Home | History | Annotate | Download | only in annotator 
      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 "annotator/cached-features.h"
     18 
     19 #include "utils/base/logging.h"
     20 #include "utils/tensor-view.h"
     21 
     22 namespace libtextclassifier3 {
     23 
     24 namespace {
     25 
     26 int CalculateOutputFeaturesSize(const FeatureProcessorOptions* options,
     27                                 int feature_vector_size) {
     28   const bool bounds_sensitive_enabled =
     29       options->bounds_sensitive_features() &&
     30       options->bounds_sensitive_features()->enabled();
     31 
     32   int num_extracted_tokens = 0;
     33   if (bounds_sensitive_enabled) {
     34     const FeatureProcessorOptions_::BoundsSensitiveFeatures* config =
     35         options->bounds_sensitive_features();
     36     num_extracted_tokens += config->num_tokens_before();
     37     num_extracted_tokens += config->num_tokens_inside_left();
     38     num_extracted_tokens += config->num_tokens_inside_right();
     39     num_extracted_tokens += config->num_tokens_after();
     40     if (config->include_inside_bag()) {
     41       ++num_extracted_tokens;
     42     }
     43   } else {
     44     num_extracted_tokens = 2 * options->context_size() + 1;
     45   }
     46 
     47   int output_features_size = num_extracted_tokens * feature_vector_size;
     48 
     49   if (bounds_sensitive_enabled &&
     50       options->bounds_sensitive_features()->include_inside_length()) {
     51     ++output_features_size;
     52   }
     53 
     54   return output_features_size;
     55 }
     56 
     57 }  // namespace
     58 
     59 std::unique_ptr<CachedFeatures> CachedFeatures::Create(
     60     const TokenSpan& extraction_span,
     61     std::unique_ptr<std::vector<float>> features,
     62     std::unique_ptr<std::vector<float>> padding_features,
     63     const FeatureProcessorOptions* options, int feature_vector_size) {
     64   const int min_feature_version =
     65       options->bounds_sensitive_features() &&
     66               options->bounds_sensitive_features()->enabled()
     67           ? 2
     68           : 1;
     69   if (options->feature_version() < min_feature_version) {
     70     TC3_LOG(ERROR) << "Unsupported feature version.";
     71     return nullptr;
     72   }
     73 
     74   std::unique_ptr<CachedFeatures> cached_features(new CachedFeatures());
     75   cached_features->extraction_span_ = extraction_span;
     76   cached_features->features_ = std::move(features);
     77   cached_features->padding_features_ = std::move(padding_features);
     78   cached_features->options_ = options;
     79 
     80   cached_features->output_features_size_ =
     81       CalculateOutputFeaturesSize(options, feature_vector_size);
     82 
     83   return cached_features;
     84 }
     85 
     86 void CachedFeatures::AppendClickContextFeaturesForClick(
     87     int click_pos, std::vector<float>* output_features) const {
     88   click_pos -= extraction_span_.first;
     89 
     90   AppendFeaturesInternal(
     91       /*intended_span=*/ExpandTokenSpan(SingleTokenSpan(click_pos),
     92                                         options_->context_size(),
     93                                         options_->context_size()),
     94       /*read_mask_span=*/{0, TokenSpanSize(extraction_span_)}, output_features);
     95 }
     96 
     97 void CachedFeatures::AppendBoundsSensitiveFeaturesForSpan(
     98     TokenSpan selected_span, std::vector<float>* output_features) const {
     99   const FeatureProcessorOptions_::BoundsSensitiveFeatures* config =
    100       options_->bounds_sensitive_features();
    101 
    102   selected_span.first -= extraction_span_.first;
    103   selected_span.second -= extraction_span_.first;
    104 
    105   // Append the features for tokens around the left bound. Masks out tokens
    106   // after the right bound, so that if num_tokens_inside_left goes past it,
    107   // padding tokens will be used.
    108   AppendFeaturesInternal(
    109       /*intended_span=*/{selected_span.first - config->num_tokens_before(),
    110                          selected_span.first +
    111                              config->num_tokens_inside_left()},
    112       /*read_mask_span=*/{0, selected_span.second}, output_features);
    113 
    114   // Append the features for tokens around the right bound. Masks out tokens
    115   // before the left bound, so that if num_tokens_inside_right goes past it,
    116   // padding tokens will be used.
    117   AppendFeaturesInternal(
    118       /*intended_span=*/{selected_span.second -
    119                              config->num_tokens_inside_right(),
    120                          selected_span.second + config->num_tokens_after()},
    121       /*read_mask_span=*/{selected_span.first, TokenSpanSize(extraction_span_)},
    122       output_features);
    123 
    124   if (config->include_inside_bag()) {
    125     AppendBagFeatures(selected_span, output_features);
    126   }
    127 
    128   if (config->include_inside_length()) {
    129     output_features->push_back(
    130         static_cast<float>(TokenSpanSize(selected_span)));
    131   }
    132 }
    133 
    134 void CachedFeatures::AppendFeaturesInternal(
    135     const TokenSpan& intended_span, const TokenSpan& read_mask_span,
    136     std::vector<float>* output_features) const {
    137   const TokenSpan copy_span =
    138       IntersectTokenSpans(intended_span, read_mask_span);
    139   for (int i = intended_span.first; i < copy_span.first; ++i) {
    140     AppendPaddingFeatures(output_features);
    141   }
    142   output_features->insert(
    143       output_features->end(),
    144       features_->begin() + copy_span.first * NumFeaturesPerToken(),
    145       features_->begin() + copy_span.second * NumFeaturesPerToken());
    146   for (int i = copy_span.second; i < intended_span.second; ++i) {
    147     AppendPaddingFeatures(output_features);
    148   }
    149 }
    150 
    151 void CachedFeatures::AppendPaddingFeatures(
    152     std::vector<float>* output_features) const {
    153   output_features->insert(output_features->end(), padding_features_->begin(),
    154                           padding_features_->end());
    155 }
    156 
    157 void CachedFeatures::AppendBagFeatures(
    158     const TokenSpan& bag_span, std::vector<float>* output_features) const {
    159   const int offset = output_features->size();
    160   output_features->resize(output_features->size() + NumFeaturesPerToken());
    161   for (int i = bag_span.first; i < bag_span.second; ++i) {
    162     for (int j = 0; j < NumFeaturesPerToken(); ++j) {
    163       (*output_features)[offset + j] +=
    164           (*features_)[i * NumFeaturesPerToken() + j] / TokenSpanSize(bag_span);
    165     }
    166   }
    167 }
    168 
    169 int CachedFeatures::NumFeaturesPerToken() const {
    170   return padding_features_->size();
    171 }
    172 
    173 }  // namespace libtextclassifier3
    174