Spaces:
No application file
No application file
Delete layout_alignment.py
Browse files- layout_alignment.py +0 -213
layout_alignment.py
DELETED
@@ -1,213 +0,0 @@
|
|
1 |
-
from typing import Dict, List, Tuple, Union
|
2 |
-
|
3 |
-
import datasets as ds
|
4 |
-
import evaluate
|
5 |
-
import numpy as np
|
6 |
-
import numpy.typing as npt
|
7 |
-
|
8 |
-
_DESCRIPTION = """\
|
9 |
-
Computes some alignment metrics that are different to each other in previous works.
|
10 |
-
"""
|
11 |
-
|
12 |
-
_KWARGS_DESCRIPTION = """\
|
13 |
-
Args:
|
14 |
-
bbox (`list` of `lists` of `int`): A list of lists of integers representing bounding boxes.
|
15 |
-
mask (`list` of `lists` of `bool`): A list of lists of booleans representing masks.
|
16 |
-
|
17 |
-
Returns:
|
18 |
-
dictionaly: A set of alignment scores.
|
19 |
-
|
20 |
-
Examples:
|
21 |
-
|
22 |
-
Example 1: Single processing
|
23 |
-
>>> metric = evaluate.load("pytorch-layout-generation/layout-alignment")
|
24 |
-
>>> model_max_length, num_coordinates = 25, 4
|
25 |
-
>>> bbox = np.random.rand(model_max_length, num_coordinates)
|
26 |
-
>>> mask = np.random.choice(a=[True, False], size=(model_max_length,))
|
27 |
-
>>> metric.add(bbox=bbox, mask=mask)
|
28 |
-
>>> print(metric.compute())
|
29 |
-
|
30 |
-
Example 2: Batch processing
|
31 |
-
>>> metric = evaluate.load("pytorch-layout-generation/layout-alignment")
|
32 |
-
>>> batch_size, model_max_length, num_coordinates = 512, 25, 4
|
33 |
-
>>> batch_bbox = np.random.rand(batch_size, model_max_length, num_coordinates)
|
34 |
-
>>> batch_mask = np.random.choice(a=[True, False], size=(batch_size, model_max_length))
|
35 |
-
>>> metric.add_batch(bbox=batch_bbox, mask=batch_mask)
|
36 |
-
>>> print(metric.compute())
|
37 |
-
"""
|
38 |
-
|
39 |
-
_CITATION = """\
|
40 |
-
@inproceedings{lee2020neural,
|
41 |
-
title={Neural design network: Graphic layout generation with constraints},
|
42 |
-
author={Lee, Hsin-Ying and Jiang, Lu and Essa, Irfan and Le, Phuong B and Gong, Haifeng and Yang, Ming-Hsuan and Yang, Weilong},
|
43 |
-
booktitle={Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part III 16},
|
44 |
-
pages={491--506},
|
45 |
-
year={2020},
|
46 |
-
organization={Springer}
|
47 |
-
}
|
48 |
-
|
49 |
-
@article{li2020attribute,
|
50 |
-
title={Attribute-conditioned layout gan for automatic graphic design},
|
51 |
-
author={Li, Jianan and Yang, Jimei and Zhang, Jianming and Liu, Chang and Wang, Christina and Xu, Tingfa},
|
52 |
-
journal={IEEE Transactions on Visualization and Computer Graphics},
|
53 |
-
volume={27},
|
54 |
-
number={10},
|
55 |
-
pages={4039--4048},
|
56 |
-
year={2020},
|
57 |
-
publisher={IEEE}
|
58 |
-
}
|
59 |
-
|
60 |
-
@inproceedings{kikuchi2021constrained,
|
61 |
-
title={Constrained graphic layout generation via latent optimization},
|
62 |
-
author={Kikuchi, Kotaro and Simo-Serra, Edgar and Otani, Mayu and Yamaguchi, Kota},
|
63 |
-
booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
|
64 |
-
pages={88--96},
|
65 |
-
year={2021}
|
66 |
-
}
|
67 |
-
"""
|
68 |
-
|
69 |
-
|
70 |
-
def convert_xywh_to_ltrb(
|
71 |
-
batch_bbox: npt.NDArray[np.float64],
|
72 |
-
) -> Tuple[
|
73 |
-
npt.NDArray[np.float64],
|
74 |
-
npt.NDArray[np.float64],
|
75 |
-
npt.NDArray[np.float64],
|
76 |
-
npt.NDArray[np.float64],
|
77 |
-
]:
|
78 |
-
xc, yc, w, h = batch_bbox
|
79 |
-
x1 = xc - w / 2
|
80 |
-
y1 = yc - h / 2
|
81 |
-
x2 = xc + w / 2
|
82 |
-
y2 = yc + h / 2
|
83 |
-
return (x1, y1, x2, y2)
|
84 |
-
|
85 |
-
|
86 |
-
class LayoutAlignment(evaluate.Metric):
|
87 |
-
def _info(self) -> evaluate.EvaluationModuleInfo:
|
88 |
-
return evaluate.MetricInfo(
|
89 |
-
description=_DESCRIPTION,
|
90 |
-
citation=_CITATION,
|
91 |
-
inputs_description=_KWARGS_DESCRIPTION,
|
92 |
-
features=ds.Features(
|
93 |
-
{
|
94 |
-
"bbox": ds.Sequence(ds.Sequence(ds.Value("float64"))),
|
95 |
-
"mask": ds.Sequence(ds.Value("bool")),
|
96 |
-
}
|
97 |
-
),
|
98 |
-
codebase_urls=[
|
99 |
-
"https://github.com/ktrk115/const_layout/blob/master/metric.py#L167-L188",
|
100 |
-
"https://github.com/CyberAgentAILab/layout-dm/blob/main/src/trainer/trainer/helpers/metric.py#L98-L147",
|
101 |
-
],
|
102 |
-
)
|
103 |
-
|
104 |
-
def _compute_ac_layout_gan(
|
105 |
-
self,
|
106 |
-
S: int,
|
107 |
-
xl: npt.NDArray[np.float64],
|
108 |
-
xc: npt.NDArray[np.float64],
|
109 |
-
xr: npt.NDArray[np.float64],
|
110 |
-
yt: npt.NDArray[np.float64],
|
111 |
-
yc: npt.NDArray[np.float64],
|
112 |
-
yb: npt.NDArray[np.float64],
|
113 |
-
batch_mask: npt.NDArray,
|
114 |
-
) -> npt.NDArray[np.float64]:
|
115 |
-
# shape: (B, 6, S)
|
116 |
-
X = np.stack((xl, xc, xr, yt, yc, yb), axis=1)
|
117 |
-
# shape: (B, 6, S, 1) - (B, 6, 1, S) = (B, 6 S, S)
|
118 |
-
X = X[:, :, :, None] - X[:, :, None, :]
|
119 |
-
|
120 |
-
# shape: (S,)
|
121 |
-
indices = np.arange(S)
|
122 |
-
X[:, :, indices, indices] = 1.0
|
123 |
-
# shape: (B, 6, S, S -> (B, S, 6, S)
|
124 |
-
X = np.abs(X).transpose(0, 2, 1, 3)
|
125 |
-
X[~batch_mask] = 1.0
|
126 |
-
|
127 |
-
# shape: (B, S, 6, S) -> (B, S)
|
128 |
-
X = X.min(axis=(2, 3))
|
129 |
-
X[X == 1.0] = 0.0
|
130 |
-
X = -np.log(1 - X)
|
131 |
-
|
132 |
-
# shape: (B, S) -> (B,)
|
133 |
-
return X.sum(axis=1)
|
134 |
-
|
135 |
-
def _compute_layout_gan_pp(
|
136 |
-
self,
|
137 |
-
score_ac_layout_gan: npt.NDArray[np.float64],
|
138 |
-
batch_mask: npt.NDArray[np.bool_],
|
139 |
-
) -> npt.NDArray[np.float64]:
|
140 |
-
# shape: (B, S) -> (B,)
|
141 |
-
batch_mask = batch_mask.sum(axis=1)
|
142 |
-
|
143 |
-
# shape: (B,)
|
144 |
-
score_normalized = score_ac_layout_gan / batch_mask
|
145 |
-
score_normalized[np.isnan(score_normalized)] = 0.0
|
146 |
-
return score_normalized
|
147 |
-
|
148 |
-
def _compute_neural_design_network(
|
149 |
-
self,
|
150 |
-
xl: npt.NDArray[np.float64],
|
151 |
-
xc: npt.NDArray[np.float64],
|
152 |
-
xr: npt.NDArray[np.float64],
|
153 |
-
batch_mask: npt.NDArray[np.bool_],
|
154 |
-
S: int,
|
155 |
-
):
|
156 |
-
# shape: (B, 3, S)
|
157 |
-
Y = np.stack((xl, xc, xr), axis=1)
|
158 |
-
# shape: (B, 3, S, S)
|
159 |
-
Y = Y[:, :, None, :] - Y[:, :, :, None]
|
160 |
-
|
161 |
-
# shape: (B, S) -> (B, S, S)
|
162 |
-
batch_mask = ~batch_mask[:, None, :] | ~batch_mask[:, :, None]
|
163 |
-
# shape: (B,)
|
164 |
-
indices = np.arange(S)
|
165 |
-
batch_mask[:, indices, indices] = True
|
166 |
-
|
167 |
-
# shape: (B, S, S) -> (B, 1, S, S) -> (B, 3, S, S)
|
168 |
-
batch_mask = np.repeat(batch_mask[:, None, :, :], repeats=3, axis=1)
|
169 |
-
Y[batch_mask] = 1.0
|
170 |
-
|
171 |
-
# shape: (B, 3, S, S) -> (B, S, S) -> (B, S)
|
172 |
-
Y = np.abs(Y).min(axis=(1, 2))
|
173 |
-
Y[Y == 1.0] = 0.0
|
174 |
-
|
175 |
-
# shape: (B, S) -> (B,)
|
176 |
-
score = Y.sum(axis=1)
|
177 |
-
return score
|
178 |
-
|
179 |
-
def _compute(
|
180 |
-
self,
|
181 |
-
*,
|
182 |
-
bbox: Union[npt.NDArray[np.float64], List[List[int]]],
|
183 |
-
mask: Union[npt.NDArray[np.bool_], List[List[bool]]],
|
184 |
-
) -> Dict[str, npt.NDArray[np.float64]]:
|
185 |
-
# shape: (B, model_max_length, C)
|
186 |
-
bbox = np.array(bbox)
|
187 |
-
# shape: (B, model_max_length)
|
188 |
-
mask = np.array(mask)
|
189 |
-
|
190 |
-
# S: model_max_length
|
191 |
-
_, S, _ = bbox.shape
|
192 |
-
|
193 |
-
# shape: (B, S, C) -> (C, B, S)
|
194 |
-
bbox = bbox.transpose(2, 0, 1)
|
195 |
-
xl, yt, xr, yb = convert_xywh_to_ltrb(bbox)
|
196 |
-
xc, yc = bbox[0], bbox[1]
|
197 |
-
|
198 |
-
# shape: (B,)
|
199 |
-
score_ac_layout_gan = self._compute_ac_layout_gan(
|
200 |
-
S=S, xl=xl, xc=xc, xr=xr, yt=yt, yc=yc, yb=yb, batch_mask=mask
|
201 |
-
)
|
202 |
-
# shape: (B,)
|
203 |
-
score_layout_gan_pp = self._compute_layout_gan_pp(
|
204 |
-
score_ac_layout_gan=score_ac_layout_gan, batch_mask=mask
|
205 |
-
)
|
206 |
-
score_ndn = self._compute_neural_design_network(
|
207 |
-
xl=xl, xc=xc, xr=xr, batch_mask=mask, S=S
|
208 |
-
)
|
209 |
-
return {
|
210 |
-
"alignment-ACLayoutGAN": score_ac_layout_gan,
|
211 |
-
"alignment-LayoutGAN++": score_layout_gan_pp,
|
212 |
-
"alignment-NDN": score_ndn,
|
213 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|