vkml 0.0.2

High-level Vulkan-based machine learning library
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

mod common;

use half::f16;
use std::path::Path;
use vkml::{ComputeManager, Tensor};

use crate::common::get_test_data_path;

fn load_image_as_input_fp16(path: &Path) -> Box<[u8]> {
    let raw = image::open(path)
        .expect("Failed to open image")
        .into_bytes();
    let mut out: Vec<f16> = Vec::with_capacity(raw.len());
    for b in raw.into_iter() {
        let f32_val = (b as f32) / 255.0;
        out.push(f16::from_f32(f32_val));
    }
    bytemuck::cast_slice(&out).to_vec().into_boxed_slice()
}

fn argmax_fp16(data: &[u8]) -> usize {
    let f16_data: &[f16] = bytemuck::cast_slice(data);
    f16_data
        .iter()
        .map(|&h| h.to_f32())
        .enumerate()
        .max_by(|(_, a), (_, b)| a.partial_cmp(b).expect("NaN in results"))
        .map(|(i, _)| i)
        .unwrap()
}

#[test]
fn test_mnist_mlp_fp16() -> Result<(), Box<dyn std::error::Error>> {
    let file_path = get_test_data_path("mnist_mlp_fp16.onnx");
    let file = file_path.to_str().unwrap();

    let mut cm = ComputeManager::new_from_onnx_path(file)?;

    let img0_path = get_test_data_path("0.png");
    let img1_path = get_test_data_path("1.png");

    let input_ids = cm.tensor_graph.get_input_tensor_ids().to_vec();
    let input_tensor = cm.tensor_read(input_ids[0]);
    let input_desc = input_tensor.desc().clone();

    let bytes0 = load_image_as_input_fp16(&img0_path);
    let bytes1 = load_image_as_input_fp16(&img1_path);

    let batch0 = Tensor::new_cpu(input_desc.clone(), bytes0);
    let batch1 = Tensor::new_cpu(input_desc.clone(), bytes1);

    let ref0 = [
        0.997_203_7,
        3.214_188_2e-9,
        9.607_657e-6,
        5.107_659e-5,
        3.579_999_2e-6,
        0.002_036_365_2,
        0.000_560_617_4,
        6.727_680_6e-6,
        0.000_106_254_67,
        2.204_262_5e-5,
    ];
    let ref1 = [
        6.525_985_4e-5,
        0.985_778_8,
        0.002_706_837,
        0.002_157_057,
        0.000_299_26,
        0.000_512_54,
        0.000_863_986_2,
        0.001_076_165_9,
        0.006_037_95,
        0.000_502_062,
    ];

    println!("Running inference for mnist_mlp_fp16 (0.png)...");
    let out0 = cm.forward(vec![batch0])?;
    let bytes0 = cm.tensor_read(out0[0]).read();
    let res0: Vec<f32> = bytemuck::cast_slice::<u8, f16>(&bytes0)
        .iter()
        .map(|&h| h.to_f32())
        .collect();
    println!("0.png Prediction: {}", argmax_fp16(&bytes0));
    common::assert_tensors_close(&res0, &ref0, 1e-2, 1e-2, "mnist_mlp_fp16 Image 0");

    println!("Running inference for mnist_mlp_fp16 (1.png)...");
    let out1 = cm.forward(vec![batch1])?;
    let bytes1 = cm.tensor_read(out1[0]).read();
    let res1: Vec<f32> = bytemuck::cast_slice::<u8, f16>(&bytes1)
        .iter()
        .map(|&h| h.to_f32())
        .collect();
    println!("1.png Prediction: {}", argmax_fp16(&bytes1));
    common::assert_tensors_close(&res1, &ref1, 1e-2, 1e-2, "mnist_mlp_fp16 Image 1");

    Ok(())
}

#[test]
fn test_mnist_conv_fp16() -> Result<(), Box<dyn std::error::Error>> {
    let file_path = get_test_data_path("mnist_conv_fp16.onnx");
    let file = file_path.to_str().unwrap();

    let mut cm = ComputeManager::new_from_onnx_path(file)?;

    let img0_path = get_test_data_path("0.png");
    let img1_path = get_test_data_path("1.png");

    let input_ids = cm.tensor_graph.get_input_tensor_ids().to_vec();
    let input_tensor = cm.tensor_read(input_ids[0]);
    let input_desc = input_tensor.desc().clone();

    let ref0 = [
        0.387_695_3,
        0.004_970_550_5,
        0.104_858_4,
        0.139_892_58,
        0.018_615_723,
        0.134_521_48,
        0.093_200_68,
        0.022_674_56,
        0.074_584_96,
        0.018_981_934,
    ];
    let ref1 = [
        0.006_629_944,
        0.647_949_2,
        0.039_062_5,
        0.065_246_58,
        0.023_101_807,
        0.026_580_81,
        0.047_149_66,
        0.033_874_51,
        0.075_805_664,
        0.034_545_9,
    ];

    println!("Running inference for mnist_conv_fp16 (0.png)...");
    let bytes0 = load_image_as_input_fp16(&img0_path);
    let batch0 = Tensor::new_cpu(input_desc.clone(), bytes0);
    let out0 = cm.forward(vec![batch0])?;
    let bytes0 = cm.tensor_read(out0[0]).read();
    let res0: Vec<f32> = bytemuck::cast_slice::<u8, f16>(&bytes0)
        .iter()
        .map(|&h| h.to_f32())
        .collect();
    println!("0.png Prediction: {}", argmax_fp16(&bytes0));
    common::assert_tensors_close(&res0, &ref0, 1e-2, 1e-2, "mnist_conv_fp16 Image 0");

    println!("Running inference for mnist_conv_fp16 (1.png)...");
    let bytes1 = load_image_as_input_fp16(&img1_path);
    let batch1 = Tensor::new_cpu(input_desc.clone(), bytes1);
    let out1 = cm.forward(vec![batch1])?;
    let bytes1 = cm.tensor_read(out1[0]).read();
    let res1: Vec<f32> = bytemuck::cast_slice::<u8, f16>(&bytes1)
        .iter()
        .map(|&h| h.to_f32())
        .collect();
    println!("1.png Prediction: {}", argmax_fp16(&bytes1));
    common::assert_tensors_close(&res1, &ref1, 1e-2, 1e-2, "mnist_conv_fp16 Image 1");

    Ok(())
}