exg 0.0.5

EXG (EEG/ECG/EMG) preprocessing — native Rust DSP + FIF reader, numerical parity with MNE-Python
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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266

//! Safetensors I/O for the preprocessing pipeline.
//!
//! - [`RawData`]: parses `raw.safetensors` written by `scripts/read_raw.py`.
//! - [`StWriter`]: generic safetensors file builder (F32, F64, I32 tensors).
//! - [`write_batch`]: writes preprocessed epochs to `batch.safetensors`.
//!
//! For LUNA-specific epoch export/import, see the
//! [`exg-luna`](https://crates.io/crates/exg-luna) crate.
use anyhow::{bail, Context, Result};
use ndarray::Array2;
use std::collections::HashMap;
use std::path::Path;

// ── Low-level safetensors parser (no dependency on the `safetensors` crate's
//    tensor types — we just need raw bytes → ndarray). ─────────────────────────

fn parse_header(bytes: &[u8]) -> Result<(HashMap<String, serde_json::Value>, usize)> {
    if bytes.len() < 8 {
        bail!("safetensors file too small");
    }
    let n = u64::from_le_bytes(bytes[..8].try_into().unwrap()) as usize;
    let header: HashMap<String, serde_json::Value> =
        serde_json::from_slice(&bytes[8..8 + n])
            .context("failed to parse safetensors header")?;
    Ok((header, 8 + n))
}

fn read_f32_tensor(
    bytes: &[u8],
    data_start: usize,
    entry: &serde_json::Value,
) -> Result<Vec<f32>> {
    let offsets = entry["data_offsets"].as_array().unwrap();
    let s = offsets[0].as_u64().unwrap() as usize;
    let e = offsets[1].as_u64().unwrap() as usize;
    let raw = &bytes[data_start + s..data_start + e];
    Ok(raw
        .chunks_exact(4)
        .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]]))
        .collect())
}

fn shape_of(entry: &serde_json::Value) -> Vec<usize> {
    entry["shape"]
        .as_array()
        .unwrap()
        .iter()
        .map(|v| v.as_u64().unwrap() as usize)
        .collect()
}

// ── Public structs ────────────────────────────────────────────────────────────

/// Raw EEG data loaded from `raw.safetensors` (output of `scripts/read_raw.py`).
pub struct RawData {
    /// [C, T] in original units (volts or arbitrary), original sample rate.
    pub data: Array2<f32>,
    /// [C, 3] channel positions in metres.
    pub chan_pos: Array2<f32>,
    /// Original sampling rate (Hz).
    pub sfreq: f32,
    /// Channel names (may be empty if not saved).
    pub ch_names: Vec<String>,
}

impl RawData {
    /// Load raw EEG data from a safetensors file produced by `scripts/read_raw.py`.
    pub fn load(path: &Path) -> Result<Self> {
        let bytes = std::fs::read(path).context("reading raw.safetensors")?;
        let (header, data_start) = parse_header(&bytes)?;

        let data_entry = header.get("data").context("missing 'data' key")?;
        let data_shape = shape_of(data_entry);
        let data_vec = read_f32_tensor(&bytes, data_start, data_entry)?;
        let data = Array2::from_shape_vec((data_shape[0], data_shape[1]), data_vec)?;

        let pos_entry = header.get("chan_pos").context("missing 'chan_pos' key")?;
        let pos_shape = shape_of(pos_entry);
        let pos_vec = read_f32_tensor(&bytes, data_start, pos_entry)?;
        let chan_pos = Array2::from_shape_vec((pos_shape[0], pos_shape[1]), pos_vec)?;

        let sfreq_entry = header.get("sfreq").context("missing 'sfreq' key")?;
        let sfreq = read_f32_tensor(&bytes, data_start, sfreq_entry)?[0];

        // Channel names are optional.
        let ch_names = if let Some(e) = header.get("ch_names") {
            let offsets = e["data_offsets"].as_array().unwrap();
            let s = offsets[0].as_u64().unwrap() as usize;
            let end = offsets[1].as_u64().unwrap() as usize;
            let raw_str = std::str::from_utf8(&bytes[data_start + s..data_start + end])?;
            raw_str
                .split('\n')
                .filter(|s| !s.is_empty())
                .map(String::from)
                .collect()
        } else {
            vec![]
        };

        Ok(RawData { data, chan_pos, sfreq, ch_names })
    }
}

// ── Generic safetensors builder ───────────────────────────────────────────────

/// Simple safetensors file writer that handles F32, F64, and I32 tensors.
///
/// Usage:
/// ```rust,no_run
/// use exg::io::StWriter;
/// use std::path::Path;
/// let mut w = StWriter::new();
/// w.add_f32("signal", &[1.0f32, 2.0, 3.0], &[1, 3]);
/// w.add_f64("signal_d", &[1.0f64, 2.0, 3.0], &[1, 3]);
/// w.write(Path::new("/tmp/out.safetensors")).unwrap();
/// ```
#[derive(Default)]
pub struct StWriter {
    /// Internal entries: (name, bytes, dtype_str, shape).
    pub entries: Vec<(String, Vec<u8>, &'static str, Vec<usize>)>,
}

impl StWriter {
    /// Create an empty writer.
    pub fn new() -> Self {
        Self::default()
    }

    /// Add a named F32 tensor.
    pub fn add_f32(&mut self, name: &str, data: &[f32], shape: &[usize]) {
        let bytes: Vec<u8> = data.iter().flat_map(|v| v.to_le_bytes()).collect();
        self.entries.push((name.to_string(), bytes, "F32", shape.to_vec()));
    }

    /// Add a named F32 tensor from an `Array2`.
    pub fn add_f32_arr2(&mut self, name: &str, arr: &ndarray::Array2<f32>) {
        let data: Vec<f32> = arr.iter().copied().collect();
        self.add_f32(name, &data, &[arr.nrows(), arr.ncols()]);
    }

    /// Add a named F64 tensor.
    pub fn add_f64(&mut self, name: &str, data: &[f64], shape: &[usize]) {
        let bytes: Vec<u8> = data.iter().flat_map(|v| v.to_le_bytes()).collect();
        self.entries.push((name.to_string(), bytes, "F64", shape.to_vec()));
    }

    /// Add a named F64 tensor from an `Array2`.
    pub fn add_f64_arr2(&mut self, name: &str, arr: &ndarray::Array2<f64>) {
        let data: Vec<f64> = arr.iter().copied().collect();
        self.add_f64(name, &data, &[arr.nrows(), arr.ncols()]);
    }

    /// Add a named I32 tensor.
    pub fn add_i32(&mut self, name: &str, data: &[i32], shape: &[usize]) {
        let bytes: Vec<u8> = data.iter().flat_map(|v| v.to_le_bytes()).collect();
        self.entries.push((name.to_string(), bytes, "I32", shape.to_vec()));
    }

    /// Write all tensors to a safetensors file.
    pub fn write(&self, path: &Path) -> Result<()> {
        use std::io::Write;
        let mut header_map = serde_json::Map::new();
        let mut offset: usize = 0;
        for (name, data, dtype, shape) in &self.entries {
            header_map.insert(name.clone(), serde_json::json!({
                "dtype": dtype,
                "shape": shape,
                "data_offsets": [offset, offset + data.len()],
            }));
            offset += data.len();
        }
        let hdr_bytes = serde_json::to_vec(&header_map)?;
        let pad = (8 - hdr_bytes.len() % 8) % 8;
        let padded: Vec<u8> = hdr_bytes.into_iter()
            .chain(std::iter::repeat_n(b' ', pad))
            .collect();
        let mut f = std::fs::File::create(path)?;
        f.write_all(&(padded.len() as u64).to_le_bytes())?;
        f.write_all(&padded)?;
        for (_, data, _, _) in &self.entries {
            f.write_all(data)?;
        }
        Ok(())
    }
}

// ── Batch writer ──────────────────────────────────────────────────────────────

/// Write preprocessed epochs to `batch.safetensors`.
///
/// `epochs[e]`: [C, 1280]   `positions[e]`: [C, 3]
pub fn write_batch(
    epochs: &[Array2<f32>],
    positions: &[Array2<f32>],
    path: &Path,
) -> Result<()> {
    use std::io::Write;

    let n = epochs.len();
    assert_eq!(n, positions.len());

    // Collect tensors: (name, f32 data, shape).
    let mut tensors: Vec<(String, Vec<u8>, Vec<usize>)> = vec![];

    for i in 0..n {
        let eeg  = &epochs[i];
        let pos  = &positions[i];

        let eeg_bytes: Vec<u8> = eeg.iter()
            .flat_map(|v| v.to_le_bytes())
            .collect();
        tensors.push((
            format!("eeg_{i}"),
            eeg_bytes,
            vec![eeg.nrows(), eeg.ncols()],
        ));

        let pos_bytes: Vec<u8> = pos.iter()
            .flat_map(|v| v.to_le_bytes())
            .collect();
        tensors.push((
            format!("chan_pos_{i}"),
            pos_bytes,
            vec![pos.nrows(), pos.ncols()],
        ));
    }

    // n_samples scalar.
    tensors.push((
        "n_samples".into(),
        (n as i32).to_le_bytes().to_vec(),
        vec![1],
    ));

    // Build header.
    let mut header_map = serde_json::Map::new();
    let mut offset: usize = 0;
    let mut dtype_map: Vec<(String, String)> = vec![];

    for (name, data, shape) in &tensors {
        let dtype = if name == "n_samples" { "I32" } else { "F32" };
        header_map.insert(name.clone(), serde_json::json!({
            "dtype": dtype,
            "shape": shape,
            "data_offsets": [offset, offset + data.len()],
        }));
        dtype_map.push((name.clone(), dtype.to_string()));
        offset += data.len();
    }

    let header_bytes = serde_json::to_vec(&header_map)?;
    let pad = (8 - header_bytes.len() % 8) % 8;
    let header_padded: Vec<u8> = header_bytes
        .into_iter()
        .chain(std::iter::repeat_n(b' ', pad))
        .collect();

    let mut f = std::fs::File::create(path)?;
    f.write_all(&(header_padded.len() as u64).to_le_bytes())?;
    f.write_all(&header_padded)?;
    for (_, data, _) in &tensors {
        f.write_all(data)?;
    }

    Ok(())
}