File size: 4,546 Bytes
53ef34c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn

class ContextEmbedding(nn.Module):
    def __init__(self, vocab_size, embedding_dim):
        super(ContextEmbedding, self).__init__()
        
        self.embedd_fc1 = nn.Embedding(vocab_size, embedding_dim)
        self.fc1 = nn.Linear(embedding_dim, embedding_dim)
    
    def forward(self, x):
        x = self.embedd_fc1(x)
        x = self.fc1(x)
        return x

class TimeEbedding(nn.Module):
    def __init__(self, time_dim, hidden_dim, out_dim):
        super(TimeEbedding, self).__init__()
        self.fc1 = nn.Linear(time_dim, hidden_dim)
        self.gelu = nn.GELU()
        self.fc2 = nn.Linear(hidden_dim, out_dim)
    
    def forward(self, x):
        x = self.fc1(x)
        x = self.gelu(x)
        x = self.fc2(x)
        
        return x

class ConvBlock(nn.Module):
    def __init__(self, in_channels, out_channels, vocab_size, embedding_dim , residual=False):
        super(ConvBlock, self).__init__()
        
        self.conv1 = nn.Conv2d(in_channels, out_channels, 2, padding="same")
        self.conv2 = nn.Conv2d(out_channels, out_channels, 2, padding="same")
        self.downsample = nn.Conv2d(in_channels, out_channels, 3, padding="same", bias=False)
        self.relu = nn.ReLU()
        self.bn = nn.BatchNorm2d(out_channels)
        self.te = TimeEbedding(1, out_channels, out_channels)
        self.context_embedding = ContextEmbedding(vocab_size, embedding_dim)
        self.__residual = residual
    
    def forward(self, x, t, context):
        x1 = self.downsample(x)
        x = self.conv1(x)
        x = self.relu(x)
        x = self.conv2(x)
        x = self.relu(x)
        
        time_embedding = self.te(t)[:, :, None, None]
        context = self.context_embedding(context)[:, :, None, None]
        
        if self.__residual:
            return self.bn(x + x1) * context + time_embedding
        return x * context + time_embedding

class UNET(nn.Module):
    def __init__(self, in_channels, out_channels, residual=False):
        super(UNET, self).__init__()

        self.convBD1 = ConvBlock(in_channels, 64, 5, 64,residual)
        self.convBD2 = ConvBlock(64, 128, 5, 128,residual)
        self.convBD3 = ConvBlock(128, 256, 5, 256,residual)
        # self.convBD4 = ConvBlock(256, 512, 5, 512,residual)
        
        self.bottle_neck = ConvBlock(256, 512, 5, 512,residual)
        # self.bottle_neck = ConvBlock(512, 1024, 5, 1024,residual)
        
        # self.convBU1 = ConvBlock(1024, 512, 5, 512,residual)
        self.convBU2 = ConvBlock(512, 256, 5, 256,residual)
        self.convBU3 = ConvBlock(256, 128, 5, 128,residual)
        self.convBU4 = ConvBlock(128, 64, 5, 64,residual)
        
        self.convT1 = nn.ConvTranspose2d(1024, 512, 2, 2)
        self.convT2 = nn.ConvTranspose2d(512, 256, 2, 2)
        self.convT3 = nn.ConvTranspose2d(256, 128, 2, 2)
        self.convT4 = nn.ConvTranspose2d(128, 64, 2, 2)
        
        self.final = nn.Conv2d(64, out_channels, 1)

        self.maxpool = nn.MaxPool2d(2)
    def forward(self, x, t, context):
        if x.ndim == 3:
            x = x.unsqueeze(0)
        
        x1 = self.convBD1(x, t, context)
        x = self.maxpool(x1)
        
        x2 = self.convBD2(x, t, context)
        x = self.maxpool(x2)
        
        x3 = self.convBD3(x, t, context)
        x = self.maxpool(x3)
        
        # x4 = self.convBD4(x, t, context)
        # x = self.maxpool(x4)
        
        x = self.bottle_neck(x, t, context)
        
        # x = self.convT1(x)
        # x = torch.cat([x4, x], dim=1)
        # x = self.convBU1(x, t, context)
        
        x = self.convT2(x)
        x = torch.cat([x3, x], dim=1)
        x = self.convBU2(x, t, context)
        
        x = self.convT3(x)
        x = torch.cat([x2, x], dim=1)
        x = self.convBU3(x, t, context)
        
        x = self.convT4(x)
        x = torch.cat([x1, x], dim=1)
        x = self.convBU4(x, t, context)
        
        x = self.final(x)
        
        return x

import torch
import torch.nn as nn
from model import UNET
from helper import DDPM, denoise_image

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
timesteps = 500
beta1 = 1e-4
beta2 = 0.02
betas = (beta2 - beta1) * torch.linspace(0, 1, timesteps + 1) + beta1
betas = betas.to(device)
alpha = 1.0 - betas
alpha_bar = torch.cumprod(alpha, dim=0).to(device)

model = UNET(3, 3,True)
model = model.to(device)
loss_fn = nn.MSELoss()
optim = torch.optim.Adam(model.parameters(), lr=1e-3)
sampler = DDPM(betas)