mako 0.3.0

Mako is the main Sidekick AI data processing library
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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235

use crate::tokenization::hf_tokenizers::processors::byte_level::process_offsets;
use crate::tokenization::hf_tokenizers::tokenizer::{Encoding, PostProcessor, Result};
use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use std::iter::FromIterator;

#[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
#[serde(tag = "type")]
pub struct RobertaProcessing {
    sep: (String, u32),
    cls: (String, u32),
    trim_offsets: bool,
    add_prefix_space: bool,
}

impl Default for RobertaProcessing {
    fn default() -> Self {
        Self {
            sep: ("</s>".into(), 2),
            cls: ("<s>".into(), 0),
            trim_offsets: true,
            add_prefix_space: true,
        }
    }
}

impl RobertaProcessing {
    pub fn new(sep: (String, u32), cls: (String, u32)) -> Self {
        RobertaProcessing {
            sep,
            cls,
            ..Default::default()
        }
    }
    pub fn trim_offsets(mut self, v: bool) -> Self {
        self.trim_offsets = v;
        self
    }
    pub fn add_prefix_space(mut self, v: bool) -> Self {
        self.add_prefix_space = v;
        self
    }
}

impl PostProcessor for RobertaProcessing {
    fn added_tokens(&self, is_pair: bool) -> usize {
        if is_pair {
            4
        } else {
            2
        }
    }

    fn process(
        &self,
        mut encoding: Encoding,
        mut pair_encoding: Option<Encoding>,
        add_special_tokens: bool,
    ) -> Result<Encoding> {
        if self.trim_offsets {
            process_offsets(&mut encoding, self.add_prefix_space);
            encoding
                .get_overflowing_mut()
                .iter_mut()
                .for_each(|mut encoding| process_offsets(&mut encoding, self.add_prefix_space));

            if let Some(mut encoding) = pair_encoding.as_mut() {
                process_offsets(&mut encoding, self.add_prefix_space);
                encoding
                    .get_overflowing_mut()
                    .iter_mut()
                    .for_each(|mut encoding| process_offsets(&mut encoding, self.add_prefix_space));
            }
        }

        if !add_special_tokens {
            return PostProcessor::default_process(encoding, pair_encoding, add_special_tokens);
        }

        let ids = [&[self.cls.1], &encoding.get_ids()[..], &[self.sep.1]].concat();
        let type_ids = [&[0], &encoding.get_type_ids()[..], &[0]].concat();
        let tokens = [
            &[self.cls.0.clone()],
            &encoding.get_tokens()[..],
            &[self.sep.0.clone()],
        ]
        .concat();
        let words = [&[None], &encoding.get_word_ids()[..], &[None]].concat();
        let offsets = [&[(0, 0)], &encoding.get_offsets()[..], &[(0, 0)]].concat();
        let special_tokens = [&[1u32], &vec![0; encoding.get_ids().len()][..], &[1]].concat();
        let attention_mask = vec![1; ids.len()];

        // For compatibility with `TemplateProcessing`, the sequence_ranges shouldn't contain
        // the special tokens.
        let sequence_ranges = HashMap::from_iter(vec![(0, 1..ids.len() - 1)]);
        let mut new_encoding = Encoding::new(
            ids,
            type_ids,
            tokens,
            words,
            offsets,
            special_tokens,
            attention_mask,
            encoding
                .take_overflowing()
                .into_iter()
                .map(|encoding| {
                    let ids = [&[self.cls.1], &encoding.get_ids()[..], &[self.sep.1]].concat();
                    let type_ids = [&[0], &encoding.get_type_ids()[..], &[0]].concat();
                    let tokens = [
                        &[self.cls.0.clone()],
                        &encoding.get_tokens()[..],
                        &[self.sep.0.clone()],
                    ]
                    .concat();
                    let words = [&[None], &encoding.get_word_ids()[..], &[None]].concat();
                    let offsets = [&[(0, 0)], &encoding.get_offsets()[..], &[(0, 0)]].concat();
                    let special_tokens =
                        [&[1u32], &vec![0; encoding.get_ids().len()][..], &[1]].concat();
                    let attention_mask = vec![1; ids.len()];

                    // For compatibility with `TemplateProcessing`, the sequence_ranges shouldn't
                    // contain the special tokens.
                    let sequence_ranges = HashMap::from_iter(vec![(0, 1..ids.len() - 1)]);
                    Encoding::new(
                        ids,
                        type_ids,
                        tokens,
                        words,
                        offsets,
                        special_tokens,
                        attention_mask,
                        vec![],
                        sequence_ranges,
                    )
                })
                .collect(),
            sequence_ranges,
        );

        if let Some(mut encoding) = pair_encoding {
            let pair_ids = [&[self.sep.1], &encoding.get_ids()[..], &[self.sep.1]].concat();
            let pair_type_ids = vec![0; encoding.get_ids().len() + 2];
            let pair_tokens = [
                &[self.sep.0.clone()],
                &encoding.get_tokens()[..],
                &[self.sep.0.clone()],
            ]
            .concat();
            let pair_words = [&[None], &encoding.get_word_ids()[..], &[None]].concat();
            let pair_offsets = [&[(0, 0)], &encoding.get_offsets()[..], &[(0, 0)]].concat();
            let pair_special_tokens =
                [&[1], &vec![0u32; encoding.get_type_ids().len()][..], &[1]].concat();
            let pair_attention_mask = vec![1; pair_ids.len()];

            // For compatibility with `TemplateProcessing`, the sequence_ranges shouldn't contain
            // the special tokens.
            let pair_sequence_ranges = HashMap::from_iter(vec![(1, 1..pair_ids.len() - 1)]);
            let new_pair_encoding = Encoding::new(
                pair_ids,
                pair_type_ids,
                pair_tokens,
                pair_words,
                pair_offsets,
                pair_special_tokens,
                pair_attention_mask,
                encoding
                    .take_overflowing()
                    .into_iter()
                    .map(|encoding| {
                        let pair_ids =
                            [&[self.sep.1], &encoding.get_ids()[..], &[self.sep.1]].concat();
                        let pair_type_ids = vec![0; encoding.get_ids().len() + 2];
                        let pair_tokens = [
                            &[self.sep.0.clone()],
                            &encoding.get_tokens()[..],
                            &[self.sep.0.clone()],
                        ]
                        .concat();
                        let pair_words = [&[None], &encoding.get_word_ids()[..], &[None]].concat();
                        let pair_offsets =
                            [&[(0, 0)], &encoding.get_offsets()[..], &[(0, 0)]].concat();
                        let pair_special_tokens =
                            [&[1], &vec![0u32; encoding.get_type_ids().len()][..], &[1]].concat();
                        let pair_attention_mask = vec![1; pair_ids.len()];

                        // For compatibility with `TemplateProcessing`, the sequence_ranges
                        // shouldn't contain the special tokens.
                        let pair_sequence_ranges =
                            HashMap::from_iter(vec![(1, 1..pair_ids.len() - 1)]);
                        Encoding::new(
                            pair_ids,
                            pair_type_ids,
                            pair_tokens,
                            pair_words,
                            pair_offsets,
                            pair_special_tokens,
                            pair_attention_mask,
                            vec![],
                            pair_sequence_ranges,
                        )
                    })
                    .collect(),
                pair_sequence_ranges,
            );

            new_encoding.merge_with(new_pair_encoding, false);
        }

        Ok(new_encoding)
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn serde() {
        let roberta = RobertaProcessing::default();
        let roberta_r = r#"{
            "type":"RobertaProcessing",
            "sep":["</s>",2],
            "cls":["<s>",0],
            "trim_offsets":true,
            "add_prefix_space":true
        }"#
        .replace(char::is_whitespace, "");
        assert_eq!(serde_json::to_string(&roberta).unwrap(), roberta_r);
        assert_eq!(
            serde_json::from_str::<RobertaProcessing>(&roberta_r).unwrap(),
            roberta
        );
    }
}