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
173
174
175
176
177

use std::sync::Arc;

use conllx::graph::Sentence;
use failure::{err_msg, Error};
use rand::{Rng, SeedableRng};
use serde::Serialize;

use crate::sampling::ZipfRangeGenerator;
use crate::train_model::{NegativeSamples, TrainIterFrom};
use crate::util::ReseedOnCloneRng;
use crate::{
    CommonConfig, DepembedsConfig, Dependency, DependencyIterator, SimpleVocab, SimpleVocabConfig,
    SubwordVocab, SubwordVocabConfig, Trainer, Vocab,
};

/// Dependency embeddings Trainer.
///
/// The `DepembedsTrainer` holds the information and logic necessary to transform a
/// `conllx::Sentence` into an iterator of focus and context tuples. The struct is cheap to clone
/// because the vocabulary is shared between clones.
#[derive(Clone)]
pub struct DepembedsTrainer<R> {
    dep_config: DepembedsConfig,
    common_config: CommonConfig,
    input_vocab: Arc<SubwordVocab>,
    output_vocab: Arc<SimpleVocab<Dependency>>,
    range_gen: ZipfRangeGenerator<R>,
    rng: R,
}

impl<R> DepembedsTrainer<R> {
    pub fn dep_config(&self) -> DepembedsConfig {
        self.dep_config
    }
}

impl<R> DepembedsTrainer<ReseedOnCloneRng<R>>
where
    R: Rng + Clone + SeedableRng,
{
    /// Constructs a new `DepTrainer`.
    pub fn new(
        input_vocab: SubwordVocab,
        output_vocab: SimpleVocab<Dependency>,
        common_config: CommonConfig,
        dep_config: DepembedsConfig,
        rng: R,
    ) -> Self {
        let rng = ReseedOnCloneRng(rng);
        let range_gen = ZipfRangeGenerator::new_with_exponent(
            rng.clone(),
            output_vocab.len(),
            common_config.zipf_exponent,
        );
        DepembedsTrainer {
            common_config,
            dep_config,
            input_vocab: Arc::new(input_vocab),
            output_vocab: Arc::new(output_vocab),
            range_gen,
            rng,
        }
    }
}

impl<R> NegativeSamples for DepembedsTrainer<R>
where
    R: Rng,
{
    fn negative_sample(&mut self, output: usize) -> usize {
        loop {
            let negative = self.range_gen.next().unwrap();
            if negative != output {
                return negative;
            }
        }
    }
}

impl<R> TrainIterFrom<Sentence> for DepembedsTrainer<R>
where
    R: Rng,
{
    type Iter = Box<Iterator<Item = (usize, Vec<usize>)>>;
    type Contexts = Vec<usize>;

    fn train_iter_from(&mut self, sentence: &Sentence) -> Self::Iter {
        let invalid_idx = self.input_vocab.len();
        let mut tokens = vec![invalid_idx; sentence.len() - 1];
        for (idx, token) in sentence.iter().filter_map(|node| node.token()).enumerate() {
            if let Some(vocab_idx) = self.input_vocab.idx(token.form()) {
                if self.rng.gen_range(0f32, 1f32) < self.input_vocab.discard(vocab_idx) {
                    tokens[idx] = vocab_idx
                }
            }
        }

        let mut contexts = vec![Vec::new(); sentence.len() - 1];
        let graph = sentence.dep_graph();
        for (focus, dep) in DependencyIterator::new_from_config(&graph, self.dep_config)
            .filter(|(focus, _dep)| tokens[*focus] != invalid_idx)
        {
            if let Some(dep_id) = self.output_vocab.idx(&dep) {
                if self.rng.gen_range(0f32, 1f32) < self.output_vocab.discard(dep_id) {
                    contexts[focus].push(dep_id)
                }
            }
        }
        Box::new(
            tokens
                .into_iter()
                .zip(contexts.into_iter())
                .filter(move |(focus, _)| *focus != invalid_idx),
        )
    }
}

impl<R> Trainer for DepembedsTrainer<R>
where
    R: Rng,
{
    type InputVocab = SubwordVocab;
    type Metadata = DepembedsMetadata<SubwordVocabConfig, SimpleVocabConfig>;

    fn input_indices(&self, idx: usize) -> Vec<u64> {
        let mut v = self.input_vocab.subword_indices_idx(idx).unwrap().to_vec();
        v.push(idx as u64);
        v
    }

    fn input_vocab(&self) -> &SubwordVocab {
        &self.input_vocab
    }

    fn try_into_input_vocab(self) -> Result<Self::InputVocab, Error> {
        match Arc::try_unwrap(self.input_vocab) {
            Ok(vocab) => Ok(vocab),
            Err(_) => Err(err_msg("Cannot unwrap input vocab.")),
        }
    }

    fn n_input_types(&self) -> usize {
        let n_buckets = 2usize.pow(self.input_vocab().config().buckets_exp);
        n_buckets + self.input_vocab().len()
    }

    fn n_output_types(&self) -> usize {
        self.output_vocab.len()
    }

    fn config(&self) -> &CommonConfig {
        &self.common_config
    }

    fn to_metadata(&self) -> Self::Metadata {
        DepembedsMetadata {
            common_config: self.common_config,
            dep_config: self.dep_config,
            input_vocab_config: self.input_vocab.config(),
            output_vocab_config: self.output_vocab.config(),
        }
    }
}

/// Metadata for dependency embeddings.
#[derive(Clone, Copy, Debug, Serialize)]
pub struct DepembedsMetadata<IC, OC>
where
    IC: Serialize,
    OC: Serialize,
{
    common_config: CommonConfig,
    #[serde(rename = "model_config")]
    dep_config: DepembedsConfig,
    input_vocab_config: IC,
    output_vocab_config: OC,
}