opus1 0.2.1

//! FARGAN, the vocoder 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/fargan_data.h` and
//! `dnn/fargan_data.c` (`init_fargan`). Do not edit: change the generator and
//! run it again (`tools/dnn-gen/README.md`).

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

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

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

/// `COND_NET_PEMBED_OUT_SIZE`.
pub const COND_NET_PEMBED_OUT_SIZE: usize = 12;

/// `COND_NET_FDENSE1_OUT_SIZE`.
pub const COND_NET_FDENSE1_OUT_SIZE: usize = 64;

/// `COND_NET_FCONV1_OUT_SIZE`.
pub const COND_NET_FCONV1_OUT_SIZE: usize = 128;

/// `COND_NET_FCONV1_IN_SIZE`.
pub const COND_NET_FCONV1_IN_SIZE: usize = 64;

/// `COND_NET_FCONV1_STATE_SIZE`, `(64 * (2))` in C.
pub const COND_NET_FCONV1_STATE_SIZE: usize = 128;

/// `COND_NET_FCONV1_DELAY`.
pub const COND_NET_FCONV1_DELAY: usize = 1;

/// `COND_NET_FDENSE2_OUT_SIZE`.
pub const COND_NET_FDENSE2_OUT_SIZE: usize = 320;

/// `SIG_NET_COND_GAIN_DENSE_OUT_SIZE`.
pub const SIG_NET_COND_GAIN_DENSE_OUT_SIZE: usize = 1;

/// `SIG_NET_FWC0_CONV_OUT_SIZE`.
pub const SIG_NET_FWC0_CONV_OUT_SIZE: usize = 192;

/// `SIG_NET_FWC0_GLU_GATE_OUT_SIZE`.
pub const SIG_NET_FWC0_GLU_GATE_OUT_SIZE: usize = 192;

/// `SIG_NET_GRU1_OUT_SIZE`.
pub const SIG_NET_GRU1_OUT_SIZE: usize = 160;

/// `SIG_NET_GRU1_STATE_SIZE`.
pub const SIG_NET_GRU1_STATE_SIZE: usize = 160;

/// `SIG_NET_GRU2_OUT_SIZE`.
pub const SIG_NET_GRU2_OUT_SIZE: usize = 128;

/// `SIG_NET_GRU2_STATE_SIZE`.
pub const SIG_NET_GRU2_STATE_SIZE: usize = 128;

/// `SIG_NET_GRU3_OUT_SIZE`.
pub const SIG_NET_GRU3_OUT_SIZE: usize = 128;

/// `SIG_NET_GRU3_STATE_SIZE`.
pub const SIG_NET_GRU3_STATE_SIZE: usize = 128;

/// `SIG_NET_GRU1_GLU_GATE_OUT_SIZE`.
pub const SIG_NET_GRU1_GLU_GATE_OUT_SIZE: usize = 160;

/// `SIG_NET_GRU2_GLU_GATE_OUT_SIZE`.
pub const SIG_NET_GRU2_GLU_GATE_OUT_SIZE: usize = 128;

/// `SIG_NET_GRU3_GLU_GATE_OUT_SIZE`.
pub const SIG_NET_GRU3_GLU_GATE_OUT_SIZE: usize = 128;

/// `SIG_NET_SKIP_GLU_GATE_OUT_SIZE`.
pub const SIG_NET_SKIP_GLU_GATE_OUT_SIZE: usize = 128;

/// `SIG_NET_SKIP_DENSE_OUT_SIZE`.
pub const SIG_NET_SKIP_DENSE_OUT_SIZE: usize = 128;

/// `SIG_NET_SIG_DENSE_OUT_OUT_SIZE`.
pub const SIG_NET_SIG_DENSE_OUT_OUT_SIZE: usize = 40;

/// `SIG_NET_GAIN_DENSE_OUT_OUT_SIZE`.
pub const SIG_NET_GAIN_DENSE_OUT_OUT_SIZE: usize = 4;

/// The layers in the order of `init_fargan`: the arrays that each one reads,
/// and its shape. The names are the array names of the blob.
pub const LAYERS: [LayerSpec; 20] = [
    LayerSpec::Linear(LinearSpec {
        name: "cond_net_pembed",
        bias: Some("cond_net_pembed_bias"),
        subias: None,
        weights: None,
        float_weights: Some("cond_net_pembed_weights_float"),
        weights_idx: None,
        diag: None,
        scale: None,
        nb_inputs: 224,
        nb_outputs: 12,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "cond_net_fdense1",
        bias: Some("cond_net_fdense1_bias"),
        subias: None,
        weights: None,
        float_weights: Some("cond_net_fdense1_weights_float"),
        weights_idx: None,
        diag: None,
        scale: None,
        nb_inputs: 32,
        nb_outputs: 64,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "cond_net_fconv1",
        bias: Some("cond_net_fconv1_bias"),
        subias: Some("cond_net_fconv1_subias"),
        weights: Some("cond_net_fconv1_weights_int8"),
        float_weights: Some("cond_net_fconv1_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("cond_net_fconv1_scale"),
        nb_inputs: 192,
        nb_outputs: 128,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "cond_net_fdense2",
        bias: Some("cond_net_fdense2_bias"),
        subias: Some("cond_net_fdense2_subias"),
        weights: Some("cond_net_fdense2_weights_int8"),
        float_weights: Some("cond_net_fdense2_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("cond_net_fdense2_scale"),
        nb_inputs: 128,
        nb_outputs: 320,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_cond_gain_dense",
        bias: Some("sig_net_cond_gain_dense_bias"),
        subias: None,
        weights: None,
        float_weights: Some("sig_net_cond_gain_dense_weights_float"),
        weights_idx: None,
        diag: None,
        scale: None,
        nb_inputs: 80,
        nb_outputs: 1,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_fwc0_conv",
        bias: Some("sig_net_fwc0_conv_bias"),
        subias: Some("sig_net_fwc0_conv_subias"),
        weights: Some("sig_net_fwc0_conv_weights_int8"),
        float_weights: Some("sig_net_fwc0_conv_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_fwc0_conv_scale"),
        nb_inputs: 328,
        nb_outputs: 192,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_fwc0_glu_gate",
        bias: Some("sig_net_fwc0_glu_gate_bias"),
        subias: Some("sig_net_fwc0_glu_gate_subias"),
        weights: Some("sig_net_fwc0_glu_gate_weights_int8"),
        float_weights: Some("sig_net_fwc0_glu_gate_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_fwc0_glu_gate_scale"),
        nb_inputs: 192,
        nb_outputs: 192,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru1_input",
        bias: None,
        subias: Some("sig_net_gru1_input_subias"),
        weights: Some("sig_net_gru1_input_weights_int8"),
        float_weights: Some("sig_net_gru1_input_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru1_input_scale"),
        nb_inputs: 272,
        nb_outputs: 480,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru1_recurrent",
        bias: None,
        subias: Some("sig_net_gru1_recurrent_subias"),
        weights: Some("sig_net_gru1_recurrent_weights_int8"),
        float_weights: Some("sig_net_gru1_recurrent_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru1_recurrent_scale"),
        nb_inputs: 160,
        nb_outputs: 480,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru2_input",
        bias: None,
        subias: Some("sig_net_gru2_input_subias"),
        weights: Some("sig_net_gru2_input_weights_int8"),
        float_weights: Some("sig_net_gru2_input_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru2_input_scale"),
        nb_inputs: 240,
        nb_outputs: 384,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru2_recurrent",
        bias: None,
        subias: Some("sig_net_gru2_recurrent_subias"),
        weights: Some("sig_net_gru2_recurrent_weights_int8"),
        float_weights: Some("sig_net_gru2_recurrent_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru2_recurrent_scale"),
        nb_inputs: 128,
        nb_outputs: 384,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru3_input",
        bias: None,
        subias: Some("sig_net_gru3_input_subias"),
        weights: Some("sig_net_gru3_input_weights_int8"),
        float_weights: Some("sig_net_gru3_input_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru3_input_scale"),
        nb_inputs: 208,
        nb_outputs: 384,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru3_recurrent",
        bias: None,
        subias: Some("sig_net_gru3_recurrent_subias"),
        weights: Some("sig_net_gru3_recurrent_weights_int8"),
        float_weights: Some("sig_net_gru3_recurrent_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru3_recurrent_scale"),
        nb_inputs: 128,
        nb_outputs: 384,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru1_glu_gate",
        bias: Some("sig_net_gru1_glu_gate_bias"),
        subias: Some("sig_net_gru1_glu_gate_subias"),
        weights: Some("sig_net_gru1_glu_gate_weights_int8"),
        float_weights: Some("sig_net_gru1_glu_gate_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru1_glu_gate_scale"),
        nb_inputs: 160,
        nb_outputs: 160,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru2_glu_gate",
        bias: Some("sig_net_gru2_glu_gate_bias"),
        subias: Some("sig_net_gru2_glu_gate_subias"),
        weights: Some("sig_net_gru2_glu_gate_weights_int8"),
        float_weights: Some("sig_net_gru2_glu_gate_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru2_glu_gate_scale"),
        nb_inputs: 128,
        nb_outputs: 128,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gru3_glu_gate",
        bias: Some("sig_net_gru3_glu_gate_bias"),
        subias: Some("sig_net_gru3_glu_gate_subias"),
        weights: Some("sig_net_gru3_glu_gate_weights_int8"),
        float_weights: Some("sig_net_gru3_glu_gate_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_gru3_glu_gate_scale"),
        nb_inputs: 128,
        nb_outputs: 128,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_skip_glu_gate",
        bias: Some("sig_net_skip_glu_gate_bias"),
        subias: Some("sig_net_skip_glu_gate_subias"),
        weights: Some("sig_net_skip_glu_gate_weights_int8"),
        float_weights: Some("sig_net_skip_glu_gate_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_skip_glu_gate_scale"),
        nb_inputs: 128,
        nb_outputs: 128,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_skip_dense",
        bias: Some("sig_net_skip_dense_bias"),
        subias: Some("sig_net_skip_dense_subias"),
        weights: Some("sig_net_skip_dense_weights_int8"),
        float_weights: Some("sig_net_skip_dense_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_skip_dense_scale"),
        nb_inputs: 688,
        nb_outputs: 128,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_sig_dense_out",
        bias: Some("sig_net_sig_dense_out_bias"),
        subias: Some("sig_net_sig_dense_out_subias"),
        weights: Some("sig_net_sig_dense_out_weights_int8"),
        float_weights: Some("sig_net_sig_dense_out_weights_float"),
        weights_idx: None,
        diag: None,
        scale: Some("sig_net_sig_dense_out_scale"),
        nb_inputs: 128,
        nb_outputs: 40,
    }),
    LayerSpec::Linear(LinearSpec {
        name: "sig_net_gain_dense_out",
        bias: Some("sig_net_gain_dense_out_bias"),
        subias: None,
        weights: None,
        float_weights: Some("sig_net_gain_dense_out_weights_float"),
        weights_idx: None,
        diag: None,
        scale: None,
        nb_inputs: 192,
        nb_outputs: 4,
    }),
];

/// libopus `FARGAN` (`dnn/fargan_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 Fargan<'w> {
    /// `cond_net_pembed`: float weights, 224 inputs, 12 outputs.
    pub cond_net_pembed: LinearLayer<'w>,
    /// `cond_net_fdense1`: float weights, 32 inputs, 64 outputs.
    pub cond_net_fdense1: LinearLayer<'w>,
    /// `cond_net_fconv1`: int8 weights, 192 inputs, 128 outputs.
    pub cond_net_fconv1: LinearLayer<'w>,
    /// `cond_net_fdense2`: int8 weights, 128 inputs, 320 outputs.
    pub cond_net_fdense2: LinearLayer<'w>,
    /// `sig_net_cond_gain_dense`: float weights, 80 inputs, 1 output.
    pub sig_net_cond_gain_dense: LinearLayer<'w>,
    /// `sig_net_fwc0_conv`: int8 weights, 328 inputs, 192 outputs.
    pub sig_net_fwc0_conv: LinearLayer<'w>,
    /// `sig_net_fwc0_glu_gate`: int8 weights, 192 inputs, 192 outputs.
    pub sig_net_fwc0_glu_gate: LinearLayer<'w>,
    /// `sig_net_gru1_input`: int8 weights, 272 inputs, 480 outputs.
    pub sig_net_gru1_input: LinearLayer<'w>,
    /// `sig_net_gru1_recurrent`: int8 weights, 160 inputs, 480 outputs.
    pub sig_net_gru1_recurrent: LinearLayer<'w>,
    /// `sig_net_gru2_input`: int8 weights, 240 inputs, 384 outputs.
    pub sig_net_gru2_input: LinearLayer<'w>,
    /// `sig_net_gru2_recurrent`: int8 weights, 128 inputs, 384 outputs.
    pub sig_net_gru2_recurrent: LinearLayer<'w>,
    /// `sig_net_gru3_input`: int8 weights, 208 inputs, 384 outputs.
    pub sig_net_gru3_input: LinearLayer<'w>,
    /// `sig_net_gru3_recurrent`: int8 weights, 128 inputs, 384 outputs.
    pub sig_net_gru3_recurrent: LinearLayer<'w>,
    /// `sig_net_gru1_glu_gate`: int8 weights, 160 inputs, 160 outputs.
    pub sig_net_gru1_glu_gate: LinearLayer<'w>,
    /// `sig_net_gru2_glu_gate`: int8 weights, 128 inputs, 128 outputs.
    pub sig_net_gru2_glu_gate: LinearLayer<'w>,
    /// `sig_net_gru3_glu_gate`: int8 weights, 128 inputs, 128 outputs.
    pub sig_net_gru3_glu_gate: LinearLayer<'w>,
    /// `sig_net_skip_glu_gate`: int8 weights, 128 inputs, 128 outputs.
    pub sig_net_skip_glu_gate: LinearLayer<'w>,
    /// `sig_net_skip_dense`: int8 weights, 688 inputs, 128 outputs.
    pub sig_net_skip_dense: LinearLayer<'w>,
    /// `sig_net_sig_dense_out`: int8 weights, 128 inputs, 40 outputs.
    pub sig_net_sig_dense_out: LinearLayer<'w>,
    /// `sig_net_gain_dense_out`: float weights, 192 inputs, 4 outputs.
    pub sig_net_gain_dense_out: LinearLayer<'w>,
}

impl<'w> Fargan<'w> {
    /// libopus `init_fargan`: 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(Fargan {
            cond_net_pembed: LAYERS[0].linear(list)?,
            cond_net_fdense1: LAYERS[1].linear(list)?,
            cond_net_fconv1: LAYERS[2].linear(list)?,
            cond_net_fdense2: LAYERS[3].linear(list)?,
            sig_net_cond_gain_dense: LAYERS[4].linear(list)?,
            sig_net_fwc0_conv: LAYERS[5].linear(list)?,
            sig_net_fwc0_glu_gate: LAYERS[6].linear(list)?,
            sig_net_gru1_input: LAYERS[7].linear(list)?,
            sig_net_gru1_recurrent: LAYERS[8].linear(list)?,
            sig_net_gru2_input: LAYERS[9].linear(list)?,
            sig_net_gru2_recurrent: LAYERS[10].linear(list)?,
            sig_net_gru3_input: LAYERS[11].linear(list)?,
            sig_net_gru3_recurrent: LAYERS[12].linear(list)?,
            sig_net_gru1_glu_gate: LAYERS[13].linear(list)?,
            sig_net_gru2_glu_gate: LAYERS[14].linear(list)?,
            sig_net_gru3_glu_gate: LAYERS[15].linear(list)?,
            sig_net_skip_glu_gate: LAYERS[16].linear(list)?,
            sig_net_skip_dense: LAYERS[17].linear(list)?,
            sig_net_sig_dense_out: LAYERS[18].linear(list)?,
            sig_net_gain_dense_out: LAYERS[19].linear(list)?,
        })
    }

    /// Every layer, in the order of `LAYERS`.
    pub fn layers(&self) -> [Layer<'_, 'w>; 20] {
        [
            Layer::Linear(&self.cond_net_pembed),
            Layer::Linear(&self.cond_net_fdense1),
            Layer::Linear(&self.cond_net_fconv1),
            Layer::Linear(&self.cond_net_fdense2),
            Layer::Linear(&self.sig_net_cond_gain_dense),
            Layer::Linear(&self.sig_net_fwc0_conv),
            Layer::Linear(&self.sig_net_fwc0_glu_gate),
            Layer::Linear(&self.sig_net_gru1_input),
            Layer::Linear(&self.sig_net_gru1_recurrent),
            Layer::Linear(&self.sig_net_gru2_input),
            Layer::Linear(&self.sig_net_gru2_recurrent),
            Layer::Linear(&self.sig_net_gru3_input),
            Layer::Linear(&self.sig_net_gru3_recurrent),
            Layer::Linear(&self.sig_net_gru1_glu_gate),
            Layer::Linear(&self.sig_net_gru2_glu_gate),
            Layer::Linear(&self.sig_net_gru3_glu_gate),
            Layer::Linear(&self.sig_net_skip_glu_gate),
            Layer::Linear(&self.sig_net_skip_dense),
            Layer::Linear(&self.sig_net_sig_dense_out),
            Layer::Linear(&self.sig_net_gain_dense_out),
        ]
    }
}