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
/// RMSprop optimizer implementation
///
/// RMSprop maintains a moving average of squared gradients and divides the gradient
/// by the square root of this average. This helps handle different scales of gradients
/// and can lead to faster convergence than standard SGD.
#[derive(Debug, Clone)]
pub struct RMSprop {
/// Learning rate
pub learning_rate: f64,
/// Decay rate for moving average
pub decay_rate: f64,
/// Small constant for numerical stability
pub epsilon: f64,
/// Moving average of squared gradients for each parameter
cache: Vec<f64>,
}
impl RMSprop {
/// Creates a new RMSprop optimizer
///
/// # Arguments
///
/// * `learning_rate` - Step size for updates
/// * `decay_rate` - Rate for moving average (typically 0.9)
/// * `epsilon` - Small constant for numerical stability
///
/// # Example
///
/// ```
/// use algos::ml::deep::rmsprop::RMSprop;
/// let optimizer = RMSprop::new(0.001, 0.9, 1e-8);
/// ```
pub fn new(learning_rate: f64, decay_rate: f64, epsilon: f64) -> Self {
assert!(learning_rate > 0.0, "Learning rate must be positive");
assert!(
decay_rate > 0.0 && decay_rate < 1.0,
"Decay rate must be between 0 and 1"
);
assert!(epsilon > 0.0, "Epsilon must be positive");
RMSprop {
learning_rate,
decay_rate,
epsilon,
cache: Vec::new(),
}
}
/// Initializes the optimizer for a given number of parameters
///
/// # Arguments
///
/// * `param_count` - Number of parameters to optimize
pub fn initialize(&mut self, param_count: usize) {
self.cache = vec![0.0; param_count];
}
/// Updates parameters using RMSprop
///
/// # Arguments
///
/// * `params` - Parameters to update
/// * `grads` - Gradients for each parameter
///
/// # Returns
///
/// * Updated parameters
pub fn update(&mut self, params: &[f64], grads: &[f64]) -> Vec<f64> {
assert_eq!(
params.len(),
grads.len(),
"Parameters and gradients must have same length"
);
if self.cache.is_empty() {
self.initialize(params.len());
}
let mut updated_params = params.to_vec();
for i in 0..params.len() {
// Update moving average of squared gradients
self.cache[i] =
self.decay_rate * self.cache[i] + (1.0 - self.decay_rate) * grads[i].powi(2);
// Update parameters
updated_params[i] -=
self.learning_rate * grads[i] / (self.cache[i] + self.epsilon).sqrt();
}
updated_params
}
/// Updates 2D parameters (e.g., weight matrices)
///
/// # Arguments
///
/// * `params` - 2D parameters to update
/// * `grads` - 2D gradients for each parameter
///
/// # Returns
///
/// * Updated 2D parameters
pub fn update_2d(&mut self, params: &[Vec<f64>], grads: &[Vec<f64>]) -> Vec<Vec<f64>> {
assert_eq!(
params.len(),
grads.len(),
"Parameters and gradients must have same dimensions"
);
let total_params: usize = params.iter().map(|row| row.len()).sum();
if self.cache.is_empty() {
self.initialize(total_params);
}
let mut updated_params = params.to_vec();
let mut cache_idx = 0;
for i in 0..params.len() {
assert_eq!(
params[i].len(),
grads[i].len(),
"Parameter and gradient rows must have same length"
);
for j in 0..params[i].len() {
// Update moving average of squared gradients
self.cache[cache_idx] = self.decay_rate * self.cache[cache_idx]
+ (1.0 - self.decay_rate) * grads[i][j].powi(2);
// Update parameters
updated_params[i][j] -= self.learning_rate * grads[i][j]
/ (self.cache[cache_idx] + self.epsilon).sqrt();
cache_idx += 1;
}
}
updated_params
}
/// Updates 4D parameters (e.g., convolutional filters)
///
/// # Arguments
///
/// * `params` - 4D parameters to update
/// * `grads` - 4D gradients for each parameter
///
/// # Returns
///
/// * Updated 4D parameters
pub fn update_4d(
&mut self,
params: &[Vec<Vec<Vec<f64>>>],
grads: &[Vec<Vec<Vec<f64>>>],
) -> Vec<Vec<Vec<Vec<f64>>>> {
assert_eq!(
params.len(),
grads.len(),
"Parameters and gradients must have same dimensions"
);
let total_params: usize = params
.iter()
.flat_map(|x| x.iter())
.flat_map(|x| x.iter())
.map(|x| x.len())
.sum();
if self.cache.is_empty() {
self.initialize(total_params);
}
let mut updated_params = params.to_vec();
let mut cache_idx = 0;
for i in 0..params.len() {
assert_eq!(params[i].len(), grads[i].len());
for j in 0..params[i].len() {
assert_eq!(params[i][j].len(), grads[i][j].len());
for k in 0..params[i][j].len() {
assert_eq!(params[i][j][k].len(), grads[i][j][k].len());
for l in 0..params[i][j][k].len() {
// Update moving average of squared gradients
self.cache[cache_idx] = self.decay_rate * self.cache[cache_idx]
+ (1.0 - self.decay_rate) * grads[i][j][k][l].powi(2);
// Update parameters
updated_params[i][j][k][l] -= self.learning_rate * grads[i][j][k][l]
/ (self.cache[cache_idx] + self.epsilon).sqrt();
cache_idx += 1;
}
}
}
}
updated_params
}
}
#[cfg(test)]
mod tests {
use super::*;
/// Tests RMSprop initialization
#[test]
fn test_rmsprop_initialization() {
let optimizer = RMSprop::new(0.001, 0.9, 1e-8);
assert_eq!(optimizer.learning_rate, 0.001);
assert_eq!(optimizer.decay_rate, 0.9);
assert_eq!(optimizer.epsilon, 1e-8);
assert!(optimizer.cache.is_empty());
}
/// Tests invalid learning rate
#[test]
#[should_panic(expected = "Learning rate must be positive")]
fn test_invalid_learning_rate() {
RMSprop::new(-0.001, 0.9, 1e-8);
}
/// Tests invalid decay rate
#[test]
#[should_panic(expected = "Decay rate must be between 0 and 1")]
fn test_invalid_decay_rate() {
RMSprop::new(0.001, 1.5, 1e-8);
}
/// Tests parameter update
#[test]
fn test_parameter_update() {
let mut optimizer = RMSprop::new(0.1, 0.9, 1e-8);
let params = vec![1.0, 2.0, 3.0];
let grads = vec![0.1, 0.2, 0.3];
let updated = optimizer.update(¶ms, &grads);
assert_eq!(updated.len(), params.len());
for i in 0..params.len() {
assert!(updated[i] != params[i]); // Parameters should change
}
}
/// Tests 2D parameter update
#[test]
fn test_2d_parameter_update() {
let mut optimizer = RMSprop::new(0.1, 0.9, 1e-8);
let params = vec![vec![1.0, 2.0], vec![3.0, 4.0]];
let grads = vec![vec![0.1, 0.2], vec![0.3, 0.4]];
let updated = optimizer.update_2d(¶ms, &grads);
assert_eq!(updated.len(), params.len());
assert_eq!(updated[0].len(), params[0].len());
for i in 0..params.len() {
for j in 0..params[i].len() {
assert!(updated[i][j] != params[i][j]); // Parameters should change
}
}
}
/// Tests 4D parameter update
#[test]
fn test_4d_parameter_update() {
let mut optimizer = RMSprop::new(0.1, 0.9, 1e-8);
let params = vec![vec![vec![vec![1.0; 2]; 2]; 2]; 2];
let grads = vec![vec![vec![vec![0.1; 2]; 2]; 2]; 2];
let updated = optimizer.update_4d(¶ms, &grads);
assert_eq!(updated.len(), params.len());
assert_eq!(updated[0].len(), params[0].len());
assert_eq!(updated[0][0].len(), params[0][0].len());
assert_eq!(updated[0][0][0].len(), params[0][0][0].len());
// Check that parameters have been updated
assert!(updated[0][0][0][0] != params[0][0][0][0]);
}
}