runn 0.1.1

Runn is a feature-rich, easy-to-use library for building, training, and evaluating feed-forward neural networks 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

use super::{Optimizer, OptimizerConfig, OptimizerConfigClone};
use crate::{common::matrix::DMat, error::NetworkError, LearningRateScheduler};
use serde::{Deserialize, Serialize};
use typetag;

/// MomentumOptimizer is an implementation of the Momentum optimization algorithm.
/// Momentum is an optimization algorithm that accelerates gradients in the direction
/// of previous gradients, helping to overcome local minima and speed up convergence.
/// velocity = momentum * velocity + learning_rate * gradient
/// weight = weight - velocity
/// bias = bias - velocity
#[derive(Serialize, Deserialize, Clone)]
struct MomentumOptimizer {
    config: MomentumConfig,
    velocity_weights: DMat,
    velocity_biases: DMat,
}

impl MomentumOptimizer {
    pub fn new(config: MomentumConfig) -> Self {
        Self {
            config,
            velocity_weights: DMat::zeros(0, 0),
            velocity_biases: DMat::zeros(0, 0),
        }
    }
}

#[derive(Serialize, Deserialize, Clone)]
struct MomentumConfig {
    learning_rate: f32,
    momentum: f32,
    scheduler: Option<Box<dyn LearningRateScheduler>>,
}

#[typetag::serde]
impl OptimizerConfig for MomentumConfig {
    fn update_learning_rate(&mut self, learning_rate: f32) {
        self.learning_rate = learning_rate;
    }
    fn create_optimizer(&self) -> Box<dyn Optimizer> {
        Box::new(MomentumOptimizer::new(self.clone()))
    }
    fn learning_rate(&self) -> f32 {
        self.learning_rate
    }
}

#[typetag::serde]
impl Optimizer for MomentumOptimizer {
    fn initialize(&mut self, weights: &DMat, biases: &DMat) {
        self.velocity_weights = DMat::zeros(weights.rows(), weights.cols());
        self.velocity_biases = DMat::zeros(biases.rows(), biases.cols());
    }

    fn update(&mut self, weights: &mut DMat, biases: &mut DMat, d_weights: &DMat, d_biases: &DMat, epoch: usize) {
        if self.config.scheduler.is_some() {
            let scheduler = self.config.scheduler.as_ref().unwrap();
            self.config.learning_rate = scheduler.schedule(epoch, self.config.learning_rate);
        }
        weights.apply_with_indices(|r, c, v| {
            let velocity = &mut self.velocity_weights;
            let previous_velocity = velocity.at(r, c);
            // Calculate the new velocity using the Momentum update rule
            // velocity = (momentum * previous_velocity) + (learning_rate * gradient)
            let new_velocity =
                self.config.momentum * previous_velocity + self.config.learning_rate * d_weights.at(r, c);
            velocity.set(r, c, new_velocity);
            *v -= new_velocity
        });
        biases.apply_with_indices(|r, c, v| {
            let velocity = &mut self.velocity_biases;
            let previous_velocity = velocity.at(r, c);
            // Calculate the new velocity using the Momentum update rule
            // velocity = (momentum * previous_velocity) + (learning_rate * gradient)
            let new_velocity = self.config.momentum * previous_velocity + self.config.learning_rate * d_biases.at(r, c);
            velocity.set(r, c, new_velocity);
            *v -= new_velocity
        });
    }

    fn update_learning_rate(&mut self, learning_rate: f32) {
        self.config.learning_rate = learning_rate;
    }
}

///Builder for Momentum optimizer
/// MomentumOptimizer is an implementation of the Momentum optimization algorithm.
/// Momentum is an optimization algorithm that accelerates gradients in the direction
/// of previous gradients, helping to overcome local minima and speed up convergence.
/// velocity = momentum * velocity + learning_rate * gradient
/// weight = weight - velocity
/// bias = bias - velocity
pub struct Momentum {
    learning_rate: f32,
    momentum: f32,
    scheduler: Option<Result<Box<dyn LearningRateScheduler>, NetworkError>>,
}

impl Momentum {
    // Creates a new builder for Momentum optimizer.
    // Default values:
    // - learning_rate: 0.01
    // - momentum: 0.9
    // - scheduler: None
    fn new() -> Self {
        Self {
            learning_rate: 0.01,
            momentum: 0.9,
            scheduler: None,
        }
    }

    /// Set the learning rate.
    ///
    /// Controls the step size for parameter updates. Smaller values lead to slower but more stable convergence.
    /// # Parameters
    /// - `lr`: The learning rate value (e.g., 0.01).
    pub fn learning_rate(mut self, lr: f32) -> Self {
        self.learning_rate = lr;
        self
    }

    /// Set the momentum factor.
    ///
    /// Determines how much past gradients influence the current update. Higher values increase the effect of momentum.
    /// # Parameters
    /// - `momentum`: Momentum factor, typically in [0.0, 1.0] (e.g., 0.9).
    pub fn momentum(mut self, momentum: f32) -> Self {
        self.momentum = momentum;
        self
    }

    /// Set a learning rate scheduler.
    ///
    /// Optionally applies a scheduler to adjust the learning rate during training (e.g., exponential, step).
    /// # Parameters
    /// - `scheduler`: Learning rate scheduler to use.
    pub fn scheduler(mut self, scheduler: Result<Box<dyn LearningRateScheduler>, NetworkError>) -> Self {
        self.scheduler = Some(scheduler);
        self
    }

    fn validate(&self) -> Result<(), NetworkError> {
        if self.learning_rate <= 0.0 {
            return Err(NetworkError::ConfigError(format!(
                "Learning rate for Momentum must be greater than 0.0, but was {}",
                self.learning_rate
            )));
        }
        if self.momentum < 0.0 || self.momentum > 1.0 {
            return Err(NetworkError::ConfigError(format!(
                "Momentum for Momentum must be in [0.0, 1.0], but was {}",
                self.momentum
            )));
        }
        if let Some(ref scheduler) = self.scheduler {
            scheduler.as_ref().map_err(|e| e.clone())?;
        }
        Ok(())
    }

    pub fn build(self) -> Result<Box<dyn OptimizerConfig>, NetworkError> {
        self.validate()?;
        Ok(Box::new(MomentumConfig {
            learning_rate: self.learning_rate,
            momentum: self.momentum,
            scheduler: self.scheduler.map(|s| s.unwrap()),
        }))
    }
}

impl Default for Momentum {
    /// Creates a new Momentum builder with default values.
    /// Default values:
    /// - `learning_rate`: 0.01
    /// - `momentum`: 0.9
    /// - `scheduler`: None
    fn default() -> Self {
        Self::new()
    }
}

impl OptimizerConfigClone for MomentumConfig {
    fn clone_box(&self) -> Box<dyn OptimizerConfig> {
        Box::new(self.clone())
    }
}

#[cfg(test)]
mod tests {
    use crate::{step::Step, util::equal_approx};

    use super::*;

    #[test]
    fn test_initialize() {
        let config = MomentumConfig {
            learning_rate: 0.01,
            momentum: 0.9,
            scheduler: None,
        };
        let mut optimizer = MomentumOptimizer::new(config);
        let weights = DMat::new(2, 2, &[0.1, 0.2, 0.3, 0.4]);
        let biases = DMat::new(2, 1, &[0.1, 0.2]);
        optimizer.initialize(&weights, &biases);
        assert_eq!(optimizer.velocity_weights.rows(), 2);
        assert_eq!(optimizer.velocity_weights.cols(), 2);
        assert_eq!(optimizer.velocity_biases.rows(), 2);
        assert_eq!(optimizer.velocity_biases.cols(), 1);
    }

    #[test]
    fn test_update() {
        let config = MomentumConfig {
            learning_rate: 0.1,
            momentum: 0.9,
            scheduler: None,
        };
        let mut optimizer = MomentumOptimizer::new(config);
        let mut weights = DMat::new(2, 2, &[1.0, 1.0, 1.0, 1.0]);
        let mut biases = DMat::new(2, 1, &[1.0, 1.0]);
        let d_weights = DMat::new(2, 2, &[0.1, 0.1, 0.1, 0.1]);
        let d_biases = DMat::new(2, 1, &[0.1, 0.1]);
        optimizer.initialize(&weights, &biases);

        optimizer.update(&mut weights, &mut biases, &d_weights, &d_biases, 1);
        assert!(weights.at(0, 0) < 1.0);
        assert!(biases.at(0, 0) < 1.0);
    }

    #[test]
    fn test_update_learning_rate() {
        let config = MomentumConfig {
            learning_rate: 0.01,
            momentum: 0.9,
            scheduler: None,
        };
        let mut optimizer = MomentumOptimizer::new(config);
        optimizer.update_learning_rate(0.02);
        assert_eq!(optimizer.config.learning_rate, 0.02);
    }

    #[test]
    fn test_momentum_optimizer() {
        // Create mock parameter matrices
        let mut weights = DMat::new(2, 2, &[1.0, 2.0, 3.0, 4.0]);
        let mut biases = DMat::new(2, 1, &[1.0, 2.0]);

        // Create mock gradient matrices
        let d_weights = DMat::new(2, 2, &[0.1, 0.1, 0.1, 0.1]);
        let d_biases = DMat::new(2, 1, &[0.1, 0.1]);

        // Create an instance of the Momentum optimizer
        let config = MomentumConfig {
            learning_rate: 0.1,
            momentum: 0.9,
            scheduler: None,
        };
        let mut optimizer = MomentumOptimizer::new(config);
        optimizer.initialize(&weights, &biases);

        // Update the parameters using the mock gradients
        optimizer.update(&mut weights, &mut biases, &d_weights, &d_biases, 1);

        let expected_weights = DMat::new(2, 2, &[0.99, 1.99, 2.99, 3.99]);

        assert!(equal_approx(&weights, &expected_weights, 1e-3));
    }

    #[test]
    fn test_momentum_builder() {
        let optimizer = Momentum::default().learning_rate(0.01).momentum(0.9).build().unwrap();
        assert_eq!(optimizer.learning_rate(), 0.01);
    }

    #[test]
    fn test_momentum_builder_invalid_learning_rate() {
        let result = Momentum::default().learning_rate(-0.01).build();
        assert!(result.is_err());
        if let Err(e) = result {
            assert_eq!(
                e.to_string(),
                "Configuration error: Learning rate for Momentum must be greater than 0.0, but was -0.01"
            );
        }
    }

    #[test]
    fn test_momentum_builder_invalid_scheduler() {
        let result = Momentum::new()
            .momentum(0.5)
            .scheduler(Step::default().decay_rate(5.0).build())
            .build();
        assert!(result.is_err());
        if let Err(e) = result {
            assert_eq!(
                e.to_string(),
                "Configuration error: Decay rate for Step must be in the range (0, 1), but was 5"
            );
        }
    }
}