ksnn 0.3.0

ksnn, or Kosiorek's Simple Neural Networks, is a crate that simplifies the creation, training, and validation of a neural network. The crate is heavily inspired by "Neural Networks from Scratch in Python" by Harrison Kinsley & Daniel Kukieła.
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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320

//! activation_functions
//! 
//! All activation and loss functions that are needed for the networks are stored in this module.

use ndarray::Array;
use ndarray::Array2;
use ndarray::Axis;
use serde::{Serialize, Deserialize};
use crate::network_layers::DenseLayer;

#[derive(Serialize, Deserialize, Debug)]
/// A enum of pre-made activation functions that can be used in a neural network.
pub enum ActivationFunctions {
    ActivationReLU(ActivationReLU),
    SoftmaxLossCC(SoftmaxLossCC),
}

impl ActivationFunctions {
    pub fn forward(&mut self, inputs: &Array2<f64>, y_true: &Array2<i32>) {
        match self {
            ActivationFunctions::ActivationReLU(activation_re_lu) => activation_re_lu.forward(&inputs),
            ActivationFunctions::SoftmaxLossCC(softmax_loss_cc) => softmax_loss_cc.forward(&inputs, y_true),
        }
    }
    pub fn backward(&mut self, inputs: &Array2<f64>, y_true: &Array2<i32>) {
        match self {
            ActivationFunctions::ActivationReLU(activation_re_lu) => activation_re_lu.backward(&inputs),
            ActivationFunctions::SoftmaxLossCC(softmax_loss_cc) => softmax_loss_cc.backward(&inputs, y_true),
        }
    }
    pub fn get_outputs(&mut self) -> &Array2<f64> {
        match self {
            ActivationFunctions::ActivationReLU(activation_re_lu) => activation_re_lu.outputs.as_ref().unwrap(),
            ActivationFunctions::SoftmaxLossCC(softmax_loss_cc) => softmax_loss_cc.outputs.as_ref().unwrap(),
        }
    }
    pub fn get_data_loss(&mut self) -> Result<f64, &'static str> {
        match self {
            ActivationFunctions::SoftmaxLossCC(softmax_loss_cc) => Ok(softmax_loss_cc.data_loss.unwrap()),
            _ => return Err("Final activation function does not have a 'data_loss' value, 
                            consider using the 'SoftmaxLossCC' activation function as your final activation function."),
        }
    }
    pub fn get_regularization_loss(&mut self, layer: &DenseLayer) -> Result<f64, &'static str> {
        match self {
            ActivationFunctions::SoftmaxLossCC(softmax_loss_cc) => Ok(softmax_loss_cc.loss.regularization_loss(layer)),
            _ => return Err("Final activation function does not have a 'data_loss' value, 
                            consider using the 'SoftmaxLossCC' activation function as your final activation function."),
        }
    }
    pub fn get_dinputs(&mut self) -> &Array2<f64> {
        match self {
            ActivationFunctions::ActivationReLU(activation_re_lu) => activation_re_lu.dinputs.as_ref().unwrap(),
            ActivationFunctions::SoftmaxLossCC(softmax_loss_cc) => softmax_loss_cc.dinputs.as_ref().unwrap(),
        }
    }
}

#[derive(Serialize, Deserialize, Debug)]
/// An activation function used in the hidden layers of this network, 
/// the Rectified Linear Activation Function.
pub struct ActivationReLU {
    inputs: Option<Array2<f64>>,
    outputs: Option<Array2<f64>>,
    dinputs: Option<Array2<f64>>,
}

impl ActivationReLU {
    pub fn new() -> Self {
        ActivationReLU {
            inputs: None,
            outputs: None,
            dinputs: None,
        }
    }
    fn forward(&mut self, inputs: &Array2<f64>) {
        // If x <= 0, x equals 0, otherwise x = x
        let outputs = inputs.map(|x| x.max(0.0));
        let inputs = inputs.map(|x| *x);
        // Remember input values
        self.inputs = Some(inputs);
        self.outputs = Some(outputs);
    }
    fn backward(&mut self, dvalues: &Array2<f64>) {
        // drelu is a gradient with 0 as values where input values are negative
        let mut drelu: Array2<f64> = Array::zeros((dvalues.shape()[0], dvalues.shape()[1]));
        for outer in 0..drelu.shape()[0] {
            for inner in 0..drelu.shape()[1] {
                if self.inputs.as_ref().unwrap()[(outer, inner)] >= 0.0 {
                    drelu[(outer, inner)] = dvalues[(outer, inner)];
                }
            }
        }
        self.dinputs = Some(drelu);
    }
}

#[derive(Serialize, Deserialize, Debug)]
/// The softmax actiation function is used for the output layer and 
/// returns probabilities for what each input samples classification is.
pub struct ActivationSoftmax {
    outputs: Option<Array2<f64>>,
    dinputs: Option<Array2<f64>>,
}

impl ActivationSoftmax {
    fn new() -> Self {
        ActivationSoftmax {
            outputs: None,
            dinputs: None,
        }
    }
    fn forward(&mut self, inputs: &Array2<f64>) {
        let mut probabilities = Array2::<f64>::zeros(inputs.raw_dim());
        for (in_row, mut out_row) in inputs.axis_iter(Axis(0)).zip(probabilities.axis_iter_mut(Axis(0))) {
            let mut max = 0.0;
            for col in in_row.iter() {
                if col > &max {
                    max = *col;
                }
            }
            // Normlaize values between 0 and 1 before being exponentialized
            let exp = in_row.map(|x| (x - max).exp());
            let sum = exp.sum();
            out_row.assign(&(exp / sum));
        }
        self.outputs = Some(probabilities);
    }
    fn backward(&mut self, dvalues: &Array2<f64>) {
        self.dinputs = Some(Array::zeros((dvalues.shape()[0], dvalues.shape()[1])));
        let arr_sh1 = dvalues.shape()[1];

        // An array the same shape as a flattened version of the self.output array
        let mut temp_output_arr: Array2<f64> = Array::zeros((arr_sh1, 1));
        let mut temp_dvalues_arr: Array2<f64> = Array::zeros((arr_sh1, 1));
        let mut l_cntr = 0;
        let mut iter_cntr = 0;

        for (_index, (single_output, single_dvalues)) in self.outputs.as_ref().unwrap().iter().zip(dvalues).enumerate() {
            temp_output_arr[(l_cntr, 0)] = *single_output;
            temp_dvalues_arr[(l_cntr, 0)] = *single_dvalues;
            l_cntr += 1;

            if l_cntr == arr_sh1 {
                let mut jacobian_matrix: Array<f64, ndarray::Dim<[usize; 2]>> = Array::zeros((temp_output_arr.shape()[0], temp_output_arr.shape()[0]));
                
                for outer in 0..jacobian_matrix.shape()[0] {
                    for inner in 0..jacobian_matrix.shape()[1] {
                        if inner == outer {
                            jacobian_matrix[(outer, inner)] = temp_output_arr[(outer, 0)];
                        }
                    }
                }

                // Calculate jacobian matrix of the output
                jacobian_matrix = jacobian_matrix - temp_output_arr.dot(&temp_output_arr.t());

                // Calculate sample-wise gradient and add it to the array of sample gradients
                let dinputs = jacobian_matrix.dot(&temp_dvalues_arr);

                // Add sample-wise gradient to the array of sample gradients
                for inner in 0..arr_sh1 {
                    self.dinputs.as_mut().unwrap()[(iter_cntr, inner)] = dinputs[(inner, 0)];
                }
                
                temp_output_arr = Array::zeros((arr_sh1, 1));
                temp_dvalues_arr = Array::zeros((arr_sh1, 1));
                l_cntr = 0;
                iter_cntr += 1;
            }
        }
    }
}

#[derive(Serialize, Deserialize, Debug)]
/// Calculates the networks current total loss.
pub struct LossCategoricalCrossentropy {
    dinputs: Option<Array2<f64>>,
    data_loss: Option<f64>,
}

impl LossCategoricalCrossentropy {
    fn new() -> Self {
        LossCategoricalCrossentropy {
            dinputs: None,
            data_loss: None,
        }
    }

    /// Calculates loss using the sparse categorical cross entropy equation.
    fn calculate(&mut self, y_pred: &Array2<f64>, y_true: &Array2<i32>) {
        let data_loss = self.forward(y_pred, y_true);
        self.data_loss = Some(data_loss);
    }

    fn forward(&mut self, y_pred: &Array2<f64>, y_true: &Array2<i32>) -> f64 {
        // clip values in the array between 1e-7 and 1-1e-7
        let y_pred = &y_pred.map(|x| x.max(0.0000001));
        let y_pred = y_pred.map(|x| x.min(0.9999999));

        let mut sum: f64 = 0.0;
        
        let correct_confidences = (y_pred * y_true.map(|x| *x as f64)).sum_axis(Axis(1));
        
        let mut negative_log_likelihoods: Array<f64, ndarray::Dim<[usize; 1]>> = Array::zeros(correct_confidences.shape()[0]);

        // Gets the sum of the negative log likelihoods for each sample
        for i in 0..correct_confidences.len() {
            negative_log_likelihoods[i] = -correct_confidences[i].ln();
            sum += negative_log_likelihoods[i];
        }

        // Calculate total data loss
        let mean = sum / y_true.shape()[0] as f64;

        let data_loss = mean;
        data_loss
    }

    fn regularization_loss(&mut self, layer: &DenseLayer) -> f64 {
        let mut regularization_loss = 0.0;

        if layer.weight_regularizer_l1 > 0.0 {
            regularization_loss += layer.weight_regularizer_l1 * (&layer.weights.map(|x| x.abs())).sum();
        }

        if layer.weight_regularizer_l2 > 0.0 {
            regularization_loss += layer.weight_regularizer_l2 * (&layer.weights * &layer.weights).sum();
        }

        if layer.bias_regularizer_l1 > 0.0 {
            regularization_loss += layer.bias_regularizer_l1 * (&layer.biases.map(|x| x.abs())).sum();
        }

        if layer.bias_regularizer_l2 > 0.0 {
            regularization_loss += layer.bias_regularizer_l2 * (&layer.biases * &layer.biases).sum();
        }

        regularization_loss
    }

    fn backward(&mut self, dvalues: &Array2<f64>, y_true: &Array2<i32>) {
        // Number of samples
        let samples = dvalues.shape()[0];
        // Number of labels in every sample
        let _labels = dvalues.shape()[1];
        
        // Calculate gradient
        self.dinputs = Some(-y_true.map(|x| *x as f64) / dvalues);
        // Normalize gradient
        self.dinputs = Some(self.dinputs.as_ref().unwrap() / samples as f64);
    }
}

#[derive(Serialize, Deserialize, Debug)]
/// Full name ActivationSoftmaxLossCategoricalCrossentropy; A combination of the softmax activation function and the categorical loss 
/// crossentropy function. The combination allows for a simplified implementation and overall faster calculation time.
pub struct SoftmaxLossCC {
    activation: ActivationSoftmax,
    loss: LossCategoricalCrossentropy,
    outputs: Option<Array2<f64>>,
    dinputs: Option<Array2<f64>>,
    data_loss: Option<f64>,
}

impl SoftmaxLossCC {
    pub fn new() -> Self {
        SoftmaxLossCC {
            activation: ActivationSoftmax {
                outputs: None, 
                dinputs: None,
            },
            loss: LossCategoricalCrossentropy {
                dinputs: None,
                data_loss: None,
            },
            outputs: None,
            dinputs: None,
            data_loss: None,
        }
    }
    fn forward(&mut self, inputs: &Array2<f64>, y_true: &Array2<i32>) {
        // Output layers activation function
        self.activation.forward(inputs);
        // Set the output
        let outputs = self.activation.outputs.as_ref().unwrap().map(|x| *x);
        self.outputs = Some(outputs);
        // Calculate and return loss value
        self.loss.calculate(self.outputs.as_ref().unwrap(), y_true);
        self.data_loss = self.loss.data_loss;
    }
    fn backward(&mut self, dvalues: &Array2<f64>, y_true: &Array2<i32>) {
        let mut class_targets: Array<i32, ndarray::Dim<[usize; 1]>> = Array::zeros(dvalues.shape()[0]);
        let mut dinput = dvalues.map(|x| *x);
        // Gets the number of samples (or rows)
        let samples = class_targets.len();

        let mut clas_max_num: i32 = -999999999;

        // Tunrs one hot encoded values into discrete values (this loop is basically python numpys argmax on axis=1)
        for outer in 0..dinput.shape()[0] {
            for inner in 0..dinput.shape()[1] {
                if y_true[(outer, inner)] > clas_max_num {
                    clas_max_num = y_true[(outer, inner)];
                    class_targets[outer] = inner as i32;
                }
            }
            clas_max_num = -999999999;
        }

        // Calculate the gradient
        for outer in 0..dinput.shape()[0] {
            dinput[(outer, class_targets[outer] as usize)] -= 1.0;
        }

        // Normalize the gradient
        dinput = dinput / samples as f64;
        self.dinputs = Some(dinput);
    }
}