INR-SSM/models.py at main · multimodallearning/INR-SSM · GitHub

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
import numpy as np
from functools import partial
import torch
import torch.nn as nn
import torch.nn.functional as F
#from hashencoding import MultiResHashGrid

def weights_init(m, scale=30):
    with torch.no_grad():
        if isinstance(m, nn.Linear):
            hidden_ch = max(m.weight.shape[0],m.weight.shape[1])
            val = np.sqrt(6/hidden_ch)/scale if m.weight.shape[1]>4 else 1 / m.weight.shape[1]
            torch.nn.init.uniform_(m.weight,-val,val)
            if m.bias is not None:
                torch.nn.init.zeros_(m.bias)

#convlinear is work-around to have different initialisation modes for sine and relu layers
class ConvLinear(nn.Module):
    def __init__(self,in_features, out_features):
        super().__init__()
        self.conv = nn.Conv2d(in_features,out_features,1)
    def forward(self,x):
        return self.conv(x.permute(0,3,1,2)).permute(0,2,3,1)
class ReLUSiren(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        listnet = [nn.Linear(in_features,hidden_ch)]
        for i in range(num_layers):
            listnet.append(ConvLinear(hidden_ch,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self,x):
        scale = self.scale
        x = torch.sin(scale*self.net[0](x))
        x = torch.sin(scale*self.net[1](x))
        for i in range(2,len(self.net)-2):
            x = F.relu(self.net[i](x))#
        x = torch.sin(scale*self.net[-2](x))
        return self.net[-1](x)

class Siren(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        listnet = [nn.Linear(in_features,hidden_ch)]
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self,x):
        scale = self.scale
        x = torch.sin(scale*self.net[0](x))
        for i in range(1,len(self.net)-1):
            x = torch.sin(scale*self.net[i](x))
        return self.net[-1](x)

class SirenModulated(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        self.hidden_ch = hidden_ch
        listnet = [nn.Linear(2,hidden_ch)]
        modulation_listnet = []
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch,hidden_ch))
            modulation_listnet.append(nn.Linear(in_features-2,hidden_ch*2))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)
        self.mod = nn.Sequential(*modulation_listnet)


    def forward(self,x):
        scale = self.scale
        hdim = self.hidden_ch
        y = torch.sin(scale*self.net[0](x[...,:2]))
        for i in range(1,len(self.net)-1):
            phi_psi = self.mod[i-1](x[...,2:])
            y = torch.sin(scale*(phi_psi[...,:hdim]*self.net[i](y))+phi_psi[...,hdim:])
        return self.net[-1](y)

class ResSiren(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        listnet = [nn.Linear(in_features,hidden_ch)]
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self,x):
        scale = self.scale
        x = torch.sin(scale*self.net[0](x))
        for i in range(1,len(self.net)-1):
            x = torch.sin(scale*self.net[i](x)) + x
        return self.net[-1](x)

class SkipSiren(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        listnet = [nn.Linear(in_features,hidden_ch)]
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch+in_features,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self,x):
        scale = self.scale
        y = torch.sin(scale*self.net[0](x))
        for i in range(1,len(self.net)-1):
            y = torch.sin(scale*self.net[i](torch.cat((y,x),dim=-1)))
        return self.net[-1](y)

class SirenTanH(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        listnet = [nn.Linear(in_features,hidden_ch)]
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self,x):
        scale = self.scale
        x = torch.sin(scale*self.net[0](x))
        for i in range(1,len(self.net)-1):
            x = torch.sin(scale*self.net[i](x))
        return torch.tanh(self.net[-1](x))

class FINERSiren(nn.Module):
    def __init__(self,in_features, out_features, hidden_ch=256,scale=30,num_layers=3):
        super().__init__()
        self.scale = scale
        listnet = [nn.Linear(in_features,hidden_ch)]
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self,x):
        scale = self.scale
        x = torch.sin(scale*self.net[0](x))
        for i in range(1,len(self.net)-1):
            x = torch.sin(scale*self.net[i](x)*(torch.abs(x)+1))
        return self.net[-1](x)

class HashGridINR(nn.Module):
    def __init__(self,in_features, out_features, H, hidden_ch=256,num_layers=3):
        super().__init__()
        self.hashgrid = MultiResHashGrid(2, base_resolution=8, log2_hashmap_size=12, finest_resolution=H).cuda()
        L = 16
        listnet = [nn.Linear(L*2 + in_features-2,hidden_ch)]
        for i in range(num_layers):
            listnet.append(nn.Linear(hidden_ch,hidden_ch))
        listnet.append(nn.Linear(hidden_ch,out_features))
        self.net = nn.Sequential(*listnet)

    def forward(self, x):
        x_hashed = self.hashgrid(x[...,:2])

        x = torch.cat((x_hashed, x[...,2:]),dim=-1)
        x = F.relu(self.net[0](x))
        for i in range(1,len(self.net)-1):
            x = F.relu(self.net[i](x))
        return self.net[-1](x)