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
use super::super::ulp::assert_ulp_eq;
use super::*;
use proptest::prelude::*;
// ── Scalar known-answer tests ────────────────────────────────────────
/// Verify constant input produces output equal to beta during training
#[test]
fn test_batchnorm_constant_input_training() {
// All inputs constant for one channel -> output = beta (when gamma=1)
// N=4, C=1, all values = 5.0
let input = [5.0_f32, 5.0, 5.0, 5.0];
let gamma = [1.0_f32];
let beta = [3.0_f32];
let mut running_mean = [0.0_f32];
let mut running_var = [0.0_f32];
let mut output = [0.0_f32; 4];
batchnorm_scalar(
&input,
4,
1,
&gamma,
&beta,
1e-5,
&mut running_mean,
&mut running_var,
&mut output,
0.1,
true,
);
// With constant input: mean=5.0, var=0.0, (x-mean)=0
// output = gamma * 0 + beta = beta = 3.0
for (i, &o) in output.iter().enumerate() {
assert!((o - 3.0).abs() < 1e-3, "output[{i}] = {o}, expected ~3.0");
}
}
/// Verify training mode updates running mean and variance with momentum
#[test]
fn test_batchnorm_training_updates_running_stats() {
let input = [1.0_f32, 2.0, 3.0, 4.0]; // N=4, C=1
let gamma = [1.0_f32];
let beta = [0.0_f32];
let mut running_mean = [0.0_f32];
let mut running_var = [0.0_f32];
let mut output = [0.0_f32; 4];
batchnorm_scalar(
&input,
4,
1,
&gamma,
&beta,
1e-5,
&mut running_mean,
&mut running_var,
&mut output,
0.1,
true,
);
// batch_mean = 2.5, batch_var = 1.25
// running_mean = 0.9*0 + 0.1*2.5 = 0.25
// running_var = 0.9*0 + 0.1*1.25 = 0.125
assert!(
(running_mean[0] - 0.25).abs() < 1e-5,
"running_mean = {}, expected 0.25",
running_mean[0]
);
assert!(
(running_var[0] - 0.125).abs() < 1e-5,
"running_var = {}, expected 0.125",
running_var[0]
);
}
/// Verify inference mode uses running stats and leaves them unchanged
#[test]
fn test_batchnorm_inference_uses_running_stats() {
let input = [1.0_f32, 2.0, 3.0, 4.0]; // N=4, C=1
let gamma = [1.0_f32];
let beta = [0.0_f32];
let mut running_mean = [10.0_f32]; // Intentionally different from batch stats
let mut running_var = [4.0_f32];
let mut output = [0.0_f32; 4];
batchnorm_scalar(
&input,
4,
1,
&gamma,
&beta,
0.0,
&mut running_mean,
&mut running_var,
&mut output,
0.1,
false,
);
// Inference uses running stats: inv_std = 1/sqrt(4) = 0.5
// output[i] = (input[i] - 10) * 0.5
let inv_std = 1.0 / 4.0_f32.sqrt();
for (i, &x) in input.iter().enumerate() {
let expected = (x - 10.0) * inv_std;
assert!(
(output[i] - expected).abs() < 1e-5,
"output[{i}] = {}, expected {expected}",
output[i]
);
}
// Running stats should NOT change during inference
assert!((running_mean[0] - 10.0).abs() < 1e-10);
assert!((running_var[0] - 4.0).abs() < 1e-10);
}
/// Verify training and inference outputs differ when running stats diverge from batch stats
#[test]
fn test_batchnorm_inference_differs_from_training() {
let input = [1.0_f32, 2.0, 3.0, 4.0];
let gamma = [1.0_f32];
let beta = [0.0_f32];
// Training output
let mut rm_train = [5.0_f32];
let mut rv_train = [2.0_f32];
let mut train_out = [0.0_f32; 4];
batchnorm_scalar(
&input,
4,
1,
&gamma,
&beta,
1e-5,
&mut rm_train,
&mut rv_train,
&mut train_out,
0.1,
true,
);
// Inference output (with different running stats)
let mut rm_eval = [5.0_f32];
let mut rv_eval = [2.0_f32];
let mut eval_out = [0.0_f32; 4];
batchnorm_scalar(
&input,
4,
1,
&gamma,
&beta,
1e-5,
&mut rm_eval,
&mut rv_eval,
&mut eval_out,
0.1,
false,
);
// They should differ since batch stats != running stats
let mut all_equal = true;
for i in 0..4 {
if (train_out[i] - eval_out[i]).abs() > 1e-6 {
all_equal = false;
}
}
assert!(
!all_equal,
"training and inference outputs should differ when running stats != batch stats"
);
}
/// Verify batchnorm normalizes each channel independently with N=2, C=2
#[test]
fn test_batchnorm_multi_channel() {
// N=2, C=2
// Channel 0: [1.0, 3.0], mean=2.0, var=1.0
// Channel 1: [2.0, 4.0], mean=3.0, var=1.0
let input = [1.0_f32, 2.0, 3.0, 4.0]; // [sample0_ch0, sample0_ch1, sample1_ch0, sample1_ch1]
let gamma = [1.0_f32, 1.0];
let beta = [0.0_f32, 0.0];
let mut running_mean = [0.0_f32, 0.0];
let mut running_var = [0.0_f32, 0.0];
let mut output = [0.0_f32; 4];
batchnorm_scalar(
&input,
2,
2,
&gamma,
&beta,
1e-8,
&mut running_mean,
&mut running_var,
&mut output,
0.1,
true,
);
// Channel 0: mean=2, var=1, inv_std=1/sqrt(1+eps)~1
// (1-2)*1 = -1, (3-2)*1 = 1
assert!((output[0] - (-1.0)).abs() < 1e-3);
assert!((output[2] - 1.0).abs() < 1e-3);
// Channel 1: mean=3, var=1
// (2-3)*1 = -1, (4-3)*1 = 1
assert!((output[1] - (-1.0)).abs() < 1e-3);
assert!((output[3] - 1.0).abs() < 1e-3);
}
/// Verify batchnorm with batch size 1 produces output equal to beta
#[test]
fn test_batchnorm_single_sample() {
// N=1 (batch size 1) -> var=0, output=beta when gamma=1
let input = [7.0_f32, 3.0]; // N=1, C=2
let gamma = [1.0_f32, 1.0];
let beta = [5.0_f32, -5.0];
let mut running_mean = [0.0_f32, 0.0];
let mut running_var = [0.0_f32, 0.0];
let mut output = [0.0_f32; 2];
batchnorm_scalar(
&input,
1,
2,
&gamma,
&beta,
1e-5,
&mut running_mean,
&mut running_var,
&mut output,
0.1,
true,
);
// With N=1: mean=input, var=0, (x-mean)=0, output=beta
assert!((output[0] - 5.0).abs() < 1e-3, "output[0] = {}", output[0]);
assert!(
(output[1] - (-5.0)).abs() < 1e-3,
"output[1] = {}",
output[1]
);
}
/// Verify batchnorm panics on input length mismatch
#[test]
#[should_panic(expected = "input length must be n * c")]
fn test_batchnorm_input_length_mismatch() {
let input = [1.0_f32, 2.0];
let gamma = [1.0_f32];
let beta = [0.0_f32];
let mut rm = [0.0_f32];
let mut rv = [0.0_f32];
let mut output = [0.0_f32; 2];
batchnorm_scalar(
&input,
3,
1,
&gamma,
&beta,
1e-5,
&mut rm,
&mut rv,
&mut output,
0.1,
true,
);
}
/// Verify batchnorm panics when batch size or channels is zero
#[test]
#[should_panic(expected = "batchnorm requires n > 0 and c > 0")]
fn test_batchnorm_zero_batch() {
let input: [f32; 0] = [];
let gamma: [f32; 0] = [];
let beta: [f32; 0] = [];
let mut rm: [f32; 0] = [];
let mut rv: [f32; 0] = [];
let mut output: [f32; 0] = [];
batchnorm_scalar(
&input,
0,
0,
&gamma,
&beta,
1e-5,
&mut rm,
&mut rv,
&mut output,
0.1,
true,
);
}
// ── Property-based tests ─────────────────────────────────────────────
proptest! {
#[test]
fn prop_batchnorm_training_finite(
n in 2_usize..8,
c in 1_usize..4,
) {
let total = n * c;
let input: Vec<f32> = (0..total).map(|i| (i as f32) * 0.1 - 1.0).collect();
let gamma = vec![1.0_f32; c];
let beta = vec![0.0_f32; c];
let mut running_mean = vec![0.0_f32; c];
let mut running_var = vec![1.0_f32; c];
let mut output = vec![0.0_f32; total];
batchnorm_scalar(
&input, n, c, &gamma, &beta, 1e-5,
&mut running_mean, &mut running_var,
&mut output, 0.1, true,
);
for (i, &o) in output.iter().enumerate() {
prop_assert!(o.is_finite(), "output[{i}] = {o} is not finite");
}
for (i, &rv) in running_var.iter().enumerate() {
prop_assert!(rv >= 0.0, "running_var[{i}] = {rv} is negative");
}
}
#[test]
fn prop_batchnorm_running_var_nonneg(
n in 2_usize..8,
c in 1_usize..4,
iters in 1_usize..20,
) {
let total = n * c;
let mut running_mean = vec![0.0_f32; c];
let mut running_var = vec![1.0_f32; c];
for step in 0..iters {
let input: Vec<f32> = (0..total)
.map(|i| ((i + step) as f32) * 0.3 - 2.0)
.collect();
let gamma = vec![1.0_f32; c];
let beta = vec![0.0_f32; c];
let mut output = vec![0.0_f32; total];
batchnorm_scalar(
&input, n, c, &gamma, &beta, 1e-5,
&mut running_mean, &mut running_var,
&mut output, 0.1, true,
);
}
for (i, &rv) in running_var.iter().enumerate() {
prop_assert!(
rv >= 0.0,
"running_var[{i}] = {rv} after {iters} iterations"
);
}
}
}