speakrs 0.6.0

Fast Rust speaker diarization with pyannote-level accuracy and native CPU/CoreML/CUDA inference
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

use std::num::NonZeroUsize;

/// Model batch sizes supported by a loaded segmentation backend
#[derive(Clone, Copy)]
pub(super) enum Batching {
    #[cfg(any(feature = "migraphx", test))]
    Single,
    #[cfg(any(feature = "cpu", test))]
    Useful(NonZeroUsize),
    #[cfg(any(feature = "migraphx", feature = "coreml", feature = "_cuda", test))]
    Fixed(NonZeroUsize),
}

/// Output delivery whose existing fixed-tail behavior must be preserved
#[derive(Clone, Copy)]
pub(super) enum Delivery {
    Collected,
    Streaming,
}

/// A nonempty inference call, including only the outputs useful to the caller
#[derive(Clone, Copy)]
pub(super) struct BatchPlan {
    useful: NonZeroUsize,
    model: NonZeroUsize,
}

impl Batching {
    pub(super) fn plan(self, remaining: NonZeroUsize, _delivery: Delivery) -> BatchPlan {
        match self {
            #[cfg(any(feature = "migraphx", test))]
            Self::Single => BatchPlan {
                useful: NonZeroUsize::MIN,
                model: NonZeroUsize::MIN,
            },
            #[cfg(any(feature = "cpu", test))]
            Self::Useful(capacity) => {
                let useful = remaining.min(capacity);
                BatchPlan {
                    useful,
                    model: useful,
                }
            }
            #[cfg(any(feature = "migraphx", feature = "coreml", feature = "_cuda", test))]
            Self::Fixed(capacity) => {
                let one = NonZeroUsize::MIN;
                if remaining >= capacity {
                    return BatchPlan {
                        useful: capacity,
                        model: capacity,
                    };
                }

                // collected accelerator tails use the single-window model; streaming
                // tails use fixed padding, which can produce different scores
                if matches!(_delivery, Delivery::Streaming) && remaining > one {
                    return BatchPlan {
                        useful: remaining,
                        model: capacity,
                    };
                }

                BatchPlan {
                    useful: one,
                    model: one,
                }
            }
        }
    }
}

impl BatchPlan {
    pub(super) fn useful(self) -> usize {
        self.useful.get()
    }

    pub(super) fn model(self) -> usize {
        self.model.get()
    }

    pub(super) fn is_single(self) -> bool {
        self.model == NonZeroUsize::MIN
    }
}

#[cfg(test)]
mod tests {
    use super::{Batching, Delivery};
    use std::num::NonZeroUsize;

    fn calls(batching: Batching, delivery: Delivery, total: usize) -> Vec<(usize, usize, usize)> {
        let mut calls = Vec::new();
        let mut next = 0;
        while let Some(remaining) = NonZeroUsize::new(total - next) {
            let plan = batching.plan(remaining, delivery);
            calls.push((next, plan.useful(), plan.model()));
            next += plan.useful();
        }
        calls
    }

    #[test]
    fn collected_fixed_tails_keep_single_window_inference() {
        let fixed = Batching::Fixed(NonZeroUsize::new(32).unwrap());
        assert_eq!(calls(fixed, Delivery::Collected, 0), []);
        assert_eq!(calls(fixed, Delivery::Collected, 32), [(0, 32, 32)]);
        assert_eq!(
            calls(fixed, Delivery::Collected, 35),
            [(0, 32, 32), (32, 1, 1), (33, 1, 1), (34, 1, 1)]
        );
    }

    #[test]
    fn streaming_fixed_tails_pad_only_multi_window_calls() {
        let fixed = Batching::Fixed(NonZeroUsize::new(32).unwrap());
        assert_eq!(calls(fixed, Delivery::Streaming, 1), [(0, 1, 1)]);
        assert_eq!(calls(fixed, Delivery::Streaming, 31), [(0, 31, 32)]);
        assert_eq!(
            calls(fixed, Delivery::Streaming, 33),
            [(0, 32, 32), (32, 1, 1)]
        );
        assert_eq!(
            calls(fixed, Delivery::Streaming, 35),
            [(0, 32, 32), (32, 3, 32)]
        );
    }

    #[test]
    fn useful_batches_keep_partial_calls_unpadded_for_both_deliveries() {
        let useful = Batching::Useful(NonZeroUsize::new(8).unwrap());
        for delivery in [Delivery::Collected, Delivery::Streaming] {
            assert_eq!(calls(useful, delivery, 0), []);
            assert_eq!(calls(useful, delivery, 1), [(0, 1, 1)]);
            assert_eq!(calls(useful, delivery, 7), [(0, 7, 7)]);
            assert_eq!(calls(useful, delivery, 8), [(0, 8, 8)]);
            assert_eq!(calls(useful, delivery, 9), [(0, 8, 8), (8, 1, 1)]);
            assert_eq!(
                calls(useful, delivery, 19),
                [(0, 8, 8), (8, 8, 8), (16, 3, 3)]
            );
        }
    }

    #[test]
    fn single_window_backend_never_pads_or_groups_inputs() {
        for delivery in [Delivery::Collected, Delivery::Streaming] {
            assert_eq!(
                calls(Batching::Single, delivery, 3),
                [(0, 1, 1), (1, 1, 1), (2, 1, 1)]
            );
        }
    }
}