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
// SPDX-License-Identifier: MIT OR Apache-2.0
//! Max pooling layer.
//!
//! Downsamples spatial dimensions by taking the maximum value in each
//! pooling window. No trainable parameters.
use crate::neural::traits::{BackwardOutput, Layer};
/// 2D max pooling layer.
///
/// Input shape: `[batch, C, H, W]` (channels-first, flattened).
/// Output shape: `[batch, C, H_out, W_out]` where
/// `H_out = (H - pool_size) / stride + 1`.
pub struct MaxPool2D {
/// Pooling window size.
pub(crate) pool_size: usize,
/// Stride (defaults to pool_size).
pub(crate) stride: usize,
/// Number of channels (set on first forward).
pub(crate) channels: usize,
/// Input spatial height.
pub(crate) h_in: usize,
/// Input spatial width.
pub(crate) w_in: usize,
/// Output spatial height.
pub(crate) h_out: usize,
/// Output spatial width.
pub(crate) w_out: usize,
/// Index of the max element in each pooling window (for backward).
pub(crate) cache_max_indices: Vec<usize>,
pub(crate) cache_batch: usize,
}
impl MaxPool2D {
/// Create a new MaxPool2D layer.
///
/// `pool_size` is the spatial extent of the pooling window.
/// `stride` defaults to `pool_size` if `None`.
pub fn new(pool_size: usize, stride: Option<usize>) -> Self {
let stride = stride.unwrap_or(pool_size);
Self {
pool_size,
stride,
channels: 0,
h_in: 0,
w_in: 0,
h_out: 0,
w_out: 0,
cache_max_indices: Vec::new(),
cache_batch: 0,
}
}
}
impl Layer for MaxPool2D {
fn forward(&mut self, input: &[f64], batch: usize, training: bool) -> Vec<f64> {
// Infer dimensions from input.
let per_sample = input.len() / batch;
// We need to know the number of channels. Assume it was set via the
// previous layer's output. If channels is 0, try to infer: assume
// square spatial dims.
if self.channels == 0 {
// Can't infer channels without additional info.
// Default: assume channels = 1 if spatial is square.
// Otherwise the caller must set channels beforehand.
let side_sq = (per_sample as f64).sqrt() as usize;
if side_sq * side_sq == per_sample {
self.channels = 1;
self.h_in = side_sq;
self.w_in = side_sq;
} else {
// Try to find c such that per_sample / c is a perfect square.
for c in (1..=per_sample).rev() {
if per_sample % c == 0 {
let spatial = per_sample / c;
let side = (spatial as f64).sqrt() as usize;
if side * side == spatial {
self.channels = c;
self.h_in = side;
self.w_in = side;
break;
}
}
}
}
} else {
let spatial = per_sample / self.channels;
let side = (spatial as f64).sqrt() as usize;
debug_assert_eq!(side * side, spatial);
self.h_in = side;
self.w_in = side;
}
self.h_out = (self.h_in - self.pool_size) / self.stride + 1;
self.w_out = (self.w_in - self.pool_size) / self.stride + 1;
let out_per_sample = self.channels * self.h_out * self.w_out;
let mut output = Vec::with_capacity(batch * out_per_sample);
let mut max_indices = if training {
Vec::with_capacity(batch * out_per_sample)
} else {
Vec::new()
};
for b in 0..batch {
let sample = &input[b * per_sample..];
for c in 0..self.channels {
let ch_offset = c * self.h_in * self.w_in;
for oy in 0..self.h_out {
for ox in 0..self.w_out {
let mut max_val = f64::NEG_INFINITY;
let mut max_idx = 0;
for py in 0..self.pool_size {
for px in 0..self.pool_size {
let iy = oy * self.stride + py;
let ix = ox * self.stride + px;
let idx = ch_offset + iy * self.w_in + ix;
let val = sample[idx];
if val > max_val {
max_val = val;
max_idx = idx;
}
}
}
output.push(max_val);
if training {
max_indices.push(b * per_sample + max_idx);
}
}
}
}
}
if training {
self.cache_max_indices = max_indices;
self.cache_batch = batch;
}
output
}
fn backward(&self, grad_output: &[f64]) -> BackwardOutput {
let batch = self.cache_batch;
let in_per_sample = self.channels * self.h_in * self.w_in;
let mut grad_input = vec![0.0; batch * in_per_sample];
for (i, &max_idx) in self.cache_max_indices.iter().enumerate() {
grad_input[max_idx] += grad_output[i];
}
(grad_input, vec![]) // no parameters
}
fn n_param_groups(&self) -> usize {
0
}
fn params_mut(&mut self) -> Vec<(&mut Vec<f64>, &mut Vec<f64>)> {
vec![]
}
fn save_params(&self) -> Vec<(Vec<f64>, Vec<f64>)> {
vec![]
}
fn restore_params(&mut self, _saved: &[(Vec<f64>, Vec<f64>)]) {}
fn in_size(&self) -> usize {
self.channels * self.h_in * self.w_in
}
fn out_size(&self) -> usize {
self.channels * self.h_out * self.w_out
}
fn name(&self) -> &'static str {
"MaxPool2D"
}
}
#[cfg(test)]
mod tests {
use super::*;
#[allow(clippy::float_cmp)]
#[test]
fn maxpool_basic() {
let mut pool = MaxPool2D::new(2, None);
pool.channels = 1;
// 1 channel, 4×4 input:
// 1 2 3 4
// 5 6 7 8
// 9 10 11 12
// 13 14 15 16
let input: Vec<f64> = (1..=16).map(|x| x as f64).collect();
let output = pool.forward(&input, 1, false);
// 2×2 pool, stride 2 → 2×2 output
assert_eq!(output.len(), 4);
assert_eq!(output[0], 6.0); // max(1,2,5,6)
assert_eq!(output[1], 8.0); // max(3,4,7,8)
assert_eq!(output[2], 14.0); // max(9,10,13,14)
assert_eq!(output[3], 16.0); // max(11,12,15,16)
}
#[allow(clippy::float_cmp)]
#[test]
fn maxpool_backward() {
let mut pool = MaxPool2D::new(2, None);
pool.channels = 1;
let input: Vec<f64> = (1..=16).map(|x| x as f64).collect();
pool.forward(&input, 1, true);
let grad_out = vec![1.0, 2.0, 3.0, 4.0];
let (grad_input, param_grads) = pool.backward(&grad_out);
assert_eq!(grad_input.len(), 16);
assert!(param_grads.is_empty());
// Gradient should be routed to the max positions.
assert_eq!(grad_input[5], 1.0); // position of 6 (max of patch 0)
assert_eq!(grad_input[7], 2.0); // position of 8 (max of patch 1)
assert_eq!(grad_input[13], 3.0); // position of 14 (max of patch 2)
assert_eq!(grad_input[15], 4.0); // position of 16 (max of patch 3)
// Non-max positions should be zero.
assert_eq!(grad_input[0], 0.0);
assert_eq!(grad_input[4], 0.0);
}
#[test]
fn maxpool_multi_channel() {
let mut pool = MaxPool2D::new(2, None);
pool.channels = 2;
// 2 channels, 4×4 each → total 32 elements.
let mut input = Vec::with_capacity(32);
for c in 0..2 {
for i in 0..16 {
input.push((c * 16 + i + 1) as f64);
}
}
let output = pool.forward(&input, 1, false);
// Each channel produces 2×2 output → 2 * 4 = 8
assert_eq!(output.len(), 8);
}
#[test]
fn maxpool_batched() {
let mut pool = MaxPool2D::new(2, None);
pool.channels = 1;
let input: Vec<f64> = (0..32).map(|x| x as f64).collect(); // 2 samples, 4×4
let output = pool.forward(&input, 2, false);
assert_eq!(output.len(), 8); // 2 * 1 * 2 * 2
}
}