candding 0.1.0

Pure candle embeddings for Rust with no ONNX: dense, sparse, multi-vector and rerank models today, image models on the roadmap
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
166
167
168
169
170

use super::*;

/// The specials are filtered before either index path, leaving `doc_len == 2`; "cat" takes the dense path, "xyzzy" the hashed.
/// A dense index always sits below the hashed range's shift of 128 000, so the dense entry sorts first.
#[test]
fn a_row_mixing_recognized_and_unrecognized_words_uses_both_index_paths() {
  let vocab = vocab(&["cat"]);
  let stopwords = stopwords(&[]);
  let tokens = token_text(&["[CLS]", "cat", "xyzzy", "[SEP]"]);
  let matrix = Tensor::from_vec(
    vec![
      0f32, 0., 0., 0., 0., 0., 0., 0., 1.0, 0., 0., 0., 0., 0., 0., 0.,
    ],
    (2, 2, 4),
    &Device::Cpu,
  )
  .unwrap();
  let hidden_row = vec![
    vec![0.0f32, 0.0],
    vec![1.0, 0.0],
    vec![0.0, 0.0],
    vec![0.0, 0.0],
  ];

  let embedding = pooling::pool_row(
    &hidden_row,
    &[0, 1, 2, 3],
    &tokens,
    &vocab,
    &stopwords,
    &matrix,
    false,
  )
  .unwrap();

  let tf = 1.0 * (1.2 + 1.0) / (1.0 + 1.2 * (1.0 - 0.75 + 0.75 * 2.0 / 150.0));
  let dense = index::dense_index(1, 0);
  let hashed = index::hashed_index("xyzzy");

  assert!(dense < 128_000, "{dense}");
  assert!(hashed >= 128_000, "{hashed}");
  assert_eq!(embedding.indices, vec![dense, hashed]);
  for &got in &embedding.values {
    assert!((got - tf).abs() < 1e-5, "{got} vs {tf}");
  }
}

#[test]
fn a_row_of_only_unrecognized_words_produces_only_hashed_entries() {
  let vocab = vocab(&["cat"]);
  let stopwords = stopwords(&[]);
  let tokens = token_text(&["foo", "bar"]);
  let matrix = Tensor::from_vec(vec![0f32; 16], (2, 2, 4), &Device::Cpu).unwrap();
  let hidden_row = vec![vec![0.0f32, 0.0], vec![0.0, 0.0]];

  let embedding = pooling::pool_row(
    &hidden_row,
    &[0, 1],
    &tokens,
    &vocab,
    &stopwords,
    &matrix,
    false,
  )
  .unwrap();

  let tf = 1.0 * (1.2 + 1.0) / (1.0 + 1.2 * (1.0 - 0.75 + 0.75 * 2.0 / 150.0));
  let mut expected = [index::hashed_index("foo"), index::hashed_index("bar")];
  expected.sort_unstable();

  assert_eq!(embedding.indices, expected);
  assert!(
    embedding.indices.iter().all(|&i| i >= 128_000),
    "{:?}",
    embedding.indices
  );
  for &got in &embedding.values {
    assert!((got - tf).abs() < 1e-5, "{got} vs {tf}");
  }
}

/// Counted toward `doc_len`, a bare `.` would reweight every word, so "cat"'s weight is compared with a control row.
#[test]
fn a_bare_punctuation_token_never_moves_a_recognized_words_own_weight() {
  let vocab = vocab(&["cat"]);
  let stopwords = stopwords(&[]);
  let matrix = Tensor::from_vec(
    vec![
      0f32, 0., 0., 0., 0., 0., 0., 0., 1.0, 0., 0., 0., 0., 0., 0., 0.,
    ],
    (2, 2, 4),
    &Device::Cpu,
  )
  .unwrap();

  let control_tokens = token_text(&["[CLS]", "cat", "[SEP]"]);
  let control_hidden = vec![vec![0.0f32, 0.0], vec![1.0, 0.0], vec![0.0, 0.0]];
  let control = pooling::pool_row(
    &control_hidden,
    &[0, 1, 2],
    &control_tokens,
    &vocab,
    &stopwords,
    &matrix,
    false,
  )
  .unwrap();

  let punctuated_tokens = token_text(&["[CLS]", "cat", ".", "[SEP]"]);
  let punctuated_hidden = vec![
    vec![0.0f32, 0.0],
    vec![1.0, 0.0],
    vec![0.0, 0.0],
    vec![0.0, 0.0],
  ];
  let punctuated = pooling::pool_row(
    &punctuated_hidden,
    &[0, 1, 2, 3],
    &punctuated_tokens,
    &vocab,
    &stopwords,
    &matrix,
    false,
  )
  .unwrap();

  assert_eq!(
    control.indices,
    vec![index::dense_index(1, 0)],
    "control row (no punctuation)"
  );
  // "cat"'s dense index (4) sorts below every hashed index, so `values[0]` is its weight in both rows.
  assert_eq!(
    punctuated.values[0], control.values[0],
    "\"cat\"'s own weight must not move just because the row also contains a bare \".\""
  );
  assert_eq!(
    punctuated.indices, control.indices,
    "a correctly filtered punctuation token must add no index of its own"
  );

  let tf = 1.0 * (1.2 + 1.0) / (1.0 + 1.2 * (1.0 - 0.75 + 0.75 * 1.0 / 150.0));
  assert!(
    (control.values[0] - tf).abs() < 1e-5,
    "{} vs {tf}",
    control.values[0]
  );
}

#[test]
fn an_empty_row_produces_an_empty_embedding() {
  let vocab = vocab(&["cat"]);
  let stopwords = stopwords(&[]);
  let tokens = token_text(&["cat"]);
  let matrix = Tensor::from_vec(vec![0f32; 16], (2, 2, 4), &Device::Cpu).unwrap();
  let hidden_row: Vec<Vec<f32>> = vec![];

  let embedding = pooling::pool_row(
    &hidden_row,
    &[],
    &tokens,
    &vocab,
    &stopwords,
    &matrix,
    false,
  )
  .unwrap();

  assert!(embedding.is_empty());
}