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

//! Reader and writer for the word2vec binary format.
//!
//! Embeddings in the word2vec binary format are these formats are
//! read as follows:
//!
//! ```
//! use std::fs::File;
//! use std::io::BufReader;
//!
//! use finalfusion::prelude::*;
//!
//! let mut reader = BufReader::new(File::open("testdata/similarity.bin").unwrap());
//!
//! // Read the embeddings. The second arguments specifies whether
//! // the embeddings should be normalized to unit vectors.
//! let embeddings = Embeddings::read_word2vec_binary(&mut reader)
//!     .unwrap();
//!
//! // Look up an embedding.
//! let embedding = embeddings.embedding("Berlin");
//! ```

use std::io::{BufRead, Write};
use std::mem;
use std::slice::from_raw_parts_mut;

use byteorder::{LittleEndian, WriteBytesExt};
use failure::{err_msg, Error};
use ndarray::{Array2, Axis};

use crate::embeddings::Embeddings;
use crate::norms::NdNorms;
use crate::storage::{CowArray, NdArray, Storage, StorageViewMut};
use crate::util::l2_normalize_array;
use crate::vocab::{SimpleVocab, Vocab};

/// Method to construct `Embeddings` from a word2vec binary file.
///
/// This trait defines an extension to `Embeddings` to read the word embeddings
/// from a file in word2vec binary format.
pub trait ReadWord2Vec<R>
where
    Self: Sized,
    R: BufRead,
{
    /// Read the embeddings from the given buffered reader.
    fn read_word2vec_binary(reader: &mut R) -> Result<Self, Error>;
}

impl<R> ReadWord2Vec<R> for Embeddings<SimpleVocab, NdArray>
where
    R: BufRead,
{
    fn read_word2vec_binary(reader: &mut R) -> Result<Self, Error> {
        let (_, vocab, mut storage, _) = Embeddings::read_word2vec_binary_raw(reader)?.into_parts();
        let norms = l2_normalize_array(storage.view_mut());

        Ok(Embeddings::new(None, vocab, storage, NdNorms(norms)))
    }
}

/// Read raw, unnormalized embeddings.
pub(crate) trait ReadWord2VecRaw<R>
where
    Self: Sized,
    R: BufRead,
{
    /// Read the embeddings from the given buffered reader.
    fn read_word2vec_binary_raw(reader: &mut R) -> Result<Self, Error>;
}

impl<R> ReadWord2VecRaw<R> for Embeddings<SimpleVocab, NdArray>
where
    R: BufRead,
{
    fn read_word2vec_binary_raw(reader: &mut R) -> Result<Self, Error> {
        let n_words = read_number(reader, b' ')?;
        let embed_len = read_number(reader, b'\n')?;

        let mut matrix = Array2::zeros((n_words, embed_len));
        let mut words = Vec::with_capacity(n_words);

        for idx in 0..n_words {
            let word = read_string(reader, b' ')?;
            let word = word.trim();
            words.push(word.to_owned());

            let mut embedding = matrix.index_axis_mut(Axis(0), idx);

            {
                let mut embedding_raw = match embedding.as_slice_mut() {
                    Some(s) => unsafe { typed_to_bytes(s) },
                    None => return Err(err_msg("Matrix not contiguous")),
                };
                reader.read_exact(&mut embedding_raw)?;
            }
        }

        Ok(Embeddings::new_without_norms(
            None,
            SimpleVocab::new(words),
            NdArray(matrix),
        ))
    }
}

fn read_number(reader: &mut BufRead, delim: u8) -> Result<usize, Error> {
    let field_str = read_string(reader, delim)?;
    Ok(field_str.parse()?)
}

fn read_string(reader: &mut BufRead, delim: u8) -> Result<String, Error> {
    let mut buf = Vec::new();
    reader.read_until(delim, &mut buf)?;
    buf.pop();
    Ok(String::from_utf8(buf)?)
}

unsafe fn typed_to_bytes<T>(slice: &mut [T]) -> &mut [u8] {
    from_raw_parts_mut(
        slice.as_mut_ptr() as *mut u8,
        slice.len() * mem::size_of::<T>(),
    )
}

/// Method to write `Embeddings` to a word2vec binary file.
///
/// This trait defines an extension to `Embeddings` to write the word embeddings
/// to a file in word2vec binary format.
pub trait WriteWord2Vec<W>
where
    W: Write,
{
    /// Write the embeddings from the given writer.
    ///
    /// If `unnormalize` is `true`, the norms vector is used to
    /// restore the original vector magnitudes.
    fn write_word2vec_binary(&self, w: &mut W, unnormalize: bool) -> Result<(), Error>;
}

impl<W, V, S> WriteWord2Vec<W> for Embeddings<V, S>
where
    W: Write,
    V: Vocab,
    S: Storage,
{
    fn write_word2vec_binary(&self, w: &mut W, unnormalize: bool) -> Result<(), Error>
    where
        W: Write,
    {
        writeln!(w, "{} {}", self.vocab().len(), self.dims())?;

        for (word, embed_norm) in self.iter_with_norms() {
            write!(w, "{} ", word)?;

            let embed = if unnormalize {
                CowArray::Owned(embed_norm.into_unnormalized())
            } else {
                embed_norm.embedding
            };

            for v in embed.as_view() {
                w.write_f32::<LittleEndian>(*v)?;
            }

            w.write_all(&[0x0a])?;
        }

        Ok(())
    }
}