burn-core 0.22.0

Flexible and Comprehensive Deep Learning Framework in Rust
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

use burn::module::Module;
use burn_core as burn;

#[derive(Module, Debug)]
struct Envelope<M> {
    state: M,
}

#[derive(Module, Debug)]
enum Choice<M> {
    Value(M),
    Other(usize),
}

fn validation<M: Module>(module: &M) -> M {
    module.valid()
}

fn training<M: Module>(module: M) -> M {
    module.train()
}

#[test]
fn stateless_modules_transition_without_autodiff() {
    #[derive(Module, Debug)]
    struct Settings {
        value: usize,
    }

    let settings = training(validation(&Envelope {
        state: Settings { value: 42 },
    }));
    assert_eq!(settings.state.value, 42);

    let choice = training(validation(&Choice::<Settings>::Other(7)));
    assert!(matches!(choice, Choice::Other(7)));
}

#[cfg(feature = "autodiff")]
mod autodiff {
    use super::*;
    use burn::module::{Flag, Lora, Param, ParamId, RunningState};
    use burn::tensor::{Bool, Device, Int, Tensor};

    #[derive(Module, Debug)]
    struct State {
        weight: Param<Tensor<1>>,
        indices: Param<Tensor<1, Int>>,
        mask: Param<Tensor<1, Bool>>,
        raw: Tensor<1>,
        running: RunningState<Tensor<1>>,
        #[module(skip)]
        skipped: Tensor<1>,
    }

    type Nested = Envelope<(Vec<Option<[State; 1]>>, Choice<Param<Flag>>)>;

    fn state(module: &Nested) -> &State {
        &module.state.0[0].as_ref().unwrap()[0]
    }

    fn flag(module: &Nested) -> &Param<Flag> {
        let Choice::Value(flag) = &module.state.1 else {
            panic!("expected the flag variant")
        };
        flag
    }

    #[test]
    fn nested_transitions_preserve_specialized_state_hooks() {
        let device = Device::flex().autodiff();
        let running = RunningState::new(Tensor::zeros([1], &device));
        running.update(Tensor::from_floats([7.0], &device));
        let module = Envelope {
            state: (
                vec![Some([State {
                    weight: Param::from_tensor(Tensor::ones([1], &device)),
                    indices: Param::initialized(ParamId::new(), Tensor::from_ints([1], &device)),
                    mask: Param::initialized(ParamId::new(), Tensor::from_bool([true], &device)),
                    raw: Tensor::ones([1], &device),
                    running,
                    skipped: Tensor::ones([1], &device),
                }])],
                Choice::Value(Param::from_bool(true)),
            ),
        };

        let valid = validation(&module);
        assert!(!state(&valid).weight.is_require_grad());
        assert!(!state(&valid).indices.val().is_autodiff());
        assert!(!state(&valid).mask.val().is_autodiff());
        // Raw tensors are not visited by parameter mappers, but still transition.
        assert!(!state(&valid).raw.is_autodiff());
        assert!(state(&valid).skipped.is_autodiff());
        assert!(!state(&valid).running.value().is_autodiff());
        assert_eq!(state(&valid).running.value().into_scalar::<f32>(), 7.0);
        assert!(!flag(&valid).is_enabled());
        assert_eq!(flag(&valid).id, flag(&module).id);
        assert!(state(&module).weight.is_require_grad());
        assert!(flag(&module).is_enabled());

        // train() must also flush pending running-state updates.
        state(&valid)
            .running
            .update(Tensor::from_floats([9.0], &Device::flex()));
        let trained = training(valid);
        assert!(state(&trained).weight.is_require_grad());
        assert_eq!(state(&trained).weight.id, state(&module).weight.id);
        assert!(state(&trained).indices.val().is_autodiff());
        assert!(state(&trained).mask.val().is_autodiff());
        assert!(state(&trained).raw.is_autodiff());
        assert!(!state(&trained).raw.is_require_grad());
        assert!(state(&trained).running.value().is_autodiff());
        assert_eq!(state(&trained).running.value().into_scalar::<f32>(), 9.0);
        assert!(flag(&trained).is_enabled());
    }

    #[test]
    fn training_enables_nested_adapter_parameters_without_unfreezing_the_base() {
        let device = Device::flex();
        let weight =
            Param::from_tensor(Tensor::<2>::ones([4, 4], &device)).apply_lora(Lora::new(2, 4.0));
        let adapter = weight.adapter().unwrap();
        let a_id = adapter.a.id;
        let b_id = adapter.b.id;

        let trained = training(weight);
        assert!(trained.base().is_autodiff());
        assert!(!trained.base().is_require_grad());
        let adapter = trained.adapter().unwrap();
        assert_eq!(adapter.a.id, a_id);
        assert_eq!(adapter.b.id, b_id);
        assert!(adapter.a.is_require_grad());
        assert!(adapter.b.is_require_grad());
    }
}