tabicl-model 2.1.1

TabICL transformer model — column embedding, row interaction, ICL learning, KV cache.
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

//! Production depth (3+3+12) but with RANDOM-init weights (not HF).
//! If this matches Python: the bug is in the HF checkpoint
//! interpretation (e.g., wrong key naming or shape).
//! If this fails too: the bug is in my port at deep stacks.

use std::path::Path;
use tabicl_model::TabICL;
use tabicl_model::tabicl::{ColFeatureGroup, TabICLConfig};

fn read_f32(path: &Path) -> Option<Vec<f32>> {
    let bytes = std::fs::read(path).ok()?;
    let mut out = Vec::with_capacity(bytes.len() / 4);
    for chunk in bytes.chunks_exact(4) {
        out.push(f32::from_le_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]));
    }
    Some(out)
}

#[test]
fn production_depth_random_weights_parity() {
    let base = Path::new("/tmp/tabicl_d3");
    let ckpt = base.join("ckpt.json");
    if !ckpt.exists() {
        return;
    }
    let mut cfg = TabICLConfig::default();
    cfg.max_classes = 10;
    cfg.num_quantiles = 999;
    cfg.embed_dim = 128;
    cfg.col_num_blocks = 3;
    cfg.col_nhead = 8;
    cfg.col_num_inds = 16;
    cfg.col_feature_group = ColFeatureGroup::Same;
    cfg.col_feature_group_size = 3;
    cfg.col_target_aware = true;
    cfg.col_ssmax = "qassmax-mlp-elementwise".into();
    cfg.row_num_blocks = 3;
    cfg.row_nhead = 8;
    cfg.row_num_cls = 4;
    cfg.icl_num_blocks = 12;
    cfg.icl_nhead = 8;
    cfg.icl_ssmax = "qassmax-mlp-elementwise".into();

    let mut model = TabICL::new(cfg.clone());
    model.load_from_file(&ckpt).expect("load");
    let x_flat = read_f32(&base.join("input.bin")).unwrap();
    let y_ref = read_f32(&base.join("output.bin")).unwrap();
    let x = ndarray::Array3::from_shape_vec((1, 6, 4), x_flat).unwrap();
    let y_train = ndarray::Array2::from_shape_vec((1, 4), vec![0_usize, 1, 2, 3]).unwrap();
    let out = model.forward(x.view(), Some(y_train.view()), None).unwrap();
    let n_train = 4;
    let active = 4;
    let mut max_diff = 0.0_f32;
    for i in 0..2 {
        for c in 0..active {
            let r = out[(0, n_train + i, c)];
            let p = y_ref[i * active + c];
            let d = (r - p).abs();
            if d > max_diff {
                max_diff = d;
            }
        }
    }
    eprintln!("Production-depth random-weights max diff: {max_diff}");
    eprintln!("py row0: {:?}", &y_ref[..active]);
    let rust_row0: Vec<_> = (0..active).map(|c| out[(0, n_train, c)]).collect();
    eprintln!("rs row0: {rust_row0:?}");
    assert!(
        max_diff < 0.01,
        "production-depth random weights diff too large: {max_diff}"
    );
}