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

use crate as burn;

use crate::{config::Config, tensor::Tensor};
use burn_tensor::{backend::Backend, ElementConversion};

/// Gradient Clipping provides a way to mitigate exploding gradients
#[derive(Config)]
pub enum GradientClippingConfig {
    /// Clip the gradient by value.
    Value(f32),

    /// Clip the gradient by norm.
    Norm(f32),
}

impl GradientClippingConfig {
    /// Initialize the gradient clipping.
    ///
    /// # Returns
    ///
    /// The gradient clipping.
    pub fn init(&self) -> GradientClipping {
        match self {
            GradientClippingConfig::Value(val) => GradientClipping::Value(*val),
            GradientClippingConfig::Norm(val) => GradientClipping::Norm(*val),
        }
    }
}

/// Gradient Clipping provides a way to mitigate exploding gradients
/// by clipping every component of the gradient by value or by norm during
/// backpropagation.
pub enum GradientClipping {
    /// Clip the gradient by value.
    Value(f32),

    /// Clip the gradient by norm.
    Norm(f32),
}

impl GradientClipping {
    /// Clip the gradient.
    ///
    /// # Arguments
    ///
    /// * `grad` - The gradient to clip.
    ///
    /// # Returns
    ///
    /// The clipped gradient.
    pub fn clip_gradient<B: Backend, const D: usize>(&self, grad: Tensor<B, D>) -> Tensor<B, D> {
        match self {
            GradientClipping::Value(threshold) => self.clip_by_value(grad, *threshold),
            GradientClipping::Norm(max_norm) => self.clip_by_norm(grad, *max_norm),
        }
    }

    fn clip_by_value<B: Backend, const D: usize>(
        &self,
        grad: Tensor<B, D>,
        threshold: f32,
    ) -> Tensor<B, D> {
        let greater_mask = grad.clone().greater_elem(threshold);
        let lower_mask = grad.clone().lower_elem(-threshold);

        let clipped_grad = grad.mask_fill(greater_mask, threshold);

        clipped_grad.mask_fill(lower_mask, -threshold)
    }

    fn l2_norm<B: Backend, const D: usize>(tensor: Tensor<B, D>) -> Tensor<B, 1> {
        let squared = tensor.powf(2.0);
        let sum = squared.sum();

        sum.sqrt()
    }

    fn clip_by_norm<B: Backend, const D: usize>(
        &self,
        grad: Tensor<B, D>,
        threshold: f32,
    ) -> Tensor<B, D> {
        let norm = Self::l2_norm(grad.clone());
        let norm_float = norm.into_scalar().elem::<f32>();
        if norm_float > threshold {
            let scale = threshold / norm_float;
            grad.mul_scalar(scale)
        } else {
            grad
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::tensor::Tensor;
    use crate::TestBackend;

    #[test]
    fn test_clip_by_value() {
        let gradient: Tensor<TestBackend, 2> = Tensor::from_floats([
            [0.6294, 0.0940, 0.8176, 0.8824, 0.5228, 0.4310],
            [0.7152, 0.9559, 0.7893, 0.5684, 0.5939, 0.8883],
        ]);

        let clipped_gradient = GradientClipping::Value(0.5).clip_gradient(gradient);
        let clipped_gradient_data = clipped_gradient.into_data();

        for value in clipped_gradient_data.value {
            assert!(value <= 0.5);
        }
    }

    #[test]
    fn test_clip_by_norm() {
        let gradient: Tensor<TestBackend, 2> = Tensor::from_floats([
            [0.6294, 0.0940, 0.8176, 0.8824, 0.5228, 0.4310],
            [0.7152, 0.9559, 0.7893, 0.5684, 0.5939, 0.8883],
        ]);

        let clipped_gradient = GradientClipping::Norm(2.2).clip_gradient(gradient);
        let clipped_gradient_data = clipped_gradient.into_data();

        for value in clipped_gradient_data.value {
            assert!(value <= 0.88);
        }
    }
}