleAdderClass adder(y->data(), y->size());
103
104 const int x_size = x.size();
105 for (int i = 0; i < x_size; ++i) {
106 const float &scale = x[i];
107 if (apply_relu) {
108 if (scale > 0) {
109 adder.LazyScaleAdd(weights[i].data(), scale);
110 }
111 } else {
112 adder.LazyScaleAdd(weights[i].data(), scale);
113 }
114 }
115 return true;
116 }
117 } // namespace
118
ConcatEmbeddings(const std::vector<FeatureVector> & feature_vectors,Vector * concat) const119 bool EmbeddingNetwork::ConcatEmbeddings(
120 const std::vector<FeatureVector> &feature_vectors, Vector *concat) const {
121 concat->resize(concat_layer_size_);
122
123 // Invariant 1: feature_vectors contains exactly one element for each
124 // embedding space. That element is itself a FeatureVector, which may be
125 // empty, but it should be there.
126 if (feature_vectors.size() != embedding_matrices_.size()) {
127 TC_LOG(ERROR) << feature_vectors.size()
128 << " != " << embedding_matrices_.size();
129 return false;
130 }
131
132 // "es_index" stands for "embedding space index".
133 for (int es_index = 0; es_index < feature_vectors.size(); ++es_index) {
134 // Access is safe by es_index loop bounds and Invariant 1.
135 EmbeddingMatrix *const embedding_matrix =
136 embedding_matrices_[es_index].get();
137 if (embedding_matrix == nullptr) {
138 // Should not happen, hence our terse log error message.
139 TC_LOG(ERROR) << es_index;
140 return false;
141 }
142
143 // Access is safe due to es_index loop bounds.
144 const FeatureVector &feature_vector = feature_vectors[es_index];
145
146 // Access is safe by es_index loop bounds, Invariant 1, and Invariant 2.
147 const int concat_offset = concat_offset_[es_index];
148
149 if (!GetEmbeddingInternal(feature_vector, embedding_matrix, concat_offset,
150 concat->data(), concat->size())) {
151 TC_LOG(ERROR) << es_index;
152 return false;
153 }
154 }
155 return true;
156 }
157
GetEmbedding(const FeatureVector & feature_vector,int es_index,float * embedding) const158 bool EmbeddingNetwork::GetEmbedding(const FeatureVector &feature_vector,
159 int es_index, float *embedding) const {
160 EmbeddingMatrix *const embedding_matrix = embedding_matrices_[es_index].get();
161 if (embedding_matrix == nullptr) {
162 // Should not happen, hence our terse log error message.
163 TC_LOG(ERROR) << es_index;
164 return false;
165 }
166 return GetEmbeddingInternal(feature_vector, embedding_matrix, 0, embedding,
167 embedding_matrices_[es_index]->dim());
168 }
169
GetEmbeddingInternal(const FeatureVector & feature_vector,EmbeddingMatrix * const embedding_matrix,const int concat_offset,float * concat,int concat_size) const170 bool EmbeddingNetwork::GetEmbeddingInternal(
171 const FeatureVector &feature_vector,
172 EmbeddingMatrix *const embedding_matrix, const int concat_offset,
173 float *concat, int concat_size) const {
174 const int embedding_dim = embedding_matrix->dim();
175 const bool is_quantized =
176 embedding_matrix->quant_type() != QuantizationType::NONE;
177 const int num_features = feature_vector.size();
178 for (int fi = 0; fi < num_features; ++fi) {
179 // Both accesses below are safe due to loop bounds for fi.
180 const FeatureType *feature_type = feature_vector.type(fi);
181 const FeatureValue feature_value = feature_vector.value(fi);
182 const int feature_offset =
183 concat_offset + feature_type->base() * embedding_dim;
184
185 // Code below updates max(0, embedding_dim) elements from concat, starting
186 // with index feature_offset. Check below ensures these updates are safe.
187 if ((feature_offset < 0) ||
188 (feature_offset + embedding_dim > concat_size)) {
189 TC_LOG(ERROR) << fi << ": " << feature_offset << " " << embedding_dim
190 << " " << concat_size;
191 return false;
192 }
193
194 // Pointer to float / uint8 weights for relevant embedding.
195 const void *embedding_data;
196
197 // Multiplier for each embedding weight.
198 float multiplier;
199
200 if (feature_type->is_continuous()) {
201 // Continuous features (encoded as FloatFeatureValue).
202 FloatFeatureValue float_feature_value(feature_value);
203 const int id = float_feature_value.id;
204 embedding_matrix->get_embedding(id, &embedding_data, &multiplier);
205 multiplier *= float_feature_value.weight;
206 } else {
207 // Discrete features: every present feature has implicit value 1.0.
208 // Hence, after we grab the multiplier below, we don't multiply it by
209 // any weight.
210 embedding_matrix->get_embedding(feature_value, &embedding_data,
211 &multiplier);
212 }
213
214 // Weighted embeddings will be added starting from this address.
215 float *concat_ptr = concat + feature_offset;
216
217 if (is_quantized) {
218 const uint8 *quant_weights =
219 reinterpret_cast<const uint8 *>(embedding_data);
220 for (int i = 0; i < embedding_dim; ++i, ++quant_weights, ++concat_ptr) {
221 // 128 is bias for UINT8 quantization, only one we currently support.
222 *concat_ptr += (static_cast<int>(*quant_weights) - 128) * multiplier;
223 }
224 } else {
225 const float *weights = reinterpret_cast<const float *>(embedding_data);
226 for (int i = 0; i < embedding_dim; ++i, ++weights, ++concat_ptr) {
227 *concat_ptr += *weights * multiplier;
228 }
229 }
230 }
231 return true;
232 }
233
ComputeLogits(const VectorSpan<float> & input,Vector * scores) const234 bool EmbeddingNetwork::ComputeLogits(const VectorSpan<float> &input,
235 Vector *scores) const {
236 return EmbeddingNetwork::ComputeLogitsInternal(input, scores);
237 }
238
ComputeLogits(const Vector & input,Vector * scores) const239 bool EmbeddingNetwork::ComputeLogits(const Vector &input,
240 Vector *scores) const {
241 return EmbeddingNetwork::ComputeLogitsInternal(input, scores);
242 }
243
ComputeLogitsInternal(const VectorSpan<float> & input,Vector * scores) const244 bool EmbeddingNetwork::ComputeLogitsInternal(const VectorSpan<float> &input,
245 Vector *scores) const {
246 return FinishComputeFinalScoresInternal<SimpleAdder>(input, scores);
247 }
248
249 template <typename ScaleAdderClass>
FinishComputeFinalScoresInternal(const VectorSpan<float> & input,Vector * scores) const250 bool EmbeddingNetwork::FinishComputeFinalScoresInternal(
251 const VectorSpan<float> &input, Vector *scores) const {
252 // This vector serves as an alternating storage for activations of the
253 // different layers. We can't use just one vector here because all of the
254 // activations of the previous layer are needed for computation of
255 // activations of the next one.
256 std::vector<Vector> h_storage(2);
257
258 // Compute pre-logits activations.
259 VectorSpan<float> h_in(input);
260 Vector *h_out;
261 for (int i = 0; i < hidden_weights_.size(); ++i) {
262 const bool apply_relu = i > 0;
263 h_out = &(h_storage[i % 2]);
264 h_out->resize(hidden_bias_[i].size());
265 if (!SparseReluProductPlusBias<ScaleAdderClass>(
266 apply_relu, hidden_weights_[i], hidden_bias_[i], h_in, h_out)) {
267 return false;
268 }
269 h_in = VectorSpan<float>(*h_out);
270 }
271
272 // Compute logit scores.
273 if (!SparseReluProductPlusBias<ScaleAdderClass>(
274 true, softmax_weights_, softmax_bias_, h_in, scores)) {
275 return false;
276 }
277
278 return true;
279 }
280
ComputeFinalScores(const std::vector<FeatureVector> & features,Vector * scores) const281 bool EmbeddingNetwork::ComputeFinalScores(
282 const std::vector<FeatureVector> &features, Vector *scores) const {
283 return ComputeFinalScores(features, {}, scores);
284 }
285
ComputeFinalScores(const std::vector<FeatureVector> & features,const std::vector<float> extra_inputs,Vector * scores) const286 bool EmbeddingNetwork::ComputeFinalScores(
287 const std::vector<FeatureVector> &features,
288 const std::vector<float> extra_inputs, Vector *scores) const {
289 // If we haven't successfully initialized, return without doing anything.
290 if (!is_valid()) return false;
291
292 Vector concat;
293 if (!ConcatEmbeddings(features, &concat)) return false;
294
295 if (!extra_inputs.empty()) {
296 concat.reserve(concat.size() + extra_inputs.size());
297 for (int i = 0; i < extra_inputs.size(); i++) {
298 concat.push_back(extra_inputs[i]);
299 }
300 }
301
302 scores->resize(softmax_bias_.size());
303 return ComputeLogits(concat, scores);
304 }
305
EmbeddingNetwork(const EmbeddingNetworkParams * model)306 EmbeddingNetwork::EmbeddingNetwork(const EmbeddingNetworkParams *model) {
307 // We'll set valid_ to true only if construction is successful. If we detect
308 // an error along the way, we log an informative message and return early, but
309 // we do not crash.
310 valid_ = false;
311
312 // Fill embedding_matrices_, concat_offset_, and concat_layer_size_.
313 const int num_embedding_spaces = model->GetNumEmbeddingSpaces();
314 int offset_sum = 0;
315 for (int i = 0; i < num_embedding_spaces; ++i) {
316 concat_offset_.push_back(offset_sum);
317 const EmbeddingNetworkParams::Matrix matrix = model->GetEmbeddingMatrix(i);
318 if (matrix.quant_type != QuantizationType::UINT8) {
319 TC_LOG(ERROR) << "Unsupported quantization for embedding #" << i << ": "
320 << static_cast<int>(matrix.quant_type);
321 return;
322 }
323
324 // There is no way to accomodate an empty embedding matrix. E.g., there is
325 // no way for get_embedding to return something that can be read safely.
326 // Hence, we catch that error here and return early.
327 if (matrix.rows == 0) {
328 TC_LOG(ERROR) << "Empty embedding matrix #" << i;
329 return;
330 }
331 embedding_matrices_.emplace_back(new EmbeddingMatrix(matrix));
332 const int embedding_dim = embedding_matrices_.back()->dim();
333 offset_sum += embedding_dim * model->GetNumFeaturesInEmbeddingSpace(i);
334 }
335 concat_layer_size_ = offset_sum;
336
337 // Invariant 2 (trivial by the code above).
338 TC_DCHECK_EQ(concat_offset_.size(), embedding_matrices_.size());
339
340 const int num_hidden_layers = model->GetNumHiddenLayers();
341 if (num_hidden_layers < 1) {
342 TC_LOG(ERROR) << "Wrong number of hidden layers: " << num_hidden_layers;
343 return;
344 }
345 hidden_weights_.resize(num_hidden_layers);
346 hidden_bias_.resize(num_hidden_layers);
347
348 for (int i = 0; i < num_hidden_layers; ++i) {
349 const EmbeddingNetworkParams::Matrix matrix =
350 model->GetHiddenLayerMatrix(i);
351 const EmbeddingNetworkParams::Matrix bias = model->GetHiddenLayerBias(i);
352 if (!InitNonQuantizedMatrix(matrix, &hidden_weights_[i]) ||
353 !InitNonQuantizedVector(bias, &hidden_bias_[i])) {
354 TC_LOG(ERROR) << "Bad hidden layer #" << i;
355 return;
356 }
357 }
358
359 if (!model->HasSoftmaxLayer()) {
360 TC_LOG(ERROR) << "Missing softmax layer";
361 return;
362 }
363 const EmbeddingNetworkParams::Matrix softmax = model->GetSoftmaxMatrix();
364 const EmbeddingNetworkParams::Matrix softmax_bias = model->GetSoftmaxBias();
365 if (!InitNonQuantizedMatrix(softmax, &softmax_weights_) ||
366 !InitNonQuantizedVector(softmax_bias, &softmax_bias_)) {
367 TC_LOG(ERROR) << "Bad softmax layer";
368 return;
369 }
370
371 // Everything looks good.
372 valid_ = true;
373 }
374
EmbeddingSize(int es_index) const375 int EmbeddingNetwork::EmbeddingSize(int es_index) const {
376 return embedding_matrices_[es_index]->dim();
377 }
378
379 } // namespace nlp_core
380 } // namespace libtextclassifier
381