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
// SPDX-FileCopyrightText: Copyright 2025 Au-Zone Technologies
// SPDX-License-Identifier: Apache-2.0
use super::config::ConfigOutputRef;
use super::configs::{self, DecoderType, DimName};
use super::{ArrayViewDQuantized, Decoder};
use crate::DecoderError;
use ndarray::{ArrayView, ArrayViewD, Dimension};
impl Decoder {
pub(super) fn match_outputs_to_detect<'a, 'b, T>(
configs: &[configs::Detection],
outputs: &'a [ArrayViewD<'b, T>],
) -> Result<Vec<&'a ArrayViewD<'b, T>>, DecoderError> {
let mut new_output_order = Vec::new();
for c in configs {
let mut found = false;
for o in outputs {
if o.shape() == c.shape {
new_output_order.push(o);
found = true;
break;
}
}
if !found {
return Err(DecoderError::InvalidShape(format!(
"Did not find output with shape {:?}",
c.shape
)));
}
}
Ok(new_output_order)
}
pub(super) fn find_outputs_with_shape<'a, 'b, T>(
shape: &[usize],
outputs: &'a [ArrayViewD<'b, T>],
skip: &[usize],
) -> Result<(&'a ArrayViewD<'b, T>, usize), DecoderError> {
for (ind, o) in outputs.iter().enumerate() {
if skip.contains(&ind) {
continue;
}
if o.shape() == shape {
return Ok((o, ind));
}
}
Err(DecoderError::InvalidShape(format!(
"Did not find output with shape {:?}",
shape
)))
}
pub(super) fn find_outputs_with_shape_quantized<'a, 'b>(
shape: &[usize],
outputs: &'a [ArrayViewDQuantized<'b>],
skip: &[usize],
) -> Result<(&'a ArrayViewDQuantized<'b>, usize), DecoderError> {
for (ind, o) in outputs.iter().enumerate() {
if skip.contains(&ind) {
continue;
}
if o.shape() == shape {
return Ok((o, ind));
}
}
Err(DecoderError::InvalidShape(format!(
"Did not find output with shape {:?}",
shape
)))
}
/// This is split detection, need to swap axes to batch, height, width,
/// num_anchors_x_features,
pub(super) fn modelpack_det_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::NumBoxes => 1,
DimName::Padding => 2,
DimName::BoxCoords => 3,
_ => 1000, // this should be unreachable
}
}
// This is Ultralytics detection, need to swap axes to batch, num_features,
// height, width
pub(super) fn yolo_det_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::NumFeatures => 1,
DimName::NumBoxes => 2,
_ => 1000, // this should be unreachable
}
}
// This is modelpack boxes, need to swap axes to batch, num_boxes, padding,
// box_coords
pub(super) fn modelpack_boxes_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::NumBoxes => 1,
DimName::Padding => 2,
DimName::BoxCoords => 3,
_ => 1000, // this should be unreachable
}
}
/// This is Ultralytics boxes, need to swap axes to batch, box_coords,
/// num_boxes
pub(super) fn yolo_boxes_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::BoxCoords => 1,
DimName::NumBoxes => 2,
_ => 1000, // this should be unreachable
}
}
/// This is modelpack scores, need to swap axes to batch, num_boxes,
/// num_classes
pub(super) fn modelpack_scores_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::NumBoxes => 1,
DimName::NumClasses => 2,
_ => 1000, // this should be unreachable
}
}
pub(super) fn yolo_scores_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::NumClasses => 1,
DimName::NumBoxes => 2,
_ => 1000, // this should be unreachable
}
}
/// This is modelpack segmentation, need to swap axes to batch, height,
/// width, num_classes
pub(super) fn modelpack_segmentation_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::Height => 1,
DimName::Width => 2,
DimName::NumClasses => 3,
_ => 1000, // this should be unreachable
}
}
/// This is modelpack masks, need to swap axes to batch, height,
/// width
pub(super) fn modelpack_mask_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::Height => 1,
DimName::Width => 2,
_ => 1000, // this should be unreachable
}
}
/// This is yolo protos, need to swap axes to batch, height, width,
/// num_protos
pub(super) fn yolo_protos_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::Height => 1,
DimName::Width => 2,
DimName::NumProtos => 3,
_ => 1000, // this should be unreachable
}
}
/// This is yolo mask coefficients, need to swap axes to batch, num_protos,
/// num_boxes
pub(super) fn yolo_maskcoefficients_order(x: DimName) -> usize {
match x {
DimName::Batch => 0,
DimName::NumProtos => 1,
DimName::NumBoxes => 2,
_ => 1000, // this should be unreachable
}
}
pub(super) fn get_order_fn(config: ConfigOutputRef) -> fn(DimName) -> usize {
let decoder_type = config.decoder();
match (config, decoder_type) {
(ConfigOutputRef::Detection(_), DecoderType::ModelPack) => Self::modelpack_det_order,
(ConfigOutputRef::Detection(_), DecoderType::Ultralytics) => Self::yolo_det_order,
(ConfigOutputRef::Boxes(_), DecoderType::ModelPack) => Self::modelpack_boxes_order,
(ConfigOutputRef::Boxes(_), DecoderType::Ultralytics) => Self::yolo_boxes_order,
(ConfigOutputRef::Scores(_), DecoderType::ModelPack) => Self::modelpack_scores_order,
(ConfigOutputRef::Scores(_), DecoderType::Ultralytics) => Self::yolo_scores_order,
(ConfigOutputRef::Segmentation(_), _) => Self::modelpack_segmentation_order,
(ConfigOutputRef::Mask(_), _) => Self::modelpack_mask_order,
(ConfigOutputRef::Protos(_), _) => Self::yolo_protos_order,
(ConfigOutputRef::MaskCoefficients(_), _) => Self::yolo_maskcoefficients_order,
(ConfigOutputRef::Classes(_), _) => Self::yolo_scores_order,
}
}
/// Ensure a 3D protos tensor is in HWC order. When dshape is set,
/// `swap_axes_if_needed` already reorders the 4D tensor to NHWC before
/// the batch-dim slice, so the resulting 3D view is already HWC.
///
/// When dshape is empty (no named dimensions), we detect the layout by
/// comparing axis sizes: the channel/proto count (e.g. 32) is always
/// smaller than the spatial dimensions (e.g. 160×160). If axis 0 is
/// the smallest, the tensor is CHW and needs permutation to HWC;
/// otherwise it is already HWC.
///
/// **Known limitations**: the heuristic fails when the channel count is
/// not strictly the smallest dimension (e.g. protos shape `(32, 1, 1)`
/// or `(160, 5, 5)`). Set `dshape` in the config for reliable axis
/// ordering in these edge cases.
pub(super) fn protos_to_hwc<'a, T>(
protos: ArrayView<'a, T, ndarray::Ix3>,
config: &configs::Protos,
) -> ArrayView<'a, T, ndarray::Ix3> {
if config.dshape.is_empty() {
let (d0, d1, d2) = protos.dim();
log::debug!(
"protos_to_hwc: no dshape configured, using size heuristic on \
shape ({d0}, {d1}, {d2}); set dshape in config for reliable ordering"
);
if d0 < d1 && d0 < d2 {
// CHW (from NCHW) → permute to HWC
protos.permuted_axes([1, 2, 0])
} else {
// Already HWC (from NHWC) or ambiguous — keep as-is
protos
}
} else {
protos
}
}
pub(super) fn swap_axes_if_needed<'a, T, D: Dimension>(
array: &ArrayView<'a, T, D>,
config: ConfigOutputRef,
) -> ArrayView<'a, T, D> {
let mut array = array.clone();
if config.dshape().is_empty() {
return array;
}
let order_fn: fn(DimName) -> usize = Self::get_order_fn(config.clone());
let mut current_order: Vec<usize> = config
.dshape()
.iter()
.map(|x| order_fn(x.0))
.collect::<Vec<_>>();
assert_eq!(array.shape().len(), current_order.len());
// do simple bubble sort as swap_axes is inexpensive and the
// number of dimensions is small
for i in 0..current_order.len() {
let mut swapped = false;
for j in 0..current_order.len() - 1 - i {
if current_order[j] > current_order[j + 1] {
array.swap_axes(j, j + 1);
current_order.swap(j, j + 1);
swapped = true;
}
}
if !swapped {
break;
}
}
array
}
pub(super) fn match_outputs_to_detect_quantized<'a, 'b>(
configs: &[configs::Detection],
outputs: &'a [ArrayViewDQuantized<'b>],
) -> Result<Vec<&'a ArrayViewDQuantized<'b>>, DecoderError> {
let mut new_output_order = Vec::new();
for c in configs {
let mut found = false;
for o in outputs {
if o.shape() == c.shape {
new_output_order.push(o);
found = true;
break;
}
}
if !found {
return Err(DecoderError::InvalidShape(format!(
"Did not find output with shape {:?}",
c.shape
)));
}
}
Ok(new_output_order)
}
}