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
//! 1-D convolution and max-pooling: [`Tensor::conv1d`] and
//! [`Tensor::maxpool1d`].
//!
//! Some LSTM OCR network configurations put a small convolution and/or a
//! max-pool at the input stage, ahead of the recurrent layers, to reduce the
//! along-width resolution and mix neighbouring columns before the LSTM sees
//! them. These are the standard signal-processing ops (valid, no-padding
//! cross-correlation; windowed max) — clean-room standard math, not a port of
//! Tesseract's image-stacking `Convolve`/reconfig `Maxpool`, which are shaped
//! around its `StrideMap`/`NetworkIO` plumbing rather than a plain 1-D layer.
use kopitiam_core::{DType, Error, Result, Shape};
use super::Tensor;
impl Tensor {
/// 1-D convolution (cross-correlation, the ML convention — the kernel is
/// *not* flipped), no padding ("valid"), with a configurable `stride`.
///
/// `self` is the input `[in_channels, length]`; `weight` is
/// `[out_channels, in_channels, kernel]`; `bias`, if given, is
/// `[out_channels]`. The result is `[out_channels, out_length]` where
/// `out_length = (length - kernel) / stride + 1`. Each output element is
///
/// ```text
/// out[oc][t] = bias[oc] + sum_ic sum_k input[ic][t*stride + k] * weight[oc][ic][k]
/// ```
///
/// # Errors
///
/// [`Error::DTypeMismatch`] if any operand is not `f32`.
/// [`Error::ShapeMismatch`] if `self` is not rank 2, `weight` is not
/// rank 3, the channel counts disagree, `stride` is 0, or `kernel` is
/// larger than `length` (which would make `out_length` empty or
/// ill-defined).
pub fn conv1d(&self, weight: &Tensor, bias: Option<&Tensor>, stride: usize) -> Result<Tensor> {
self.require_dtype(DType::F32)?;
weight.require_dtype(DType::F32)?;
if self.rank() != 2 || weight.rank() != 3 || stride == 0 {
return Err(Error::ShapeMismatch { expected: self.shape.clone(), actual: weight.shape.clone() });
}
let (in_channels, length) = (self.shape.dims()[0], self.shape.dims()[1]);
let wdims = weight.shape.dims();
let (out_channels, w_in_channels, kernel) = (wdims[0], wdims[1], wdims[2]);
if w_in_channels != in_channels || kernel == 0 || kernel > length {
return Err(Error::ShapeMismatch { expected: self.shape.clone(), actual: weight.shape.clone() });
}
let out_length = (length - kernel) / stride + 1;
// Materialize both operands in logical row-major order so the inner
// loops can index flat slices directly (correct for a transposed or
// narrowed input view too), the same approach matmul/norm take.
let input = self.to_vec_f32()?;
let w = weight.to_vec_f32()?;
let bias_vec = match bias {
Some(b) => {
b.require_dtype(DType::F32)?;
if b.elem_count() != out_channels {
return Err(Error::ShapeMismatch { expected: weight.shape.clone(), actual: b.shape.clone() });
}
b.to_vec_f32()?
}
None => vec![0.0f32; out_channels],
};
let mut out = vec![0f32; out_channels * out_length];
for oc in 0..out_channels {
for t in 0..out_length {
let start = t * stride;
let mut acc = bias_vec[oc];
for ic in 0..in_channels {
let in_row = &input[ic * length + start..ic * length + start + kernel];
let w_row = &w[(oc * in_channels + ic) * kernel..(oc * in_channels + ic) * kernel + kernel];
for (x, wv) in in_row.iter().zip(w_row) {
acc += x * wv;
}
}
out[oc * out_length + t] = acc;
}
}
Tensor::from_f32(out, Shape::new([out_channels, out_length]))
}
/// 1-D max-pooling over a sliding window, no padding, with a configurable
/// `stride`. Applied independently per channel.
///
/// `self` is `[channels, length]`; the result is `[channels, out_length]`
/// where `out_length = (length - kernel) / stride + 1` and
/// `out[c][t] = max(input[c][t*stride .. t*stride + kernel])`.
///
/// # Errors
///
/// [`Error::DTypeMismatch`] if `self` is not `f32`.
/// [`Error::ShapeMismatch`] if `self` is not rank 2, `stride` or `kernel`
/// is 0, or `kernel` exceeds `length`.
pub fn maxpool1d(&self, kernel: usize, stride: usize) -> Result<Tensor> {
self.require_dtype(DType::F32)?;
if self.rank() != 2 || stride == 0 || kernel == 0 {
return Err(Error::ShapeMismatch { expected: self.shape.clone(), actual: self.shape.clone() });
}
let (channels, length) = (self.shape.dims()[0], self.shape.dims()[1]);
if kernel > length {
return Err(Error::ShapeMismatch { expected: self.shape.clone(), actual: self.shape.clone() });
}
let out_length = (length - kernel) / stride + 1;
let input = self.to_vec_f32()?;
let mut out = vec![0f32; channels * out_length];
for c in 0..channels {
for t in 0..out_length {
let start = t * stride;
let window = &input[c * length + start..c * length + start + kernel];
out[c * out_length + t] = window.iter().copied().fold(f32::NEG_INFINITY, f32::max);
}
}
Tensor::from_f32(out, Shape::new([channels, out_length]))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn conv1d_single_channel_matches_hand_computation() {
// input [1,5] = [1,2,3,4,5], kernel [1,1,3] = [1,0,-1], stride 1.
// valid positions: out_len = (5-3)/1 + 1 = 3.
// out[0] = 1*1 + 2*0 + 3*(-1) = -2
// out[1] = 2*1 + 3*0 + 4*(-1) = -2
// out[2] = 3*1 + 4*0 + 5*(-1) = -2
let x = Tensor::from_f32(vec![1.0, 2.0, 3.0, 4.0, 5.0], [1, 5]).unwrap();
let w = Tensor::from_f32(vec![1.0, 0.0, -1.0], [1, 1, 3]).unwrap();
let out = x.conv1d(&w, None, 1).unwrap();
assert_eq!(out.shape().dims(), &[1, 3]);
assert_eq!(out.to_vec_f32().unwrap(), vec![-2.0, -2.0, -2.0]);
}
#[test]
fn conv1d_applies_bias_and_stride() {
// input [1,5] = [1,2,3,4,5], kernel [1,1,2] = [1,1], bias [10], stride 2.
// out_len = (5-2)/2 + 1 = 2. positions t=0 (idx 0..2), t=1 (idx 2..4).
// out[0] = (1+2) + 10 = 13 ; out[1] = (3+4) + 10 = 17
let x = Tensor::from_f32(vec![1.0, 2.0, 3.0, 4.0, 5.0], [1, 5]).unwrap();
let w = Tensor::from_f32(vec![1.0, 1.0], [1, 1, 2]).unwrap();
let bias = Tensor::from_f32(vec![10.0], [1]).unwrap();
let out = x.conv1d(&w, Some(&bias), 2).unwrap();
assert_eq!(out.shape().dims(), &[1, 2]);
assert_eq!(out.to_vec_f32().unwrap(), vec![13.0, 17.0]);
}
#[test]
fn conv1d_sums_over_input_channels_and_across_output_channels() {
// 2 in-channels, length 3; 2 out-channels; kernel 2. stride 1 -> out_len 2.
// input: ch0 = [1,2,3], ch1 = [4,5,6]
// oc0 weights: ic0=[1,0], ic1=[0,1]; oc1 weights: ic0=[1,1], ic1=[1,1]
// oc0,t0 = (1*1+2*0) + (4*0+5*1) = 1 + 5 = 6
// oc0,t1 = (2*1+3*0) + (5*0+6*1) = 2 + 6 = 8
// oc1,t0 = (1+2) + (4+5) = 12 ; oc1,t1 = (2+3) + (5+6) = 16
let x = Tensor::from_f32(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], [2, 3]).unwrap();
let w = Tensor::from_f32(
vec![
1.0, 0.0, 0.0, 1.0, // oc0: ic0=[1,0], ic1=[0,1]
1.0, 1.0, 1.0, 1.0, // oc1: ic0=[1,1], ic1=[1,1]
],
[2, 2, 2],
)
.unwrap();
let out = x.conv1d(&w, None, 1).unwrap();
assert_eq!(out.shape().dims(), &[2, 2]);
assert_eq!(out.to_vec_f32().unwrap(), vec![6.0, 8.0, 12.0, 16.0]);
}
#[test]
fn conv1d_rejects_a_channel_count_mismatch() {
let x = Tensor::from_f32(vec![1.0; 5], [1, 5]).unwrap();
let w = Tensor::from_f32(vec![1.0; 6], [1, 2, 3]).unwrap(); // wants 2 in-channels.
assert!(matches!(x.conv1d(&w, None, 1), Err(Error::ShapeMismatch { .. })));
}
#[test]
fn conv1d_rejects_a_kernel_wider_than_the_input() {
let x = Tensor::from_f32(vec![1.0, 2.0], [1, 2]).unwrap();
let w = Tensor::from_f32(vec![1.0; 3], [1, 1, 3]).unwrap();
assert!(matches!(x.conv1d(&w, None, 1), Err(Error::ShapeMismatch { .. })));
}
#[test]
fn conv1d_rejects_non_f32_input() {
let x = Tensor::from_i32(vec![1; 5], [1, 5]).unwrap();
let w = Tensor::from_f32(vec![1.0; 3], [1, 1, 3]).unwrap();
assert!(matches!(x.conv1d(&w, None, 1), Err(Error::DTypeMismatch { .. })));
}
#[test]
fn maxpool1d_matches_hand_computation() {
// input [1,6] = [1,3,2,5,4,0], kernel 2, stride 2.
// out_len = (6-2)/2 + 1 = 3. windows: [1,3]->3, [2,5]->5, [4,0]->4
let x = Tensor::from_f32(vec![1.0, 3.0, 2.0, 5.0, 4.0, 0.0], [1, 6]).unwrap();
let out = x.maxpool1d(2, 2).unwrap();
assert_eq!(out.shape().dims(), &[1, 3]);
assert_eq!(out.to_vec_f32().unwrap(), vec![3.0, 5.0, 4.0]);
}
#[test]
fn maxpool1d_overlapping_windows_per_channel() {
// 2 channels, length 4, kernel 2, stride 1 -> out_len 3 (overlapping).
// ch0 = [1,3,2,4]: [1,3]->3, [3,2]->3, [2,4]->4
// ch1 = [9,0,7,7]: [9,0]->9, [0,7]->7, [7,7]->7
let x = Tensor::from_f32(vec![1.0, 3.0, 2.0, 4.0, 9.0, 0.0, 7.0, 7.0], [2, 4]).unwrap();
let out = x.maxpool1d(2, 1).unwrap();
assert_eq!(out.shape().dims(), &[2, 3]);
assert_eq!(out.to_vec_f32().unwrap(), vec![3.0, 3.0, 4.0, 9.0, 7.0, 7.0]);
}
#[test]
fn maxpool1d_handles_negative_values() {
// All-negative window: the max must be the least-negative, not 0.
let x = Tensor::from_f32(vec![-3.0, -1.0, -5.0, -2.0], [1, 4]).unwrap();
let out = x.maxpool1d(2, 2).unwrap();
assert_eq!(out.to_vec_f32().unwrap(), vec![-1.0, -2.0]);
}
#[test]
fn maxpool1d_rejects_a_kernel_wider_than_the_input() {
let x = Tensor::from_f32(vec![1.0, 2.0], [1, 2]).unwrap();
assert!(matches!(x.maxpool1d(3, 1), Err(Error::ShapeMismatch { .. })));
}
#[test]
fn maxpool1d_rejects_non_f32_input() {
let x = Tensor::from_i32(vec![1; 4], [1, 4]).unwrap();
assert!(matches!(x.maxpool1d(2, 2), Err(Error::DTypeMismatch { .. })));
}
}