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
154

//! Fixed bidirectional ONNX LSTM with IOFC gates

use ndarray::{Array2, ArrayView2, s};

use super::super::gemm::matmul;
use crate::inference::native_model::segmentation::{HIDDEN, LstmLayer};
use crate::inference::{InferenceError, TensorShapeError};

pub(super) struct Direction {
    w: Array2<f32>,
    r: Array2<f32>,
    bias: Vec<f32>,
}

pub(super) struct Lstm([Direction; 2]);

pub(super) struct LstmScratch {
    projection: Array2<f32>,
    hidden: Array2<f32>,
    cell: Vec<f32>,
    gates: Array2<f32>,
}

impl LstmScratch {
    pub(super) fn new() -> Self {
        Self {
            projection: Array2::zeros((589, 4 * HIDDEN)),
            hidden: Array2::zeros((1, HIDDEN)),
            cell: vec![0.0; HIDDEN],
            gates: Array2::zeros((1, 4 * HIDDEN)),
        }
    }
}

impl Lstm {
    pub(super) fn new(layer: LstmLayer) -> Result<Self, TensorShapeError> {
        for (actual, expected) in [
            (layer.w.len(), 2 * 4 * HIDDEN * layer.input),
            (layer.r.len(), 2 * 4 * HIDDEN * HIDDEN),
            (layer.b.len(), 2 * 8 * HIDDEN),
        ] {
            if actual != expected {
                return Err(TensorShapeError::LengthMismatch {
                    context: "CPU recurrent weights",
                    expected,
                    actual,
                });
            }
        }
        let direction = |direction: usize| -> Result<Direction, TensorShapeError> {
            let w_len = 4 * HIDDEN * layer.input;
            let r_len = 4 * HIDDEN * HIDDEN;
            let b = &layer.b[direction * 8 * HIDDEN..(direction + 1) * 8 * HIDDEN];
            Ok(Direction {
                w: array(
                    (4 * HIDDEN, layer.input),
                    layer.w[direction * w_len..(direction + 1) * w_len].to_vec(),
                )?
                .t()
                .as_standard_layout()
                .to_owned(),
                r: array(
                    (4 * HIDDEN, HIDDEN),
                    layer.r[direction * r_len..(direction + 1) * r_len].to_vec(),
                )?
                .t()
                .as_standard_layout()
                .to_owned(),
                bias: (0..4 * HIDDEN)
                    .map(|index| b[index] + b[index + 4 * HIDDEN])
                    .collect(),
            })
        };
        Ok(Self([direction(0)?, direction(1)?]))
    }

    pub(super) fn forward(
        &self,
        input: ArrayView2<'_, f32>,
        output: &mut Array2<f32>,
        scratch: &mut LstmScratch,
    ) -> Result<(), InferenceError> {
        let expected = [589, self.0[0].w.nrows()];
        if input.shape() != expected || output.dim() != (589, 2 * HIDDEN) {
            return Err(crate::inference::TensorShapeError::ShapeMismatch {
                context: "CPU recurrent input/output",
                expected: expected.to_vec(),
                actual: input.shape().to_vec(),
            }
            .into());
        }
        for (direction, weights) in self.0.iter().enumerate() {
            matmul(input, weights.w.view(), scratch.projection.view_mut(), 0.0)?;
            scratch.hidden.fill(0.0);
            scratch.cell.fill(0.0);
            for step in 0..input.nrows() {
                let time = if direction == 0 {
                    step
                } else {
                    input.nrows() - 1 - step
                };
                matmul(
                    scratch.hidden.view(),
                    weights.r.view(),
                    scratch.gates.view_mut(),
                    0.0,
                )?;
                for gate in 0..4 * HIDDEN {
                    scratch.gates[[0, gate]] +=
                        scratch.projection[[time, gate]] + weights.bias[gate];
                }
                for unit in 0..HIDDEN {
                    let (hidden, cell) = step_iofc(
                        [
                            scratch.gates[[0, unit]],
                            scratch.gates[[0, HIDDEN + unit]],
                            scratch.gates[[0, 2 * HIDDEN + unit]],
                            scratch.gates[[0, 3 * HIDDEN + unit]],
                        ],
                        scratch.cell[unit],
                    );
                    scratch.cell[unit] = cell;
                    scratch.hidden[[0, unit]] = hidden;
                }
                output
                    .slice_mut(s![time, direction * HIDDEN..(direction + 1) * HIDDEN])
                    .assign(&scratch.hidden.row(0));
            }
        }
        Ok(())
    }
}

fn step_iofc(gates: [f32; 4], cell: f32) -> (f32, f32) {
    let sigmoid = |value: f32| 1.0 / (1.0 + (-value).exp());
    let [i, o, f, c] = gates;
    let cell = sigmoid(f) * cell + sigmoid(i) * c.tanh();
    (sigmoid(o) * cell.tanh(), cell)
}

pub(super) fn array(
    shape: (usize, usize),
    values: Vec<f32>,
) -> Result<Array2<f32>, TensorShapeError> {
    let actual = values.len();
    Array2::from_shape_vec(shape, values).map_err(|_| TensorShapeError::ShapeMismatch {
        context: "CPU packed matrix",
        expected: vec![shape.0, shape.1],
        actual: vec![actual],
    })
}

#[cfg(test)]
mod tests;