burn-nn 0.22.0

Neural network building blocks for the Burn deep learning framework
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

use burn_core as burn;

use burn::module::{Module, Quantizer};
use burn::tensor::{
    Distribution, Tensor, Tolerance,
    quantization::{Calibration, QuantScheme, QuantValue, ScaleDtype},
};
use burn_nn::{
    LinearConfig,
    transformer::{TransformerEncoder, TransformerEncoderConfig, TransformerEncoderInput},
};

fn should_quantize_module<M: Module, const D: usize, F: Fn(&M) -> Tensor<D>>(
    module: M,
    scheme: QuantScheme,
    func: F,
    tolerance: Tolerance<f32>,
) {
    let result = func(&module);

    let calibration = Calibration::MinMax;
    let mut quantizer = Quantizer::new(calibration, scheme);
    let q_module = module.quantize_weights(&mut quantizer);
    let q_result = func(&q_module);

    result
        .into_data()
        .assert_approx_eq::<f32>(&q_result.into_data(), tolerance);
}

#[test]
fn should_quantize_transformer() {
    let device = Default::default();
    let transformer: TransformerEncoder =
        TransformerEncoderConfig::new(128, 256, 2, 2).init(&device);
    let signal = Tensor::random([2, 32, 128], Distribution::Default, &device);
    let scheme = device
        .settings()
        .quantization
        .scheme
        .with_value(QuantValue::Q8S)
        .per_block([32], ScaleDtype::F32);

    should_quantize_module(
        transformer,
        scheme,
        |tr| tr.forward(TransformerEncoderInput::new(signal.clone())),
        Tolerance::rel_abs(1e-2, 2e-2), // slightly higher abs tolerance (permissive: 1e-2)
    );
}

#[test]
fn should_quantize_linear_128_256() {
    let device = Default::default();
    let transformer = LinearConfig::new(128, 256).with_bias(false).init(&device);
    let signal = Tensor::<2>::random([1, 128], Distribution::Default, &device);
    let scheme = device
        .settings()
        .quantization
        .scheme
        .with_value(QuantValue::Q8S);

    should_quantize_module(
        transformer,
        scheme,
        |tr| tr.forward(signal.clone()),
        Tolerance::permissive(),
    );
}

#[test]
fn should_quantize_linear() {
    let device = Default::default();
    let transformer = LinearConfig::new(32, 32).with_bias(false).init(&device);
    let signal = Tensor::<2>::random([1, 32], Distribution::Default, &device);
    // Default scheme should select supported QuantStore default
    // TODO: set native if dtype is supported by the test backend
    let scheme = device
        .settings()
        .quantization
        .scheme
        // .with_store(QuantStore::Native)
        .with_value(QuantValue::Q8S);

    should_quantize_module(
        transformer,
        scheme,
        |tr| tr.forward(signal.clone()),
        Tolerance::permissive(),
    );
}

#[test]
fn should_quantize_linear_weights() {
    let device = Default::default();
    let transformer = LinearConfig::new(32, 32).with_bias(false).init(&device);
    let scheme = device
        .settings()
        .quantization
        .scheme
        .with_value(QuantValue::Q8S);

    should_quantize_module(
        transformer,
        scheme,
        |tr| tr.weight.val().dequantize(),
        Tolerance::permissive(),
    );
}

#[test]
fn should_quantize_linear_blocks() {
    let device = Default::default();
    let transformer = LinearConfig::new(32, 32).with_bias(false).init(&device);
    let signal = Tensor::<2>::random([1, 32], Distribution::Default, &device);
    let scheme = device
        .settings()
        .quantization
        .scheme
        .with_value(QuantValue::Q8S)
        // .with_store(QuantStore::Native)
        .per_block([16], ScaleDtype::F32);

    should_quantize_module(
        transformer,
        scheme,
        |tr| tr.forward(signal.clone()),
        Tolerance::permissive(),
    );
}

#[test]
fn should_quantize_linear_weights_blocks() {
    let device = Default::default();
    let transformer = LinearConfig::new(32, 32).with_bias(false).init(&device);
    let scheme = device
        .settings()
        .quantization
        .scheme
        .with_value(QuantValue::Q8S)
        // .with_store(QuantStore::Native)
        .per_block([16], ScaleDtype::F32);

    should_quantize_module(
        transformer,
        scheme,
        |tr| tr.weight.val().dequantize(),
        Tolerance::permissive(),
    );
}