rust-mlp 0.1.0

Small, from-scratch MLP (dense feed-forward) for Rust: predictable, allocation-free hot path, batched training, and optional fast GEMM backend.
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

use std::alloc::{GlobalAlloc, Layout, System};
use std::sync::atomic::{AtomicUsize, Ordering};

use rust_mlp::{Activation, Dataset, FitConfig, Loss, LrSchedule, MlpBuilder, Optimizer, Shuffle};

struct CountingAlloc {
    allocs: AtomicUsize,
    reallocs: AtomicUsize,
    deallocs: AtomicUsize,
    bytes: AtomicUsize,
}

impl CountingAlloc {
    const fn new() -> Self {
        Self {
            allocs: AtomicUsize::new(0),
            reallocs: AtomicUsize::new(0),
            deallocs: AtomicUsize::new(0),
            bytes: AtomicUsize::new(0),
        }
    }

    fn reset(&self) {
        self.allocs.store(0, Ordering::Relaxed);
        self.reallocs.store(0, Ordering::Relaxed);
        self.deallocs.store(0, Ordering::Relaxed);
        self.bytes.store(0, Ordering::Relaxed);
    }

    fn snapshot(&self) -> AllocSnapshot {
        AllocSnapshot {
            allocs: self.allocs.load(Ordering::Relaxed),
            reallocs: self.reallocs.load(Ordering::Relaxed),
            deallocs: self.deallocs.load(Ordering::Relaxed),
            bytes: self.bytes.load(Ordering::Relaxed),
        }
    }

    fn alloc_events(&self) -> usize {
        self.allocs.load(Ordering::Relaxed) + self.reallocs.load(Ordering::Relaxed)
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
struct AllocSnapshot {
    allocs: usize,
    reallocs: usize,
    deallocs: usize,
    bytes: usize,
}

unsafe impl GlobalAlloc for CountingAlloc {
    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
        self.allocs.fetch_add(1, Ordering::Relaxed);
        self.bytes.fetch_add(layout.size(), Ordering::Relaxed);
        unsafe { System.alloc(layout) }
    }

    unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
        self.allocs.fetch_add(1, Ordering::Relaxed);
        self.bytes.fetch_add(layout.size(), Ordering::Relaxed);
        unsafe { System.alloc_zeroed(layout) }
    }

    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
        self.deallocs.fetch_add(1, Ordering::Relaxed);
        unsafe { System.dealloc(ptr, layout) }
    }

    unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
        self.reallocs.fetch_add(1, Ordering::Relaxed);
        // Approximate accounting: record the new size.
        self.bytes.fetch_add(new_size, Ordering::Relaxed);
        unsafe { System.realloc(ptr, layout, new_size) }
    }
}

#[global_allocator]
static ALLOC: CountingAlloc = CountingAlloc::new();

fn make_dataset(len: usize, input_dim: usize, target_dim: usize) -> Dataset {
    let inputs = vec![0.1_f32; len * input_dim];
    let targets = vec![0.0_f32; len * target_dim];
    Dataset::from_flat(inputs, targets, input_dim, target_dim).unwrap()
}

#[test]
fn fit_does_not_allocate_per_step_for_batched_training() {
    if cfg!(feature = "matrixmultiply") {
        // The `matrixmultiply` backend may allocate internal scratch buffers.
        // This test focuses on the crate's own training loop behavior.
        return;
    }

    let input_dim = 32;
    let hidden = 64;
    let output_dim = 8;
    let batch_size = 16;

    let base = MlpBuilder::new(input_dim)
        .unwrap()
        .add_layer(hidden, Activation::Tanh)
        .unwrap()
        .add_layer(output_dim, Activation::Identity)
        .unwrap()
        .build_with_seed(0)
        .unwrap();

    let train_small = make_dataset(batch_size, input_dim, output_dim);
    let train_large = make_dataset(batch_size * 64, input_dim, output_dim);

    let cfg = FitConfig {
        epochs: 1,
        lr: 1e-2,
        batch_size,
        shuffle: Shuffle::None,
        lr_schedule: LrSchedule::Constant,
        optimizer: Optimizer::Sgd,
        weight_decay: 0.0,
        grad_clip_norm: None,
        loss: Loss::Mse,
        metrics: vec![],
    };

    let mut mlp_small = base.clone();
    ALLOC.reset();
    let before_small = ALLOC.snapshot();
    mlp_small.fit(&train_small, None, cfg.clone()).unwrap();
    let alloc_small = ALLOC.alloc_events();
    let after_small = ALLOC.snapshot();

    let mut mlp_large = base;
    ALLOC.reset();
    let before_large = ALLOC.snapshot();
    mlp_large.fit(&train_large, None, cfg).unwrap();
    let alloc_large = ALLOC.alloc_events();
    let after_large = ALLOC.snapshot();

    assert_eq!(
        alloc_small, alloc_large,
        "expected allocation event count to be independent of steps.\n\
small: before={before_small:?} after={after_small:?}\n\
large: before={before_large:?} after={after_large:?}"
    );
}