opus1 0.3.0

Pure-Rust implementation of the Opus audio codec (RFC 6716), following libopus 1.x
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

//! PitchDNN, the pitch estimator of deep packet loss concealment (deep PLC):
//! its layers and the constants of its C header.
//!
//! Generated by `tools/dnn-gen` from libopus 1.6.1 `dnn/pitchdnn_data.h` and
//! `dnn/pitchdnn_data.c` (`init_pitchdnn`). Do not edit: change the generator
//! and run it again (`tools/dnn-gen/README.md`).

use crate::Error;
use super::{Conv2dSpec, Layer, LayerSpec, LinearSpec};
use crate::dnn::nnet::{Conv2dLayer, LinearLayer};
use crate::dnn::weights::WeightList;

/// The file of this model's weights in `dnn-weights/`.
pub const BLOB_FILE: &str = "pitchdnn.bin";

// The constants of `dnn/pitchdnn_data.h`, in the same order.

/// `DENSE_IF_UPSAMPLER_1_OUT_SIZE`.
pub const DENSE_IF_UPSAMPLER_1_OUT_SIZE: usize = 64;

/// `DENSE_IF_UPSAMPLER_2_OUT_SIZE`.
pub const DENSE_IF_UPSAMPLER_2_OUT_SIZE: usize = 64;

/// `DENSE_DOWNSAMPLER_OUT_SIZE`.
pub const DENSE_DOWNSAMPLER_OUT_SIZE: usize = 64;

/// `DENSE_FINAL_UPSAMPLER_OUT_SIZE`.
pub const DENSE_FINAL_UPSAMPLER_OUT_SIZE: usize = 192;

/// `GRU_1_OUT_SIZE`.
pub const GRU_1_OUT_SIZE: usize = 64;

/// `GRU_1_STATE_SIZE`.
pub const GRU_1_STATE_SIZE: usize = 64;

/// `PITCH_DNN_MAX_RNN_UNITS`.
pub const PITCH_DNN_MAX_RNN_UNITS: usize = 64;

/// The layers in the order of `init_pitchdnn`: the arrays that each one reads,
/// and its shape. The names are the array names of the blob.
pub const LAYERS: [LayerSpec; 8] = [
    LayerSpec::Linear(LinearSpec {
        name: "dense_if_upsampler_1",
        bias: Some("dense_if_upsampler_1_bias"),
        subias: Some("dense_if_upsampler_1_subias"),
        weights: Some("dense_if_upsampler_1_weights_int8"),
        float_weights: Some("dense_if_upsampler_1_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("dense_if_upsampler_1_scale"),
        nb_inputs: 88,
        nb_outputs: 64,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "dense_if_upsampler_2",
        bias: Some("dense_if_upsampler_2_bias"),
        subias: Some("dense_if_upsampler_2_subias"),
        weights: Some("dense_if_upsampler_2_weights_int8"),
        float_weights: Some("dense_if_upsampler_2_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("dense_if_upsampler_2_scale"),
        nb_inputs: 64,
        nb_outputs: 64,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "dense_downsampler",
        bias: Some("dense_downsampler_bias"),
        subias: Some("dense_downsampler_subias"),
        weights: Some("dense_downsampler_weights_int8"),
        float_weights: Some("dense_downsampler_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("dense_downsampler_scale"),
        nb_inputs: 288,
        nb_outputs: 64,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "dense_final_upsampler",
        bias: Some("dense_final_upsampler_bias"),
        subias: Some("dense_final_upsampler_subias"),
        weights: Some("dense_final_upsampler_weights_int8"),
        float_weights: Some("dense_final_upsampler_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("dense_final_upsampler_scale"),
        nb_inputs: 64,
        nb_outputs: 192,
    }),
    LayerSpec::Conv2d(Conv2dSpec {
        name: "conv2d_1",
        bias: Some("conv2d_1_bias"),
        float_weights: Some("conv2d_1_weight_float"),
        in_channels: 1,
        out_channels: 4,
        ktime: 3,
        kheight: 3,
    }),
    LayerSpec::Conv2d(Conv2dSpec {
        name: "conv2d_2",
        bias: Some("conv2d_2_bias"),
        float_weights: Some("conv2d_2_weight_float"),
        in_channels: 4,
        out_channels: 1,
        ktime: 3,
        kheight: 3,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "gru_1_input",
        bias: Some("gru_1_input_bias"),
        subias: Some("gru_1_input_subias"),
        weights: Some("gru_1_input_weights_int8"),
        float_weights: Some("gru_1_input_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("gru_1_input_scale"),
        nb_inputs: 64,
        nb_outputs: 192,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "gru_1_recurrent",
        bias: Some("gru_1_recurrent_bias"),
        subias: Some("gru_1_recurrent_subias"),
        weights: Some("gru_1_recurrent_weights_int8"),
        float_weights: Some("gru_1_recurrent_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("gru_1_recurrent_scale"),
        nb_inputs: 64,
        nb_outputs: 192,
    }),
];

/// libopus `PitchDNN` (`dnn/pitchdnn_data.h`): the layers of the model, in the
/// order of the C struct. Each layer borrows its arrays from a weight list.
#[derive(Clone, Copy, Default)]
pub struct PitchDnn<'w> {
    /// `dense_if_upsampler_1`: int8 weights, 88 inputs, 64 outputs.
    pub dense_if_upsampler_1: LinearLayer<'w>,
    /// `dense_if_upsampler_2`: int8 weights, 64 inputs, 64 outputs.
    pub dense_if_upsampler_2: LinearLayer<'w>,
    /// `dense_downsampler`: int8 weights, 288 inputs, 64 outputs.
    pub dense_downsampler: LinearLayer<'w>,
    /// `dense_final_upsampler`: int8 weights, 64 inputs, 192 outputs.
    pub dense_final_upsampler: LinearLayer<'w>,
    /// `conv2d_1`: conv2d, 1 input channel and 4 output channels, kernel 3 × 3
    /// (time × height).
    pub conv2d_1: Conv2dLayer<'w>,
    /// `conv2d_2`: conv2d, 4 input channels and 1 output channel, kernel 3 × 3
    /// (time × height).
    pub conv2d_2: Conv2dLayer<'w>,
    /// `gru_1_input`: int8 weights, 64 inputs, 192 outputs.
    pub gru_1_input: LinearLayer<'w>,
    /// `gru_1_recurrent`: int8 weights, 64 inputs, 192 outputs.
    pub gru_1_recurrent: LinearLayer<'w>,
}

impl<'w> PitchDnn<'w> {
    /// libopus `init_pitchdnn`: builds every layer from the arrays of `list`,
    /// in the order of `LAYERS`. Fails at the first layer whose arrays are
    /// missing, have the wrong size or are malformed; then nothing is built.
    pub fn init(list: &WeightList<'w>) -> Result<Self, Error> {
        Ok(PitchDnn {
            dense_if_upsampler_1: LAYERS[0].linear(list)?,
            dense_if_upsampler_2: LAYERS[1].linear(list)?,
            dense_downsampler: LAYERS[2].linear(list)?,
            dense_final_upsampler: LAYERS[3].linear(list)?,
            conv2d_1: LAYERS[4].conv2d(list)?,
            conv2d_2: LAYERS[5].conv2d(list)?,
            gru_1_input: LAYERS[6].linear(list)?,
            gru_1_recurrent: LAYERS[7].linear(list)?,
        })
    }

    /// Every layer, in the order of `LAYERS`.
    pub fn layers(&self) -> [Layer<'_, 'w>; 8] {
        [
            Layer::Linear(&self.dense_if_upsampler_1),
            Layer::Linear(&self.dense_if_upsampler_2),
            Layer::Linear(&self.dense_downsampler),
            Layer::Linear(&self.dense_final_upsampler),
            Layer::Conv2d(&self.conv2d_1),
            Layer::Conv2d(&self.conv2d_2),
            Layer::Linear(&self.gru_1_input),
            Layer::Linear(&self.gru_1_recurrent),
        ]
    }
}