dspk-synapse-ml 0.0.9

Pretrained neuro models (DARTsort, SpikeNet2, UnitRefine) and their runtime for dspk, each with its provenance
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

//! DARTsort pretrained single-channel waveform denoiser (`dartsort/singlechan-denoiser-v1`).

use std::path::Path;

use dspk_core::{ComputeTarget, DspError, DspResult};
use dspk_synapse::{SnippetBatch, WaveformDenoiser};

use crate::manifest::{SafetensorsMap, TensorIoSpec};
use crate::runtime::{
    burn_conv1d, burn_linear_2d, validate_tensor_port,
};

pub const DARTSORT_DENOISER_MODEL_ID: &str = "dartsort/singlechan-denoiser-v1";

/// Pretrained DARTsort single-channel waveform denoiser loaded from `.safetensors`.
pub struct DartsortWaveformDenoiser {
    weights: SafetensorsMap,
    io_spec: Option<TensorIoSpec>,
    target: ComputeTarget,
}

impl DartsortWaveformDenoiser {
    /// Pulls and initializes `dartsort/singlechan-denoiser-v1` from [`crate::runtime::pull_model`].
    #[cfg(feature = "hub")]
    pub fn from_hub(target: ComputeTarget) -> DspResult<Self> {
        let (manifest, path) = crate::runtime::pull_model(DARTSORT_DENOISER_MODEL_ID)?;
        let compute_target = target.checked().map_err(|e| DspError::ComputeError(e.to_string()))?;
        let weights = SafetensorsMap::from_file(&path)
            .map_err(|e| DspError::InvalidConfig(e.to_string()))?;
        Ok(Self {
            weights,
            io_spec: manifest.io_spec,
            target: compute_target,
        })
    }

    /// Loads a DARTsort `.safetensors` denoiser file from disk.
    pub fn from_safetensors_file(
        path: impl AsRef<Path>,
        target: ComputeTarget,
    ) -> DspResult<Self> {
        let compute_target = target.checked().map_err(|e| DspError::ComputeError(e.to_string()))?;
        let weights = SafetensorsMap::from_file(path)
            .map_err(|e| DspError::InvalidConfig(e.to_string()))?;
        Ok(Self {
            weights,
            io_spec: None,
            target: compute_target,
        })
    }

    pub fn target(&self) -> ComputeTarget {
        self.target
    }
}

impl WaveformDenoiser for DartsortWaveformDenoiser {
    fn denoise(&self, batch: &SnippetBatch) -> DspResult<SnippetBatch> {
        let [n, k, t] = batch.shape();
        if n == 0 || k == 0 || t == 0 {
            return Ok(batch.clone());
        }

        let single_chan_batches = n * k;
        if let Some(io) = &self.io_spec {
            if let Some(in_port) = io.inputs.first() {
                validate_tensor_port(in_port, &[single_chan_batches, 1, t])?;
            }
        }

        let (conv_w_shape, conv_w) = self
            .weights
            .get_raw("conv1.weight")
            .map_err(|e| DspError::InvalidConfig(e.to_string()))?;
        if conv_w_shape.len() != 3 {
            return Err(DspError::InvalidConfig(format!(
                "conv1.weight expected 3D shape, got {:?}",
                conv_w_shape
            )));
        }
        let out_ch = conv_w_shape[0];
        let kernel_size = conv_w_shape[2];
        let conv_b = self.weights.get_raw("conv1.bias").ok().map(|(_, d)| d);
        let pad = kernel_size / 2;

        let (mut hidden, out_len) = burn_conv1d(
            self.target,
            &batch.data,
            single_chan_batches,
            1,
            t,
            conv_w,
            out_ch,
            kernel_size,
            conv_b,
            1,
            pad,
            1,
            1,
        )?;
        for v in &mut hidden {
            *v = v.max(0.0);
        }

        let (lin_w_shape, lin_w) = self
            .weights
            .get_raw("fc.weight")
            .map_err(|e| DspError::InvalidConfig(e.to_string()))?;
        if lin_w_shape.len() != 2 {
            return Err(DspError::InvalidConfig(format!(
                "fc.weight expected 2D shape, got {:?}",
                lin_w_shape
            )));
        }
        let lin_b = self.weights.get_raw("fc.bias").ok().map(|(_, d)| d);
        let denoised = burn_linear_2d(
            self.target,
            &hidden,
            single_chan_batches,
            out_ch * out_len,
            lin_w,
            lin_w_shape[0],
            lin_b,
        )?;

        Ok(SnippetBatch {
            data: denoised,
            num_spikes: n,
            num_channels: k,
            num_samples: t,
            // Denoising keeps the alignment
            peak_index: batch.peak_index,
            primary_channels: batch.primary_channels.clone(),
            center_samples: batch.center_samples.clone(),
            subsample_offsets: batch.subsample_offsets.clone(),
            channel_ids: batch.channel_ids.clone(),
        })
    }
}