zenu-layer 0.1.1

A simple neural network layer library.
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

use std::collections::HashMap;

use crate::{Module, Parameters};
use rand_distr::{Distribution, StandardNormal};
use zenu_autograd::{
    creator::{rand::normal, zeros::zeros},
    functions::{matmul::matmul, transpose::transpose},
    Variable,
};
use zenu_matrix::{device::Device, num::Num};

pub struct Linear<T: Num, D: Device> {
    in_features: usize,
    out_features: usize,
    pub weight: Variable<T, D>,
    pub bias: Option<Variable<T, D>>,
}

impl<T: Num, D: Device> Module<T, D> for Linear<T, D> {
    type Input = Variable<T, D>;
    type Output = Variable<T, D>;
    fn call(&self, input: Variable<T, D>) -> Variable<T, D> {
        let weight_t = transpose(self.weight.clone());
        let output = matmul(input, weight_t);
        if let Some(bias) = &self.bias {
            output.set_name("linear.intermediate_output");
            output + bias.clone()
        } else {
            output
        }
    }
}

impl<T: Num, D: Device> Parameters<T, D> for Linear<T, D> {
    fn weights(&self) -> HashMap<String, Variable<T, D>> {
        let mut weights = HashMap::new();
        weights.insert("linear.weight".to_string(), self.weight.clone());
        weights
    }

    fn biases(&self) -> HashMap<String, Variable<T, D>> {
        let mut biases = HashMap::new();
        if let Some(bias) = &self.bias {
            biases.insert("linear.bias".to_string(), bias.clone());
        }
        biases
    }
}

impl<T: Num, D: Device> Linear<T, D> {
    #[must_use]
    pub fn new(in_features: usize, out_features: usize, use_bias: bool) -> Self
    where
        StandardNormal: Distribution<T>,
    {
        let weight = normal(T::zero(), T::one(), None, [out_features, in_features]);
        weight
            .get_data_mut()
            .to_ref_mut()
            .div_scalar_assign(T::from_usize(in_features).sqrt());
        let bias = if use_bias {
            let bias = zeros([out_features]);
            bias.set_name("linear.bias");
            bias.set_is_train(true);
            Some(bias)
        } else {
            None
        };

        weight.set_is_train(true);
        weight.set_name("linear.weight");

        Self {
            in_features,
            out_features,
            weight,
            bias,
        }
    }

    #[must_use]
    pub fn to<Dout: Device>(self) -> Linear<T, Dout> {
        Linear {
            in_features: self.in_features,
            out_features: self.out_features,
            weight: self.weight.to(),
            bias: self.bias.map(|b| b.to()),
        }
    }
}

// #[cfg(test)]
// mod linear {
//     use zenu_autograd::creator::rand::normal;
//     use zenu_matrix::{device::Device, dim::DimTrait, operation::mul::matmul};
//     use zenu_test::{assert_mat_eq_epsilon, assert_val_eq, run_test};
//
//     use crate::{Module, StateDict};
//
//     use super::Linear;
//
//     fn with_bias<D: Device>() {
//         let layer = Linear::<f32, D>::new(3, 2, true);
//         let input = normal::<_, _, D>(0., 1., None, [5, 3]);
//         let output = layer.call(input.clone());
//         assert_eq!(output.get_data().shape().slice(), [5, 2]);
//
//         let parameters = layer.to_json();
//
//         let ans = matmul(
//             &input.get_data().to_ref(),
//             &layer.weight.get_data().to_ref(),
//         ) + &layer.bias.unwrap().get_data().to_ref();
//
//         assert_val_eq!(output.clone(), ans, 1e-4);
//
//         let new_layer = Linear::<f32, D>::from_json(&parameters);
//         let new_output = new_layer.call(input.clone());
//
//         assert_mat_eq_epsilon!(output.get_data(), new_output.get_data(), 1e-4);
//     }
//     run_test!(with_bias, with_bias_cpu, with_bias_gpu);
//
//     fn without_bias<D: Device>() {
//         let layer = Linear::<f32, D>::new(3, 2, false);
//         let input = normal::<_, _, D>(0., 1., None, [5, 3]);
//         let output = layer.call(input.clone());
//         assert_eq!(output.get_data().shape().slice(), [5, 2]);
//
//         let parameters = layer.to_json();
//         let weight = layer.weight.clone();
//
//         let ans = matmul(&input.get_data().to_ref(), &weight.get_data().to_ref());
//         assert_val_eq!(output.clone(), ans, 1e-4);
//
//         let new_layer = Linear::<f32, D>::from_json(&parameters);
//         let new_output = new_layer.call(input.clone());
//
//         assert_mat_eq_epsilon!(output.get_data(), new_output.get_data(), 1e-4);
//     }
//     run_test!(without_bias, without_bias_cpu, without_bias_gpu);
// }