ct2rs 0.9.18

Rust bindings for OpenNMT/CTranslate2
Documentation 
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

#pragma once

#include "ctranslate2/layers/encoder.h"
#include "ctranslate2/layers/decoder.h"
#include "ctranslate2/models/model.h"
#include "ctranslate2/encoding.h"
#include "ctranslate2/generation.h"
#include "ctranslate2/scoring.h"

namespace ctranslate2 {
  namespace models {

    // Base class for language models.
    class LanguageModel : public Model {
    public:
      LanguageModel();

      const Vocabulary& get_vocabulary() const;

      // The returned cache is thread safe.
      layers::DecoderStateCache& get_state_cache() const;

    protected:
      void initialize(ModelReader& model_reader) override;

    private:
      std::shared_ptr<const Vocabulary> _vocabulary;
      std::shared_ptr<layers::DecoderStateCache> _state_cache;
    };


    // Base class for generative language models.
    class SequenceGeneratorReplica : public ModelReplica {
    public:
      SequenceGeneratorReplica(const std::shared_ptr<const LanguageModel>& model)
        : ModelReplica(model)
        , _model(model)
      {
      }

      static std::unique_ptr<SequenceGeneratorReplica> create_from_model(const Model& model) {
        return model.as_sequence_generator();
      }

      std::vector<ScoringResult>
      score(const std::vector<std::vector<std::string>>& tokens,
            const ScoringOptions& options = ScoringOptions());

      std::vector<GenerationResult>
      generate(const std::vector<std::vector<std::string>>& start_tokens,
               const GenerationOptions& options = GenerationOptions());

      StorageView forward(const std::vector<std::vector<std::string>>& tokens,
                          const bool return_log_probs);
      StorageView forward(const std::vector<std::vector<size_t>>& ids,
                          const bool return_log_probs);
      StorageView forward(const StorageView& ids,
                          const StorageView& lengths,
                          const bool return_log_probs);

    protected:
      virtual bool skip_scoring(const std::vector<std::string>& tokens,
                                const ScoringOptions& options,
                                ScoringResult& result) {
        (void)tokens;
        (void)options;
        (void)result;
        return false;
      }

      virtual std::vector<ScoringResult>
      run_scoring(const std::vector<std::vector<std::string>>& tokens,
                  const ScoringOptions& options) = 0;

      virtual std::vector<GenerationResult>
      run_generation(const std::vector<std::vector<std::string>>& start_tokens,
                     const GenerationOptions& options) = 0;

      virtual StorageView forward(const StorageView& ids, const StorageView& lengths) = 0;

    private:
      const std::shared_ptr<const LanguageModel> _model;
    };


    // A model generating sequences with a decoder.
    class DecoderReplica : public SequenceGeneratorReplica {
    public:
      DecoderReplica(const std::shared_ptr<const LanguageModel>& model,
                     std::unique_ptr<layers::Decoder> decoder);

    protected:
      bool skip_scoring(const std::vector<std::string>& tokens,
                        const ScoringOptions& options,
                        ScoringResult& result) override;

      std::vector<ScoringResult>
      run_scoring(const std::vector<std::vector<std::string>>& tokens,
                  const ScoringOptions& options) override;

      std::vector<GenerationResult>
      run_generation(const std::vector<std::vector<std::string>>& start_tokens,
                     const GenerationOptions& options) override;

      StorageView forward(const StorageView& ids, const StorageView& lengths) override;

    private:
      const std::shared_ptr<const LanguageModel> _model;
      const std::unique_ptr<layers::Decoder> _decoder;
    };


    // Base class for sequence encoders.
    class SequenceEncoderReplica : public ModelReplica {
    public:
      SequenceEncoderReplica(const std::shared_ptr<const LanguageModel>& model)
        : ModelReplica(model)
        , _model(model)
      {
      }

      static std::unique_ptr<SequenceEncoderReplica> create_from_model(const Model& model) {
        return model.as_sequence_encoder();
      }

      EncoderForwardOutput forward(const std::vector<std::vector<std::string>>& tokens,
                                   const std::vector<std::vector<size_t>>& token_type_ids = {});
      EncoderForwardOutput forward(const std::vector<std::vector<size_t>>& ids,
                                   const std::vector<std::vector<size_t>>& token_type_ids = {});
      EncoderForwardOutput forward(const StorageView& ids,
                                   const StorageView& lengths,
                                   const std::vector<std::vector<size_t>>& token_type_ids = {});

    protected:
      virtual EncoderForwardOutput
      forward_impl(const StorageView& ids,
                   const StorageView& lengths,
                   const StorageView& token_type_ids) = 0;

    private:
      const std::shared_ptr<const LanguageModel> _model;
    };


    // A model encoding sequences with an encoder layer.
    class EncoderReplica : public SequenceEncoderReplica {
    public:
      EncoderReplica(const std::shared_ptr<const LanguageModel>& model,
                     std::unique_ptr<layers::Encoder> encoder);

    protected:
      EncoderForwardOutput
      forward_impl(const StorageView& ids,
                   const StorageView& lengths,
                   const StorageView& token_type_ids) override;

    private:
      const std::shared_ptr<const LanguageModel> _model;
      const std::unique_ptr<layers::Encoder> _encoder;
      const ops::ActivationType _pooler_activation;
      const std::unique_ptr<layers::Dense> _pooler_dense;
    };

  }
}