svod-model 0.1.0-alpha.3

Pretrained models inference abstraction.
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

use crate::jit::InputSpec;
use crate::resnet::{OutputMode, ResNet, ResNetConfig, ResNetDepth, ResNetJit};

fn build_classifier_jit(max_batch: usize, num_classes: usize) -> ResNetJit {
    let config =
        ResNetConfig::new(ResNetDepth::R18, OutputMode::Classification { num_classes }).with_max_batch_size(max_batch);
    let model = ResNet::with_zero_weights(config);
    ResNetJit::new(model)
}

/// Realize-based JIT smoke gated behind `--ignored`. The default suite covers
/// the JIT mechanics through the toy `jit_recurrent` tests and the GigaAM
/// batch suite (which actually exercises variable rebinding); this is here
/// for hands-on ResNet work.
#[test]
#[ignore = "heavy: full ResNet-18 graph compile through the CPU backend"]
fn prepare_and_execute_at_max_batch() {
    let max_batch = 2;
    let mut jit = build_classifier_jit(max_batch, 5);
    jit.prepare(InputSpec::f32(&[max_batch, 3, 32, 32])).unwrap();
    jit.execute_with_vars(&[("b", max_batch as i64)]).unwrap();
    let out = jit.output().unwrap();
    assert_eq!(out.size(), 2 * 5 * std::mem::size_of::<f32>());
}

#[test]
#[ignore = "heavy: three executes of the full ResNet-18 graph"]
fn rebind_batch_without_reprepare() {
    let max_batch = 4;
    let mut jit = build_classifier_jit(max_batch, 3);
    jit.prepare(InputSpec::f32(&[max_batch, 3, 32, 32])).unwrap();

    for b in [1, 2, 4] {
        jit.execute_with_vars(&[("b", b as i64)]).unwrap();
        let out = jit.output().unwrap();
        assert!(out.size() > 0, "output buffer empty for b={b}");
    }
}

#[test]
#[ignore = "heavy: full ResNet-18 graph compile through the CPU backend"]
fn features_mode_returns_spatial_map() {
    let max_batch = 1;
    let config = ResNetConfig::new(ResNetDepth::R18, OutputMode::Features).with_max_batch_size(max_batch);
    let model = ResNet::with_zero_weights(config);
    let mut jit = ResNetJit::new(model);
    jit.prepare(InputSpec::f32(&[max_batch, 3, 32, 32])).unwrap();
    jit.execute_with_vars(&[("b", 1)]).unwrap();
    let out = jit.output().unwrap();
    assert_eq!(out.size(), 512 * std::mem::size_of::<f32>());
}