Hugging Face's logo

BlinkDL
/
clip-guided-binary-autoencoder

Model card Files Community
1
clip-guided-binary-autoencoder
File size: 6,400 Bytes 
f7f9895
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ee2c15
6f9f76f
dd4d734
f7f9895
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd4d734
f7f9895
 
dd4d734
 
f7f9895
dd4d734
f7f9895
 
 
 
 
 
 
 
 
 
 
 
6f9f76f
 
f7f9895
6f9f76f
 
 
 
 
 
 
 
 
 
 
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
 
########################################################################################################
# The RWKV Language Model - https://github.com/BlinkDL/RWKV-LM
########################################################################################################

import torch, types, os
import numpy as np
from PIL import Image
import torch.nn as nn
from torch.nn import functional as F
import torchvision as vision
import torchvision.transforms as transforms
np.set_printoptions(precision=4, suppress=True, linewidth=200)
print(f'loading...')

########################################################################################################

model_prefix = 'out-v7c_d8_256-224-13bit-OB32x0.5-745'
input_imgs = ['lena.png', 'genshin.png', 'kodim14-modified.png', 'kodim19-modified.png', 'kodim24-modified.png']
device = 'cpu' # cpu cuda

########################################################################################################

class ToBinary(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x):
        return torch.floor(x + 0.5) # no need for noise when we have plenty of data

    @staticmethod
    def backward(ctx, grad_output):
        return grad_output.clone() # pass-through

class R_ENCODER(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.args = args
        dd = 8
        self.Bxx = nn.BatchNorm2d(dd*64)

        self.CIN = nn.Conv2d(3, dd, kernel_size=3, padding=1)
        self.Cx0 = nn.Conv2d(dd, 32, kernel_size=3, padding=1)
        self.Cx1 = nn.Conv2d(32, dd, kernel_size=3, padding=1)

        self.B00 = nn.BatchNorm2d(dd*4)
        self.C00 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
        self.C01 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)
        self.C02 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
        self.C03 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)

        self.B10 = nn.BatchNorm2d(dd*16)
        self.C10 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
        self.C11 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)
        self.C12 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
        self.C13 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)

        self.B20 = nn.BatchNorm2d(dd*64)
        self.C20 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
        self.C21 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)
        self.C22 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
        self.C23 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)

        self.COUT = nn.Conv2d(dd*64, args.my_img_bit, kernel_size=3, padding=1)

    def forward(self, img):
        ACT = F.mish

        x = self.CIN(img)
        xx = self.Bxx(F.pixel_unshuffle(x, 8))
        x = x + self.Cx1(ACT(self.Cx0(x)))

        x = F.pixel_unshuffle(x, 2)
        x = x + self.C01(ACT(self.C00(ACT(self.B00(x)))))
        x = x + self.C03(ACT(self.C02(x)))

        x = F.pixel_unshuffle(x, 2)
        x = x + self.C11(ACT(self.C10(ACT(self.B10(x)))))
        x = x + self.C13(ACT(self.C12(x)))

        x = F.pixel_unshuffle(x, 2)
        x = x + self.C21(ACT(self.C20(ACT(self.B20(x)))))
        x = x + self.C23(ACT(self.C22(x)))

        x = self.COUT(x + xx)
        return torch.sigmoid(x)

class R_DECODER(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.args = args
        dd = 8
        self.CIN = nn.Conv2d(args.my_img_bit, dd*64, kernel_size=3, padding=1)

        self.B00 = nn.BatchNorm2d(dd*64)
        self.C00 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
        self.C01 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)
        self.C02 = nn.Conv2d(dd*64, 256, kernel_size=3, padding=1)
        self.C03 = nn.Conv2d(256, dd*64, kernel_size=3, padding=1)

        self.B10 = nn.BatchNorm2d(dd*16)
        self.C10 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
        self.C11 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)
        self.C12 = nn.Conv2d(dd*16, 256, kernel_size=3, padding=1)
        self.C13 = nn.Conv2d(256, dd*16, kernel_size=3, padding=1)

        self.B20 = nn.BatchNorm2d(dd*4)
        self.C20 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
        self.C21 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)
        self.C22 = nn.Conv2d(dd*4, 256, kernel_size=3, padding=1)
        self.C23 = nn.Conv2d(256, dd*4, kernel_size=3, padding=1)

        self.Cx0 = nn.Conv2d(dd, 32, kernel_size=3, padding=1)
        self.Cx1 = nn.Conv2d(32, dd, kernel_size=3, padding=1)
        self.COUT = nn.Conv2d(dd, 3, kernel_size=3, padding=1)

    def forward(self, code):
        ACT = F.mish
        x = self.CIN(code)

        x = x + self.C01(ACT(self.C00(ACT(self.B00(x)))))
        x = x + self.C03(ACT(self.C02(x)))
        x = F.pixel_shuffle(x, 2)

        x = x + self.C11(ACT(self.C10(ACT(self.B10(x)))))
        x = x + self.C13(ACT(self.C12(x)))
        x = F.pixel_shuffle(x, 2)

        x = x + self.C21(ACT(self.C20(ACT(self.B20(x)))))
        x = x + self.C23(ACT(self.C22(x)))
        x = F.pixel_shuffle(x, 2)

        x = x + self.Cx1(ACT(self.Cx0(x)))
        x = self.COUT(x)
        
        return torch.sigmoid(x)

########################################################################################################

print(f'building model {model_prefix}...')
args = types.SimpleNamespace()
args.my_img_bit = 13
encoder = R_ENCODER(args).eval().to(device)
decoder = R_DECODER(args).eval().to(device)

zpow = torch.tensor([2**i for i in range(0,13)]).reshape(13,1,1).to(device).long()

encoder.load_state_dict(torch.load(f'{model_prefix}-E.pth'))
decoder.load_state_dict(torch.load(f'{model_prefix}-D.pth'))

########################################################################################################

img_transform = transforms.Compose([
    transforms.PILToTensor(),
    transforms.ConvertImageDtype(torch.float),
    transforms.Resize((224, 224))
])

for input_img in input_imgs:
    print(f'test image {input_img}...')

    with torch.no_grad():
        img = img_transform(Image.open(f'img_test/{input_img}')).unsqueeze(0).to(device)
        z = encoder(img)
        z = ToBinary.apply(z)

        zz = torch.sum(z.squeeze().long() * zpow, dim=0)
        print(f'Code shape = {zz.shape}\n{zz.cpu().numpy()}\n')
        
        out = decoder(z)
        vision.utils.save_image(out, f"img_test/{input_img.split('.')[0]}-out-13bit.png")