Create analogy_input_processor.py
Browse files
analogy_input_processor/analogy_input_processor.py
ADDED
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
ADOBE CONFIDENTIAL
|
3 |
+
Copyright 2024 Adobe
|
4 |
+
All Rights Reserved.
|
5 |
+
NOTICE: All information contained herein is, and remains
|
6 |
+
the property of Adobe and its suppliers, if any. The intellectual
|
7 |
+
and technical concepts contained herein are proprietary to Adobe
|
8 |
+
and its suppliers and are protected by all applicable intellectual
|
9 |
+
property laws, including trade secret and copyright laws.
|
10 |
+
Dissemination of this information or reproduction of this material
|
11 |
+
is strictly forbidden unless prior written permission is obtained
|
12 |
+
from Adobe.
|
13 |
+
"""
|
14 |
+
|
15 |
+
import torch as th
|
16 |
+
from torchvision import transforms
|
17 |
+
from diffusers import ModelMixin
|
18 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
19 |
+
|
20 |
+
DINO_SIZE = 224
|
21 |
+
DINO_MEAN = [0.485, 0.456, 0.406]
|
22 |
+
DINO_STD = [0.229, 0.224, 0.225]
|
23 |
+
|
24 |
+
SIGLIP_SIZE = 256
|
25 |
+
SIGLIP_MEAN = [0.5]
|
26 |
+
SIGLIP_STD = [0.5]
|
27 |
+
|
28 |
+
|
29 |
+
class AnalogyInputProcessor(ModelMixin, ConfigMixin):
|
30 |
+
|
31 |
+
@register_to_config
|
32 |
+
def __init__(self,):
|
33 |
+
super(AnalogyInputProcessor, self).__init__()
|
34 |
+
|
35 |
+
self.dino_transform = transforms.Compose(
|
36 |
+
[
|
37 |
+
transforms.Resize((DINO_SIZE, DINO_SIZE)),
|
38 |
+
transforms.ToTensor(),
|
39 |
+
transforms.Normalize(DINO_MEAN, DINO_STD),
|
40 |
+
]
|
41 |
+
)
|
42 |
+
|
43 |
+
self.siglip_transform = transforms.Compose(
|
44 |
+
[
|
45 |
+
transforms.Resize((SIGLIP_SIZE, SIGLIP_SIZE)),
|
46 |
+
transforms.ToTensor(),
|
47 |
+
transforms.Normalize(SIGLIP_MEAN, SIGLIP_STD),
|
48 |
+
]
|
49 |
+
)
|
50 |
+
|
51 |
+
dino_mean = th.tensor(DINO_MEAN).view(1, 3, 1, 1)
|
52 |
+
dino_std = th.tensor(DINO_STD).view(1, 3, 1, 1)
|
53 |
+
siglip_mean = [SIGLIP_MEAN[0],] * 3
|
54 |
+
siglip_std = [SIGLIP_STD[0],] * 3
|
55 |
+
siglip_mean = th.tensor(siglip_mean).view(1, 3, 1, 1)
|
56 |
+
siglip_std = th.tensor(siglip_std).view(1, 3, 1, 1)
|
57 |
+
self.register_buffer("dino_mean", dino_mean)
|
58 |
+
self.register_buffer("dino_std", dino_std)
|
59 |
+
self.register_buffer("siglip_mean", siglip_mean)
|
60 |
+
self.register_buffer("siglip_std", siglip_std)
|
61 |
+
|
62 |
+
def __call__(self, analogy_prompt):
|
63 |
+
# List of tuples of (A, A*, B)
|
64 |
+
img_a_dino = []
|
65 |
+
img_a_siglip = []
|
66 |
+
img_a_star_dino = []
|
67 |
+
img_a_star_siglip = []
|
68 |
+
img_b_dino = []
|
69 |
+
img_b_siglip = []
|
70 |
+
|
71 |
+
for im_set in analogy_prompt:
|
72 |
+
img_a, img_a_star, img_b = im_set
|
73 |
+
img_a_dino.append(self.dino_transform(img_a))
|
74 |
+
img_a_siglip.append(self.siglip_transform(img_a))
|
75 |
+
img_a_star_dino.append(self.dino_transform(img_a_star))
|
76 |
+
img_a_star_siglip.append(self.siglip_transform(img_a_star))
|
77 |
+
img_b_dino.append(self.dino_transform(img_b))
|
78 |
+
img_b_siglip.append(self.siglip_transform(img_b))
|
79 |
+
|
80 |
+
img_a_dino = th.stack(img_a_dino, 0)
|
81 |
+
img_a_siglip = th.stack(img_a_siglip, 0)
|
82 |
+
img_a_star_dino = th.stack(img_a_star_dino, 0)
|
83 |
+
img_a_star_siglip = th.stack(img_a_star_siglip, 0)
|
84 |
+
img_b_dino = th.stack(img_b_dino, 0)
|
85 |
+
img_b_siglip = th.stack(img_b_siglip, 0)
|
86 |
+
|
87 |
+
dino_combined_input = th.stack([img_b_dino, img_a_dino, img_a_star_dino], 0)
|
88 |
+
siglip_combined_input = th.stack([img_b_siglip, img_a_siglip, img_a_star_siglip], 0)
|
89 |
+
|
90 |
+
return dino_combined_input, siglip_combined_input
|
91 |
+
def get_negative(self, dino_in, siglip_in):
|
92 |
+
|
93 |
+
dino_i = ((dino_in * 0 + 0.5) - self.dino_mean) / self.dino_std
|
94 |
+
siglip_i = ((siglip_in * 0 + 0.5) - self.siglip_mean) / self.siglip_std
|
95 |
+
return dino_i, siglip_i
|
96 |
+
|