unda 0.2.2

General purpose machine learning crate for neural network development and analysis
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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263

use std::ops::Range;

use crate::core::data::{matrix::Matrix, input::Input};

use super::{layers::Layer, methods::{distributions::Distributions, pair::GradientPair, activations::Activations}};
use rand::RngCore;
use serde::{Deserialize, Serialize};

///A Dense Neural Network Layer of a model, containing just nodes, weights, biases and an
///activation function
#[derive(Serialize, Deserialize, Clone)]
pub struct Dense{
    pub weights: Matrix,   
    pub biases: Matrix,
    pub data: Matrix,
    loss: f32,

    pub activation_fn: Activations,
    learning_rate: f32,

    beta1: f32,
    beta2: f32,
    epsilon: f32,
    time: usize,

    m_weights: Matrix,
    v_weights: Matrix,
    m_biases: Matrix,
    v_biases: Matrix
}

impl Dense{
    pub fn new_ser(rows: usize, cols: usize, flat_weight: Vec<f32>, flat_bias: Vec<f32>) -> Dense {
        let weight_shape: Matrix = Matrix::from_sized(flat_weight, rows, cols);
        let bias_shape: Matrix = Matrix::from_sized(flat_bias, rows, 1);

        Dense {
            weights: weight_shape,
            biases: bias_shape,
            data: Matrix::new_empty(0, 0),
            loss: 1.0,
            activation_fn: Activations::SIGMOID,
            learning_rate: 0.01,
            beta1: 0.99,
            beta2: 0.99,
            epsilon: 1e-16,
            time: 1,
            m_weights: Matrix::new_empty(0, 0),
            v_weights: Matrix::new_empty(0, 0),
            m_biases: Matrix::new_empty(0, 0),
            v_biases: Matrix::new_empty(0, 0)
        }
    }
    pub fn new(rows: usize, cols: usize, activation: Activations, learning_rate: f32, rng: &mut Box<dyn RngCore>) -> Dense{
        let distribution: Distributions = match activation{
            Activations::ELU(_) | Activations::RELU | Activations::LEAKYRELU  => Distributions::He(cols),
            Activations::TANH | Activations::SIGMOID => Distributions::Xavier(rows, cols),
            _ => Distributions::Ranged(-1.0..1.0)
        };
        let mut res = Dense { 
            loss: 1.0,
            weights: Matrix::new_random(cols, rows, rng, &distribution),
            biases: Matrix::new_random(cols, 1, rng, &Distributions::Ranged(-1.0..1.0)),

            m_weights: Matrix::new_empty(cols, rows),
            v_weights: Matrix::new_empty(cols, rows),

            m_biases: Matrix::new_empty(cols, 1),
            v_biases: Matrix::new_empty(cols, 1),

            data: Matrix::new_empty(0, 0),
            activation_fn: activation,
            learning_rate,
            beta1: 0.0,
            beta2: 0.0,
            epsilon: 0.0,
            time: 1
        };
        (res.beta1, res.beta2) = res.get_betas();
        res.epsilon = res.get_epsilon();

        res
    }
    fn get_betas(&self) -> (f32, f32){
        (0.9, 0.999)
    }
    fn get_epsilon(&self) -> f32{
        1e-10
    }
}


#[typetag::serde]
impl Layer for Dense{
    fn update_gradients(&mut self, gradient_pair: (&Box<dyn Input>, &Box<dyn Input>), clip: Option<Range<f32>>) {//, noise: &f32) {
        let bias_gradient = Matrix::from(gradient_pair.0.to_param_2d()) * self.learning_rate; //+ noise;
        let weight_gradient = Matrix::from(gradient_pair.1.to_param_2d()) * self.learning_rate; //+ noise;

        self.time += 1;

        self.m_weights = self.m_weights.clone() * self.beta1 + &(weight_gradient.clone() * (1.0 - self.beta1));
        self.v_weights = self.v_weights.clone() * self.beta2 + &((weight_gradient^2) * (1.0 - self.beta2));

        self.m_biases = self.m_biases.clone() * self.beta1 + &(bias_gradient.clone() * (1.0 - self.beta1));
        self.v_biases = self.v_biases.clone() * self.beta2 + &((bias_gradient.clone()^2) * (1.0 - self.beta2));

        let m_weights_hat = self.m_weights.clone() / (1.0 - self.beta1.powi(self.time as i32));
        let v_weights_hat = self.v_weights.clone() / (1.0 - self.beta2.powi(self.time as i32));

        let m_bias_hat = self.m_biases.clone() / (1.0 - self.beta1.powi(self.time as i32));
        let v_bias_hat = self.v_biases.clone() / (1.0 - self.beta2.powi(self.time as i32));

        let mut weights_update = m_weights_hat.clone() / &(v_weights_hat.sqrt() + self.epsilon);
        let mut bias_update = m_bias_hat.clone() / &(v_bias_hat.sqrt() + self.epsilon);

        if let Some(clip_range) = clip{
            bias_update.clip(&clip_range);
            weights_update.clip(&clip_range);
        }
        self.biases = self.biases.clone() + &bias_update;
        self.weights = self.weights.clone() + &weights_update;
    }
    fn avg_gradient(&self, gradients: Vec<&Box<dyn Input>>) -> Box<dyn Input>{
        let _len = gradients.len();
        let gradients_mat = gradients.into_iter()
            .map(|gradient| Matrix::from(gradient.to_param_2d()));
        let sum: Matrix = gradients_mat.sum();
        let avg = sum;
        Box::new(avg) 
    }
    fn get_gradients(&self, data: &Box<dyn Input>, data_at: &Box<dyn Input>, errors: &Box<dyn Input>) -> GradientPair {
        let gradient = self.activation_fn.apply_fn(Matrix::from(data.to_param_2d()));
        let errors_mat = Matrix::from(errors.to_param_2d());
        let mut gradients_mat = Matrix::from(gradient.to_param_2d());
        let mut data_mat = Matrix::from(data_at.to_param_2d());

        gradients_mat = gradients_mat.dot_multiply(&errors_mat);// * self.learning_rate; 

        if gradients_mat.columns != data_mat.rows {
            data_mat = data_mat.transpose();
        }

        let weight_gradient = gradients_mat.clone() * &(data_mat.clone());

        GradientPair(Box::new(gradients_mat), Box::new(weight_gradient))
    }
    fn get_data(&self) -> Box<dyn Input>{
        Box::new(self.data.clone())
    }
    fn set_data(&mut self, data: &Box<dyn Input>) {
        self.data = Matrix::from(data.to_param_2d())
    }
    ///Moves the DNN forward through the weights and biases of this current layer
    ///Maps an activation function and then returns the resultant Matrix
    fn forward(&self, inputs: &Box<dyn Input>) -> Box<dyn Input> {
        let new_data = self.activation_fn.apply_fn(self.weights.clone() * &Matrix::from(inputs.to_param().to_param_2d()).transpose() + &self.biases);

        Box::new(new_data)
    }
    fn update_errors(&self, errors: Box<dyn Input>) -> Box<dyn Input> {
        let errors_mat = Matrix::from(errors.to_param_2d());
        Box::new(self.weights.clone().transpose() * &errors_mat)
    }
    ///Does Back Propegation according to simple Dense network rules
    ///Finds the error of the previous layer and returns what the updated weights and biases should
    ///be in that layer, updates the gradients and errors to move backwards once
    fn backward(&mut self, gradients: Box<dyn Input>, errors: Box<dyn Input>, data: Box<dyn Input>) -> Box<dyn Input> {
        let mut gradients_mat: Matrix = Matrix::from(gradients.to_param_2d());
        let mut errors_mat = Matrix::from(errors.to_param_2d());
        let data_mat = Matrix::from(data.to_param_2d());

        gradients_mat = gradients_mat.dot_multiply(&errors_mat);
        errors_mat = self.weights.clone().transpose() * &errors_mat;

        self.loss = 0.0;
        errors_mat.to_param().iter().for_each(|error| {
            self.loss += error.powi(2);
        });

        self.loss = self.loss / errors_mat.to_param().len() as f32;

        self.time += 1;

        let weight_gradient = gradients_mat.clone() * &(data_mat.clone().transpose()) * self.learning_rate;
        gradients_mat = gradients_mat * self.learning_rate;

        self.m_weights = self.m_weights.clone() * self.beta1 + &(weight_gradient.clone() * (1.0 - self.beta1));
        self.v_weights = self.v_weights.clone() * self.beta2 + &((weight_gradient^2) * (1.0 - self.beta2));

        self.m_biases = self.m_biases.clone() * self.beta1 + &(gradients_mat.clone() * (1.0 - self.beta1));
        self.v_biases = self.v_biases.clone() * self.beta2 + &((gradients_mat.clone()^2) * (1.0 - self.beta2));

        let m_weights_hat = self.m_weights.clone() / (1.0 - self.beta1.powi(self.time as i32));
        let v_weights_hat = self.v_weights.clone() / (1.0 - self.beta2.powi(self.time as i32));

        let m_bias_hat = self.m_biases.clone() / (1.0 - self.beta1.powi(self.time as i32));
        let v_bias_hat = self.v_biases.clone() / (1.0 - self.beta2.powi(self.time as i32));

        let weights_update = m_weights_hat.clone() / &(v_weights_hat.sqrt() + self.epsilon);
        let bias_update = m_bias_hat.clone() / &(v_bias_hat.sqrt() + self.epsilon);

        self.biases = self.biases.clone() - &bias_update;
        self.weights = self.weights.clone() - &weights_update;

        Box::new(errors_mat)
    }

    fn update_gradient(&self) -> Box<dyn Input> {
        Box::new(self.activation_fn.apply_derivative(self.data.clone()))
    }

    fn get_activation(&self) -> Option<Activations> {
        Some(self.activation_fn.clone())
    }
    fn shape(&self) -> (usize, usize, usize){
        (self.weights.columns, 1, 1)
    }
    fn get_loss(&self) -> f32{
        self.loss
    }
    fn get_weights(&self) -> Box<dyn Input> {
        self.weights.to_box()
    }
    fn get_biases(&self) -> Box<dyn Input>{
        self.biases.to_box()
    }
}

#[cfg(test)]
mod test{
    use crate::core::{layer::layers::Layer, data::input::Input};

    use super::Dense;

    #[test]
    fn dense_shape() {
        let new_dense = Dense::new_ser(10, 2, vec![0.0; 20], vec![0.0; 10]);
        assert_eq!(new_dense.shape(), (2, 1, 1));
    }
    #[test]
    fn forward_shape(){
        let new_dense = Dense::new_ser(10, 2, vec![0.0; 20], vec![0.0; 10]);
        let input_test: Box<dyn Input> = Box::new(vec![1.0,1.0]);
        let input_res = new_dense.forward(&input_test);
        assert_eq!(input_res.shape(), (10, 1, 1));
    }
    #[test]
    #[should_panic]
    fn forward_panic(){
        let new_dense = Dense::new_ser(5, 4, vec![0.0; 20], vec![0.0; 5]);
        let test: Box<dyn Input> = Box::new(vec![0.0; 5]);
        new_dense.forward(&test);
    }
    #[test]
    fn forward_result(){
        let new_dense = Dense::new_ser(2,3, vec![1.0,1.0,2.0,5.0,0.75,6.1], vec![0.0; 2]);
        let test: Box<dyn Input> = Box::new(vec![0.5, 6.0, 3.0]);

        let res = new_dense.forward(&test);
        assert_eq!(res.to_param_2d(), vec![vec![0.9999963], vec![1.0]]);
    }

}