sproutml 0.1.1

A simple Machine Learning Library built in Rust
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
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401

use std::vec;
use serde_derive::*;

use crate::
    {activation::
        {Activation, ActivationFunction}, 
        conv_params::
            {ConvParams, PaddingType}, 
        dense_params::
            DenseParams
    };

#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub enum LayerType {
    Dense,
    Convolutional,
    Pooling
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Layer {
    pub activation: Activation,
    pub layer_type: LayerType,
    pub conv_params: Option<ConvParams>,
    pub dense_params: Option<DenseParams>,
}

impl Layer {
    pub fn dense(nodes: [usize; 2], activation_fn: ActivationFunction) -> Self {
        let dense_params = Some(DenseParams::new(nodes[0], nodes[1]));

        let mut layer = Layer {
            dense_params,
            activation: Activation::new(activation_fn),
            layer_type: LayerType::Dense,
            conv_params: None
        };
        layer.dense_params.as_mut().unwrap().init(layer.activation.function.clone());
        layer
    }

    pub fn conv(kernel: usize, padding_type: PaddingType, stride: usize, activation_fn: ActivationFunction) -> Self {
        let conv_params = Some(ConvParams::new(kernel, padding_type, stride));
        let layer = Layer {
            dense_params: None,
            activation: Activation::new(activation_fn),
            layer_type: LayerType::Convolutional,
            conv_params
        };
        layer
    }

    pub fn pool(kernel: usize, stride: usize) -> Self {
        let conv_params = Some(ConvParams::new(kernel, PaddingType::Valid, stride));
        let layer = Layer {
            dense_params: None,
            activation: Activation::new(ActivationFunction::ReLU),
            layer_type: LayerType::Pooling,
            conv_params
        };
        layer
    }

    pub fn conv_forward(&mut self, inputs: Vec<Vec<Vec<f64>>>) -> Vec<Vec<Vec<f64>>> {
        let params = self.conv_params.as_mut().unwrap();
        params.inputs = inputs;
        if params.padding_type == PaddingType::Valid {
            params.data = params.inputs.clone();
        }
        params.add_padding(self.activation.function.clone());
        let output_dims = params.get_output_dims();

        let mut weighted_inputs = vec![vec![vec![0.0; output_dims[0]]; output_dims[1]]; params.data.len()];
        let img = &params.data;

        let mut activation = vec![vec![vec![0.0; output_dims[0]]; output_dims[1]]; params.data.len()];

        for i in 0..img.len() { //each channel
            for j in (0..weighted_inputs[i].len()) { //each img row
                if j + params.kernel > params.data[i].len() {
                    break;
                }
                for k in (0..weighted_inputs[i][j].len()) { //each img column
                    if k + params.kernel > params.data[i][0].len() {
                        break;
                    }
                    for kern_row in 0..params.kernel { //Kernel rows
                        for kern_col in 0..params.kernel { //Kernel Columns
                            weighted_inputs[i][j][k] += (img[i][j * params.stride + kern_row][k * params.stride + kern_col] * params.weights[i][kern_row][kern_col]);
                            weighted_inputs[i][j][k] += params.bias;
                        }
                    }
                }
            }
            for j in 0..weighted_inputs[i].len() { 
                activation[i][j] = self.activation.function(weighted_inputs[i][j].clone());
            }
        }
        params.outputs = activation.clone();
        activation
    }

    pub fn pool_forward(&mut self, inputs: Vec<Vec<Vec<f64>>>) -> Vec<Vec<Vec<f64>>> {
        let params = self.conv_params.as_mut().unwrap();
        params.inputs = inputs;
        params.add_padding(self.activation.function.clone());

        let output_dims = params.get_output_dims();
        let img = params.inputs.clone();

        let mut output = vec![vec![vec![0.0; output_dims[0]]; output_dims[1]]; params.data.len()];

        for i in 0..img.len() { //each channel
            for j in (0..output[i].len()) { //each output img row
                if (j) * params.kernel >= img[i].len() {
                    break;
                }
                for k in (0..output[i][j].len()) { //each output img column
                    if (k) * params.kernel >= img[i][0].len() {
                        break;
                    }
                    let mut max = f64::NEG_INFINITY;
                    for kern_row in 0..params.kernel { //Kernel rows
                        let row_i = j * params.stride + kern_row;
                        if row_i >= img[i].len() {
                            break;
                        }
                        for kern_col in 0..params.kernel { //Kernel Columns
                            let col_i = k * params.stride + kern_col;
                            if col_i >= img[i][0].len() {
                                break;
                            }
                            let val = img[i][row_i][col_i];
                            if val > max {
                                max = val;
                            }
                        }
                    }
                    output[i][j][k] = max;
                }
            }
        }
        params.outputs = output.clone();
        output
    }

    pub fn dense_forward(&mut self, inputs: Vec<f64>) -> Vec<f64> {
        self.dense_params.as_mut().unwrap().inputs = inputs.clone();
        let mut weighted_inputs = self.dense_params.as_mut().unwrap().biases.clone();
        for i in 0..self.dense_params.as_mut().unwrap().nodes_out {
            for j in 0..self.dense_params.as_mut().unwrap().nodes_in {
                weighted_inputs[i] += inputs[j] * self.dense_params.as_mut().unwrap().weights[j][i];
            }
        }

        let activation = self.activation.function(weighted_inputs);
        self.dense_params.as_mut().unwrap().outputs = activation.clone();
        activation
    }

    pub fn conv_backward(&mut self, errors: Vec<Vec<Vec<f64>>>, learning_rate: f64) -> Vec<Vec<Vec<f64>>> {
        let params = self.conv_params.as_mut().unwrap();
        let channels = params.data.len();
        let mut delta_output = errors;
        let mut weight_gradients = vec![vec![vec![0.0; params.kernel]; params.kernel]; channels];
        let img = params.data.clone();
        let mut avg_bias_gradient = 0.0;
        let mut next_delta = vec![]; //3x3
        let kernel = params.kernel;

        for i in 0..channels { //each channel
            for j in 0..delta_output[i].len() {
                if self.activation.function == ActivationFunction::SoftMax {
                    break;
                }
                let activation_derivatives = self.activation.derivative(params.outputs[i][j].clone());
                for k in 0..delta_output[i][j].len() {
                    delta_output[i][j][k] *= activation_derivatives[k];
                }
            }

            for j in (0..delta_output[i].len()) { //each img row
                if j + kernel == img[i].len() {
                    break;
                }
                for k in (0..delta_output[i][j].len()) { //each img column
                    if k + kernel == img[i][j].len() {
                        break;
                    }
                    for kern_row in 0..kernel { //Kernel rows
                        for kern_col in 0..kernel { //Kernel Columns
                            weight_gradients[i][kern_row][kern_col] += (img[i][j * params.stride + kern_row][k * params.stride + kern_col] * delta_output[i][j][k]);
                        }
                    }
                }
            }
            for j in 0..params.weights[i].len() {
                for k in 0..params.weights[i][j].len() {
                    params.weights[i][j][k] -= learning_rate *  weight_gradients[i][j][k];
                }
            }
            // Update bias using the average of the gradients
            for j in 0..delta_output[i].len() {
                for k in 0..delta_output[i][j].len() {
                    avg_bias_gradient += delta_output[i][j][k];
                }
            }
            avg_bias_gradient /= (delta_output[i].len() * delta_output[i][0].len()) as f64;

            let padded_gradients = Self::add_padding_matrix(kernel - 1, &delta_output[i]);

            let rotated_kernel = Self::rotate180(&params.weights[i]);
            let mut gradient_channel = vec![];
            for j in (0..padded_gradients.len()) { //each img row
                if j + params.kernel > padded_gradients.len() {
                    break;
                }
                let mut gradient_row = vec![];
                for k in (0..padded_gradients[j].len()) { //each img column
                    if k + params.kernel > padded_gradients[0].len() {
                        break;
                    }
                    let mut sum = 0.0;
                    for kern_row in 0..params.kernel { //Kernel rows
                        for kern_col in 0..params.kernel { //Kernel Columns
                            sum += (rotated_kernel[kern_row][kern_col] * padded_gradients[j * params.stride + kern_row][k * params.stride + kern_col]);
                        }
                    }
                    gradient_row.push(sum);
                }
                gradient_channel.push(gradient_row);
            }
            next_delta.push(gradient_channel);
        }
        params.bias -= learning_rate * avg_bias_gradient;

        next_delta
    }

    pub fn pool_backward(&mut self, errors: Vec<Vec<Vec<f64>>>) -> Vec<Vec<Vec<f64>>> {
        let params = self.conv_params.as_mut().unwrap();
        let channels = params.data.len();
        let delta_output = errors;
        let img = params.data.clone();
        let kernel = params.kernel;
        let inputs = params.inputs.clone();
        let mut indecies = vec![[0, 0]; delta_output.len() * delta_output[0].len() + delta_output[0][0].len()];

        let mut next_delta = vec![vec![vec![0.0; inputs[0][0].len()]; inputs[0].len()]; inputs.len()]; //4x4

        for i in 0..channels { //each channel
            for j in (0..delta_output[i].len()) { //each img row
                if j * params.stride >= img[i].len() {
                    break;
                }
                for k in (0..delta_output[i][j].len()) { //each img column
                    if k * params.stride >= img[i][j].len() {
                        break;
                    }
                    let mut max = f64::NEG_INFINITY;
                    let mut max_indx = [j * params.stride, k * params.stride];
                    for kern_row in 0..kernel { //Kernel rows
                        let row_i = j * params.stride + kern_row;
                        if row_i >= img[i].len() {
                            break;
                        }
                        for kern_col in 0..kernel { //Kernel Columns
                            let col_i = k * params.stride + kern_col;
                            if col_i >= img[i][0].len() {
                                break;
                            }
                            let val = img[i][row_i][col_i];
                            if val > max {
                                max_indx = [row_i, col_i]; //each kernel
                                max = val;
                            }
                        }
                        // indecies.push(max_indx); //each channel
                    }
                    next_delta[i][max_indx[0]][max_indx[1]] = delta_output[i][j][k];
                }
            }
        }
        next_delta
    }

    pub fn rotate180(mat: &Vec<Vec<f64>>) -> Vec<Vec<f64>> {
        let mut new_mat = mat.clone();
        for row in &mut new_mat {
            row.reverse();
        }
        new_mat.reverse();
        new_mat
    }

    pub fn add_padding_matrix(padding: usize, matrix: &Vec<Vec<f64>>) -> Vec<Vec<f64>> {
        let height = matrix.len();
        let width = matrix[0].len();
        let padded_height = height + 2 * padding;
        let padded_width = width + 2 * padding;
        
        let mut padded_image = vec![vec![0.0; padded_width]; padded_height];

        for i in 0..height {
            for j in 0..width {
                padded_image[i + padding][j + padding] = matrix[i][j];
            }
        }
        padded_image
    }

    pub fn dense_backward(&mut self, errors: Vec<f64>, learning_rate: f64) -> Vec<f64> {
        let mut delta_output = errors.clone();

        let activation_gradients = self.activation.derivative(self.dense_params.as_mut().unwrap().outputs.clone());
        for i in 0..delta_output.len() {
            if self.activation.function == ActivationFunction::SoftMax {
                break;
            }
            delta_output[i] *= activation_gradients[i];
        }

        for i in 0..self.dense_params.as_mut().unwrap().weights.len() {
            for j in 0..self.dense_params.as_mut().unwrap().weights[i].len() {
                self.dense_params.as_mut().unwrap().weights[i][j] -= learning_rate * (self.dense_params.as_mut().unwrap().inputs[i] * delta_output[j]);
            }
        }

        for i in 0..self.dense_params.as_mut().unwrap().biases.len() {
            self.dense_params.as_mut().unwrap().biases[i] -= learning_rate * delta_output[i];
        }

        let mut next_delta = vec![0.0; self.dense_params.as_mut().unwrap().nodes_in];
        for i in 0..self.dense_params.as_mut().unwrap().weights.len() {
            for j in 0..self.dense_params.as_mut().unwrap().weights[i].len() {
                next_delta[i] += (self.dense_params.as_mut().unwrap().weights[i][j] * delta_output[j] * self.dense_params.as_mut().unwrap().inputs[i]);
            }   
        }

        next_delta
    }

    pub fn get_dense_weights(&self) -> Vec<Vec<f64>> {
        return self.dense_params.as_ref().unwrap().weights.clone();
    }

    pub fn get_dense_biases(&self) -> Vec<f64> {
        return self.dense_params.as_ref().unwrap().biases.clone();
    }

    pub fn get_conv_weights(&self) -> Vec<Vec<Vec<f64>>> {
        return self.conv_params.as_ref().unwrap().weights.clone();
    }

    pub fn get_conv_biases(&self) -> f64 {
        return self.conv_params.as_ref().unwrap().bias.clone();
    }

    pub fn get_nodes(&self) -> usize {
        if self.layer_type == LayerType::Dense {
            return self.dense_params.as_ref().unwrap().nodes_out.clone();
        }
        0
    }

    pub fn get_input_nodes(&self) -> usize {
        if self.layer_type == LayerType::Dense {
            return self.dense_params.as_ref().unwrap().nodes_in.clone();
        }
        0
    }

    pub fn get_dense_outputs(&self) -> Vec<f64> {
        self.dense_params.as_ref().unwrap().outputs.clone()
    }

    pub fn get_conv_outputs(&self) -> Vec<Vec<Vec<f64>>> {
        self.conv_params.as_ref().unwrap().outputs.clone()
    }

    pub fn get_layer_type(&self) -> LayerType {
        self.layer_type.clone()
    }

    pub fn reset(&mut self) {
        if self.layer_type == LayerType::Dense {
            self.dense_params.as_mut().unwrap().inputs = vec![];
            self.dense_params.as_mut().unwrap().outputs = vec![];
            self.dense_params.as_mut().unwrap().init(self.activation.function.clone());
        } else {
            self.conv_params.as_mut().unwrap().inputs = vec![];
            self.conv_params.as_mut().unwrap().outputs = vec![];
            self.conv_params.as_mut().unwrap().init(self.activation.function.clone());
        }
    }

    pub fn set_params(&mut self, weights: Vec<Vec<f64>>, biases: Vec<f64>) {
        self.dense_params.as_mut().unwrap().weights = weights;
        self.dense_params.as_mut().unwrap().biases = biases;
    }
}