svod-model 0.1.0-alpha.3

Pretrained models inference abstraction.
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

use std::path::Path;
use std::time::{SystemTime, UNIX_EPOCH};

use svod_arch::ctc::CtcDecoder;

use crate::gigaam::{ConvNormType, GigaAmConfig, SubsamplingMode};

#[test]
fn test_config_from_json() {
    let config = GigaAmConfig::from_json(Path::new("tests/gigaam/ctc_config.json")).unwrap();

    assert_eq!(config.n_mels, 64);
    assert_eq!(config.max_batch_size, 32);
    assert_eq!(config.d_model, 768);
    assert_eq!(config.n_heads, 16);
    assert_eq!(config.n_layers, 16);
    assert_eq!(config.d_ff, 3072);
    assert_eq!(config.conv_kernel, 5);
    assert_eq!(config.vocab_size, 34);
    assert_eq!(config.sample_rate, 16000);
    assert_eq!(config.n_fft, 320);
    assert_eq!(config.hop_length, 160);
    assert_eq!(config.win_length, 320);
    assert!(!config.mel_center);
    assert!(matches!(config.subsampling_mode, SubsamplingMode::Conv1d));
    assert!(matches!(config.conv_norm_type, ConvNormType::LayerNorm));
    assert_eq!(config.subs_kernel_size, 5);
    assert_eq!(config.subsampling_factor, 4);
    assert_eq!(config.max_encoder_frames, 5000);
    // No explicit `max_mel_frames` / `max_seq_len` in this config — the loader
    // derives the pre-subsampling bound from the post-subsampling encoder bound
    // by scaling up by `subsampling_factor`, so audio approaching the encoder
    // cap isn't rejected at the JIT input stage.
    assert_eq!(config.max_mel_frames, 5000 * config.subsampling_factor);

    // CTC decoder built from the `decoding` section: 33 Russian glyphs + blank,
    // for `vocab_size = 34` total classes.
    assert!(matches!(config.decoder, CtcDecoder::Greedy(_)));
    assert_eq!(config.decoder.vocabulary().len(), 33);
    assert_eq!(config.decoder.blank_id(), 33);
    assert_eq!(config.decoder.total_vocab(), config.vocab_size);
    assert_eq!(config.decoder.vocabulary()[0], " ");
    assert_eq!(config.decoder.vocabulary()[32], "я");
}

#[test]
fn test_rnnt_config_from_json() {
    let config = GigaAmConfig::from_json(Path::new("tests/gigaam/rnnt_config.json")).unwrap();
    let transducer = config.transducer.as_ref().expect("rnnt transducer config");

    assert_eq!(config.vocab_size, 1025);
    assert_eq!(transducer.pred_hidden, 320);
    assert_eq!(transducer.pred_rnn_layers, 1);
    assert_eq!(transducer.joint_hidden, 320);
    assert_eq!(transducer.num_classes, 1025);
    assert_eq!(transducer.max_symbols_per_step, 10);
    assert!(transducer.vocabulary.is_empty());
    assert!(transducer.sentencepiece);
}

#[test]
fn test_config_rejects_unsupported_attention() {
    let err = expect_config_err(parse_temp_config(&minimal_config("rel_pos", 4, "htk", "null", 8, 8)));
    assert!(err.to_string().contains("self_attention_model"));
}

#[test]
fn test_config_rejects_unsupported_subsampling_factor() {
    let err = expect_config_err(parse_temp_config(&minimal_config("rotary", 2, "htk", "null", 8, 8)));
    assert!(err.to_string().contains("subsampling_factor"));
}

#[test]
fn test_config_rejects_unsupported_mel_frontend() {
    let err = expect_config_err(parse_temp_config(&minimal_config("rotary", 4, "slaney", "null", 8, 8)));
    assert!(err.to_string().contains("mel_scale"));

    let err = expect_config_err(parse_temp_config(&minimal_config("rotary", 4, "htk", r#""slaney""#, 8, 8)));
    assert!(err.to_string().contains("mel_norm"));

    let err = expect_config_err(parse_temp_config(&minimal_config("rotary", 4, "htk", "null", 16, 8)));
    assert!(err.to_string().contains("n_fft"));
}

#[test]
fn test_config_rejects_mel_bound_exceeding_rope_cache() {
    let json = minimal_config("rotary", 4, "htk", "null", 8, 8)
        .replace(r#""pos_emb_max_len": 8,"#, r#""pos_emb_max_len": 8, "max_mel_frames": 40,"#);
    let err = expect_config_err(parse_temp_config(&json));
    assert!(err.to_string().contains("max_mel_frames"));
}

fn expect_config_err(result: Result<GigaAmConfig, crate::gigaam::Error>) -> crate::gigaam::Error {
    match result {
        Ok(_) => panic!("expected config error"),
        Err(err) => err,
    }
}

fn parse_temp_config(json: &str) -> Result<GigaAmConfig, crate::gigaam::Error> {
    let stamp = SystemTime::now().duration_since(UNIX_EPOCH).unwrap().as_nanos();
    let path = std::env::temp_dir().join(format!("svod_gigaam_config_{stamp}.json"));
    std::fs::write(&path, json).unwrap();
    let out = GigaAmConfig::from_json(&path);
    let _ = std::fs::remove_file(path);
    out
}

fn minimal_config(
    self_attention_model: &str,
    subsampling_factor: usize,
    mel_scale: &str,
    mel_norm: &str,
    n_fft: usize,
    win_length: usize,
) -> String {
    format!(
        r#"{{
  "cfg": {{
    "model": {{
      "cfg": {{
        "preprocessor": {{
          "sample_rate": 16000,
          "features": 4,
          "win_length": {win_length},
          "hop_length": 4,
          "mel_scale": "{mel_scale}",
          "n_fft": {n_fft},
          "mel_norm": {mel_norm},
          "center": false
        }},
        "encoder": {{
          "d_model": 8,
          "n_layers": 1,
          "subsampling": "conv1d",
          "subs_kernel_size": 5,
          "subsampling_factor": {subsampling_factor},
          "ff_expansion_factor": 4,
          "self_attention_model": "{self_attention_model}",
          "pos_emb_max_len": 8,
          "n_heads": 2,
          "conv_kernel_size": 5,
          "conv_norm_type": "layer_norm"
        }},
        "head": {{ "num_classes": 2 }},
        "decoding": {{ "_target_": "CTCGreedyDecoding", "vocabulary": ["a"] }}
      }}
    }}
  }}
}}"#
    )
}