puffpastry 0.1.0

basic rust neural network framework
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
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

use crate::{
    vec_tools::{self, Transpose, ValidNumber},
    Activation, Tensor,
};

#[derive(Debug, Clone, PartialEq)]
pub enum Loss {
    MeanSquaredError,
    CategoricalCrossEntropy,
}

impl Loss {
    pub fn calculate_loss<T: ValidNumber<T>>(
        &self,
        result: Tensor<T>,
        expected: Tensor<T>,
    ) -> Result<T, ()> {
        match self {
            Loss::MeanSquaredError => Loss::mean_squared_error(result, expected).ok_or_else(|| ()),
            Loss::CategoricalCrossEntropy => {
                Loss::categorical_cross_entropy(result, expected).ok_or_else(|| ())
            }
        }
    }

    pub fn get_gradient<T: ValidNumber<T>>(
        &self,
        result: Tensor<T>,
        expected: Tensor<T>,
    ) -> Result<Tensor<T>, ()> {
        match self {
            Loss::MeanSquaredError => Loss::mean_squared_error_grad(result, expected),
            Loss::CategoricalCrossEntropy => Loss::categorical_cross_entropy_grad(result, expected),
        }
    }

    fn mean_squared_error<T: ValidNumber<T>>(result: Tensor<T>, expected: Tensor<T>) -> Option<T> {
        if result.shape() != expected.shape() {
            return None;
        }
        let total: f64 = result
            .data
            .iter()
            .zip(expected.data.iter())
            .fold(T::from(0.0), |error, (res, expec)| {
                error + (*expec - *res) * (*expec - *res)
            })
            .into();

        Some(T::from(total / result.data.len() as f64))
    }

    fn categorical_cross_entropy<T: ValidNumber<T>>(
        result: Tensor<T>,
        expected: Tensor<T>,
    ) -> Option<T> {
        if result.shape() != expected.shape() || result.shape()[1] != 1 && expected.shape()[1] != 1
        {
            return None;
        }
        // Requires the output to be one-hot encoded.
        Some(
            T::from(-1.0)
                * result
                    .data
                    .into_iter()
                    .zip(expected.data.into_iter())
                    .map(|(res, expec)| (res.into(), expec.into()))
                    .find(|(_, expec): &(f64, f64)| *expec == 1.0)
                    .and_then(|(res, _)| {
                        // clipped to 1000.0
                        let out = res.ln();
                        if 1000.0 < -1.0 * out || out.is_nan() {
                            return Some(T::from(-1000.0));
                        }
                        Some(T::from(out))
                    })
                    .unwrap(),
        )
    }

    fn mean_squared_error_grad<T: ValidNumber<T>>(
        result: Tensor<T>,
        expected: Tensor<T>,
    ) -> Result<Tensor<T>, ()> {
        // Result and expected should be column vectors!
        if result.shape() != expected.shape() || result.shape()[1] != 1 {
            return Err(());
        }

        let data: Vec<Vec<T>> = result
            .as_rows()
            .iter()
            .zip(expected.as_rows().iter())
            .map(|(res, expec)| vec![T::from(-2.0) * (expec.data[0] - res.data[0])])
            .collect();

        Ok(Tensor::from(data))
    }

    fn categorical_cross_entropy_grad<T: ValidNumber<T>>(
        result: Tensor<T>,
        expected: Tensor<T>,
    ) -> Result<Tensor<T>, ()> {
        // Result and expected should be column vectors!
        if result.shape() != expected.shape() || result.shape()[1] != 1 {
            return Err(());
        }
        // clipped to 1000.0 as max
        let data: Vec<Vec<T>> = result
            .as_rows()
            .iter()
            .zip(expected.as_rows().iter())
            .map(|(res, expec)| (res.data[0].into(), expec.data[0].into()))
            .map(|(res, expec): (f64, f64)| {
                vec![T::from(
                    if 1000.0 < expec * (1.0 / res) || (1.0 / res).is_nan() {
                        -1000.0
                    } else {
                        -1.0 * expec * (1.0 / res)
                    },
                )]
            })
            .collect();

        Ok(Tensor::from(data))
    }
}

#[cfg(test)]
mod test {
    use super::*;

    #[test]
    fn mean_sq_test() {
        let v1 = Tensor::from(vec![1.0, 2.0, 3.0]);
        let v2 = Tensor::from(vec![2.0, 3.0, 4.0]);
        let res = Loss::mean_squared_error(v1, v2);

        assert_eq!(res, Some(1.0))
    }

    #[test]
    fn cce_test() {
        let v1 = Tensor::from(vec![vec![0.3], vec![0.0], vec![0.0]]);
        let v2 = Tensor::from(vec![vec![1.0], vec![0.0], vec![0.8]]);
        let res = Loss::categorical_cross_entropy(v1, v2);

        assert_eq!(res, Some(-1.0 * 0.3f64.ln()));
    }
}