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
//! Weight initialization strategies for neural networks
use crate::error::{NeuralError, Result};
use scirs2_core::ndarray::{Array, Dimension, IxDyn};
use scirs2_core::numeric::{Float, NumAssign};
use scirs2_core::random::Rng;
use std::fmt::Debug;
/// Initialization strategies for neural network weights
#[derive(Debug, Clone, Copy)]
pub enum Initializer {
/// Zero initialization
Zeros,
/// One initialization
Ones,
/// Uniform random initialization
Uniform {
/// Minimum value
min: f64,
/// Maximum value
max: f64,
},
/// Normal random initialization
Normal {
/// Mean
mean: f64,
/// Standard deviation
std: f64,
},
/// Xavier/Glorot initialization
Xavier,
/// He initialization
He,
/// LeCun initialization
LeCun,
}
impl Initializer {
/// Initialize weights according to the strategy
///
/// # Arguments
/// * `shape` - Shape of the weights array
/// * `fan_in` - Number of input connections (for Xavier, He, LeCun)
/// * `fan_out` - Number of output connections (for Xavier)
/// * `rng` - Random number generator
/// # Returns
/// * Initialized weights array
pub fn initialize<F: Float + Debug, R: Rng>(
&self,
shape: IxDyn,
fan_in: usize,
fan_out: usize,
rng: &mut R,
) -> Result<Array<F, IxDyn>> {
let size = shape.as_array_view().iter().product();
match self {
Initializer::Zeros => Ok(Array::zeros(shape)),
Initializer::Ones => {
let ones: Vec<F> = (0..size).map(|_| F::one()).collect();
Array::from_shape_vec(shape, ones).map_err(|e| {
NeuralError::InvalidArchitecture(format!("Failed to create array: {e}"))
})
}
Initializer::Uniform { min, max } => {
let values: Vec<F> = (0..size)
.map(|_| {
let val = rng.random_range(*min..*max);
F::from(val).ok_or_else(|| {
NeuralError::InvalidArchitecture(
"Failed to convert random value".to_string(),
)
})
})
.collect::<Result<Vec<F>>>()?;
Array::from_shape_vec(shape, values).map_err(|e| {
NeuralError::InvalidArchitecture(format!("Failed to create array: {e}"))
})
}
Initializer::Normal { mean, std } => {
let values: Vec<F> = (0..(size / 2 + 1))
.flat_map(|_| {
// Box-Muller transform to generate normal distribution
let u1 = rng.random_range(0.0..1.0);
let u2 = rng.random_range(0.0..1.0);
let z0 = (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).cos();
let z1 = (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).sin();
let val0 = mean + std * z0;
let val1 = mean + std * z1;
vec![
F::from(val0).unwrap_or(F::zero()),
F::from(val1).unwrap_or(F::zero()),
]
})
.take(size)
.collect();
Array::from_shape_vec(shape, values).map_err(|e| {
NeuralError::InvalidArchitecture(format!("Failed to create array: {e}"))
})
}
Initializer::Xavier => {
let limit = (6.0 / (fan_in + fan_out) as f64).sqrt();
let values: Vec<F> = (0..size)
.map(|_| {
let val = rng.random_range(-limit..limit);
F::from(val).unwrap_or(F::zero())
})
.collect();
Array::from_shape_vec(shape, values).map_err(|e| {
NeuralError::InvalidArchitecture(format!("Failed to create array: {e}"))
})
}
Initializer::He => {
let std = (2.0 / fan_in as f64).sqrt();
let values: Vec<F> = (0..(size / 2 + 1))
.flat_map(|_| {
// Box-Muller transform for He initialization
let u1 = rng.random_range(0.0..1.0);
let u2 = rng.random_range(0.0..1.0);
let z0 = (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).cos();
let z1 = (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).sin();
let val0 = std * z0;
let val1 = std * z1;
vec![
F::from(val0).unwrap_or(F::zero()),
F::from(val1).unwrap_or(F::zero()),
]
})
.take(size)
.collect();
Array::from_shape_vec(shape, values).map_err(|e| {
NeuralError::InvalidArchitecture(format!("Failed to create array: {e}"))
})
}
Initializer::LeCun => {
let std = (1.0 / fan_in as f64).sqrt();
let values: Vec<F> = (0..(size / 2 + 1))
.flat_map(|_| {
// Box-Muller transform for LeCun initialization
let u1 = rng.random_range(0.0..1.0);
let u2 = rng.random_range(0.0..1.0);
let z0 = (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).cos();
let z1 = (-2.0 * u1.ln()).sqrt() * (2.0 * std::f64::consts::PI * u2).sin();
let val0 = std * z0;
let val1 = std * z1;
vec![
F::from(val0).unwrap_or(F::zero()),
F::from(val1).unwrap_or(F::zero()),
]
})
.take(size)
.collect();
Array::from_shape_vec(shape, values).map_err(|e| {
NeuralError::InvalidArchitecture(format!("Failed to create array: {e}"))
})
}
}
}
}
/// Xavier/Glorot uniform initialization
///
/// # Arguments
/// * `shape` - Shape of the weights array
/// # Returns
/// * Initialized weights array
#[allow(dead_code)]
pub fn xavier_uniform<F: Float + Debug + NumAssign>(shape: IxDyn) -> Result<Array<F, IxDyn>> {
let fan_in = match shape.ndim() {
0 => 1,
1 => shape[0],
_ => shape[0],
};
let fan_out = match shape.ndim() {
1 => 1,
_ => shape[1],
};
let mut rng = scirs2_core::random::rng();
Initializer::Xavier.initialize(shape, fan_in, fan_out, &mut rng)
}