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
//! Initialization strategies for new tensors.
use crate::error::{Error, Result};
use numr::dtype::DType;
use numr::runtime::Runtime;
use numr::tensor::Tensor;
/// Initialization strategy for new tensors.
#[derive(Debug, Clone, Copy)]
pub enum Init {
/// All zeros
Zeros,
/// All ones
Ones,
/// Constant value
Const(f32),
/// Uniform random in `[-bound, bound]`
Uniform(f32),
/// Kaiming uniform (PyTorch Linear default): U(-1/sqrt(in), 1/sqrt(in))
PyTorchLinear,
/// PyTorch Embedding default: N(0, 1)
PyTorchEmbedding,
/// Kaiming (He) normal: N(0, sqrt(2 / fan_in))
///
/// Standard initialization for ReLU networks. fan_in is the product of
/// all dimensions except the last (output) dimension.
Kaiming,
/// Xavier (Glorot) normal: N(0, sqrt(2 / (fan_in + fan_out)))
///
/// Standard initialization for Sigmoid/Tanh networks. Used in some
/// attention weight initializations.
Xavier,
/// Normal distribution with given mean and standard deviation.
Randn { mean: f64, stdev: f64 },
/// Truncated normal: N(mean, stdev) clamped to [mean - 2*stdev, mean + 2*stdev]
///
/// Used by GPT-2, BERT, and most modern LLMs for training stability.
TruncatedNormal { mean: f64, stdev: f64 },
}
impl Init {
/// Create a tensor initialized according to this strategy.
///
/// # Arguments
/// * `shape` - Shape of the tensor to create
/// * `dtype` - Data type
/// * `device` - Device to create on
/// * `client` - Runtime client (needed for random ops)
pub fn init_tensor<R, C>(
&self,
shape: &[usize],
dtype: DType,
device: &R::Device,
client: &C,
) -> Result<Tensor<R>>
where
R: Runtime<DType = DType>,
C: numr::runtime::RuntimeClient<R>
+ numr::ops::RandomOps<R>
+ numr::ops::ScalarOps<R>
+ numr::ops::BinaryOps<R>
+ numr::ops::CompareOps<R>
+ numr::ops::TensorOps<R>,
{
// Trait bounds on the function provide the methods
match *self {
Init::Zeros => Ok(Tensor::zeros(shape, dtype, device)),
Init::Ones => Ok(Tensor::ones(shape, dtype, device)),
Init::Const(val) => {
let t = Tensor::zeros(shape, dtype, device);
client.add_scalar(&t, val as f64).map_err(Error::Numr)
}
Init::Uniform(bound) => {
// U(-bound, bound) = rand() * 2*bound - bound
let r = client.rand(shape, dtype).map_err(Error::Numr)?;
let scaled = client
.mul_scalar(&r, 2.0 * bound as f64)
.map_err(Error::Numr)?;
client
.add_scalar(&scaled, -(bound as f64))
.map_err(Error::Numr)
}
Init::PyTorchLinear => {
// Kaiming uniform: U(-1/sqrt(fan_in), 1/sqrt(fan_in))
let fan_in = shape[0];
let bound = 1.0 / (fan_in as f64).sqrt();
let r = client.rand(shape, dtype).map_err(Error::Numr)?;
let scaled = client.mul_scalar(&r, 2.0 * bound).map_err(Error::Numr)?;
client.add_scalar(&scaled, -bound).map_err(Error::Numr)
}
Init::PyTorchEmbedding => {
// PyTorch `nn.Embedding` default initializes weights with N(0, 1).
client.randn(shape, dtype).map_err(Error::Numr)
}
Init::Kaiming => {
// Kaiming/He normal: N(0, sqrt(2 / fan_in))
let fan_in = if shape.len() >= 2 {
shape[..shape.len() - 1].iter().product::<usize>()
} else {
shape[0]
};
let std = (2.0 / fan_in as f64).sqrt();
let r = client.randn(shape, dtype).map_err(Error::Numr)?;
client.mul_scalar(&r, std).map_err(Error::Numr)
}
Init::Xavier => {
// Xavier/Glorot normal: N(0, sqrt(2 / (fan_in + fan_out)))
let (fan_in, fan_out) = if shape.len() >= 2 {
let fi = shape[..shape.len() - 1].iter().product::<usize>();
let fo = shape[shape.len() - 1];
(fi, fo)
} else {
(shape[0], shape[0])
};
let std = (2.0 / (fan_in + fan_out) as f64).sqrt();
let r = client.randn(shape, dtype).map_err(Error::Numr)?;
client.mul_scalar(&r, std).map_err(Error::Numr)
}
Init::Randn { mean, stdev } => {
let r = client.randn(shape, dtype).map_err(Error::Numr)?;
let scaled = client.mul_scalar(&r, stdev).map_err(Error::Numr)?;
if mean != 0.0 {
client.add_scalar(&scaled, mean).map_err(Error::Numr)
} else {
Ok(scaled)
}
}
Init::TruncatedNormal { mean, stdev } => {
// Generate N(0, 1), clamp to [-2, 2], then scale by stdev and shift by mean
let r = client.randn(shape, dtype).map_err(Error::Numr)?;
let clamped = client.clamp(&r, -2.0, 2.0).map_err(Error::Numr)?;
let scaled = client.mul_scalar(&clamped, stdev).map_err(Error::Numr)?;
if mean != 0.0 {
client.add_scalar(&scaled, mean).map_err(Error::Numr)
} else {
Ok(scaled)
}
}
}
}
}
#[cfg(test)]
mod tests {
use super::*;
use numr::runtime::cpu::{CpuDevice, CpuRuntime};
fn device() -> CpuDevice {
CpuDevice::new()
}
fn client() -> numr::runtime::cpu::CpuClient {
let d = device();
CpuRuntime::default_client(&d)
}
#[test]
fn test_init_zeros() {
let d = device();
let c = client();
let t = Init::Zeros
.init_tensor(&[2, 3], DType::F32, &d, &c)
.unwrap();
assert_eq!(t.shape(), &[2, 3]);
let data: Vec<f32> = t.to_vec();
assert!(data.iter().all(|&v| v == 0.0));
}
#[test]
fn test_init_kaiming() {
let d = device();
let c = client();
// [out=64, in=128] → fan_in=128, std=sqrt(2/128)≈0.125
let t = Init::Kaiming
.init_tensor(&[64, 128], DType::F32, &d, &c)
.unwrap();
assert_eq!(t.shape(), &[64, 128]);
let data: Vec<f32> = t.to_vec();
let mean: f32 = data.iter().sum::<f32>() / data.len() as f32;
// Mean should be close to 0
assert!(mean.abs() < 0.1, "Kaiming mean too large: {mean}");
// Std should be close to sqrt(2/128) ≈ 0.125
let var: f32 = data.iter().map(|x| (x - mean).powi(2)).sum::<f32>() / data.len() as f32;
let std = var.sqrt();
// fan_in = product of all dims except last = 64
let expected_std = (2.0f32 / 64.0).sqrt();
assert!(
(std - expected_std).abs() < 0.05,
"Kaiming std {std} vs expected {expected_std}"
);
}
#[test]
fn test_init_xavier() {
let d = device();
let c = client();
// [256, 512] → fan_in=256, fan_out=512, std=sqrt(2/768)≈0.051
let t = Init::Xavier
.init_tensor(&[256, 512], DType::F32, &d, &c)
.unwrap();
assert_eq!(t.shape(), &[256, 512]);
let data: Vec<f32> = t.to_vec();
let mean: f32 = data.iter().sum::<f32>() / data.len() as f32;
assert!(mean.abs() < 0.05, "Xavier mean too large: {mean}");
}
#[test]
fn test_init_randn() {
let d = device();
let c = client();
let t = Init::Randn {
mean: 5.0,
stdev: 0.1,
}
.init_tensor(&[1000], DType::F32, &d, &c)
.unwrap();
let data: Vec<f32> = t.to_vec();
let mean: f32 = data.iter().sum::<f32>() / data.len() as f32;
assert!((mean - 5.0).abs() < 0.1, "Randn mean {mean} should be ~5.0");
}
#[test]
fn test_init_truncated_normal() {
let d = device();
let c = client();
let t = Init::TruncatedNormal {
mean: 0.0,
stdev: 0.02,
}
.init_tensor(&[10000], DType::F32, &d, &c)
.unwrap();
let data: Vec<f32> = t.to_vec();
// All values should be within [-0.04, 0.04] (2*stdev)
for &v in &data {
assert!(
(-0.04..=0.04).contains(&v),
"Truncated normal value {v} out of range [-0.04, 0.04]"
);
}
}
}