-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
236 lines (179 loc) · 8.07 KB
/
models.py
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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple
from typing import List
from utils import NCCLoss
class SinusoidalPositionEmbeddings(nn.Module):
def __init__(self, dim: int):
super().__init__()
self.dim = dim
def forward(self, time):
device = time.device
half_dim = self.dim // 2
embeddings = math.log(10000) / (half_dim - 1)
embeddings = torch.exp(torch.arange(half_dim, device=device) * -embeddings)
embeddings = time[:, None] * embeddings[None, :]
embeddings = torch.cat((embeddings.sin(), embeddings.cos()), dim=-1)
return embeddings
class Block3d(nn.Module):
def __init__(self, in_channels: int, out_channels: int, time_emb_dim: int, up: bool=False):
super().__init__()
self.time_mlp = nn.Linear(time_emb_dim, out_channels)
if up:
# up-sampling (decoder part)
self.conv1 = nn.Conv3d(2*in_channels, out_channels, 3, padding=1)
self.transform = nn.Sequential(nn.Conv3d(out_channels, out_channels, 3, padding=1),
nn.Upsample(scale_factor=2, mode='trilinear'))
else:
# down-sampling (encoder part)
self.conv1 = nn.Conv3d(in_channels, out_channels, 3, padding=1)
self.transform = nn.Conv3d(out_channels, out_channels, 4, 2, 1)
self.conv2 = nn.Conv3d(out_channels, out_channels, 3, padding=1)
self.bnorm1 = nn.BatchNorm3d(out_channels)
self.bnorm2 = nn.BatchNorm3d(out_channels)
def forward(self, x:torch.Tensor, t:torch.Tensor):
# First Conv
h = self.bnorm1(F.silu(self.conv1(x)))
# Time embedding
time_emb = self.time_mlp(t)
# Extend last 3 dimensions
time_emb = time_emb[(..., ) + (None, ) * 3]
# Add time channel
h = h + time_emb
# Second Conv
h = self.bnorm2(F.silu(self.conv2(h)))
# Down or Upsample
out = F.silu(self.transform(h))
return out
class UNet3d(nn.Module):
"""
A simplified variant of the Unet architecture.
"""
def __init__(self, in_channels: int=1, out_channels: int=1,
down_channels: List|Tuple=(32, 32, 32),
up_channels: List|Tuple=(32, 32, 32),
time_emb_dim: int=32) -> None:
super().__init__()
# Time embedding
self.time_mlp = nn.Sequential(
SinusoidalPositionEmbeddings(time_emb_dim),
nn.Linear(time_emb_dim, time_emb_dim),
nn.SiLU()
)
# Initial projection
self.conv0 = nn.Conv3d(in_channels, down_channels[0], 3, padding=1)
# Downsample
self.downs = nn.ModuleList([Block3d(down_channels[i], down_channels[i+1], time_emb_dim) for i in range(len(down_channels)-1)])
# Upsample
self.ups = nn.ModuleList([Block3d(up_channels[i], up_channels[i+1], time_emb_dim, up=True) for i in range(len(up_channels)-1)])
# Final projection
self.output = nn.Conv3d(up_channels[-1], out_channels, 3, padding=1)
def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:
# Embedd time
t = self.time_mlp(timestep)
# Initial conv
x = self.conv0(x)
# Unet
residual_inputs = []
for down in self.downs:
x = down(x, t)
residual_inputs.append(x)
for up in self.ups:
residual_x = residual_inputs.pop()
# Add residual x as additional channels
x = torch.cat((x, residual_x), dim=1)
x = up(x, t)
return self.output(x)
class FlowNet3D(nn.Module):
"""
Implementation of both Time-Independet and Time-Dependent Phi network based on interpolative cmposition
"""
def __init__(self,
down_channels: List|Tuple=(32, 32, 32),
up_channels: List|Tuple=(32, 32, 32),
loss_type: str='ncc') -> None:
super().__init__()
self.loss_type = loss_type
self.ncc_loss = NCCLoss(win=11)
self.net = UNet3d(in_channels=2,
out_channels=3,
down_channels=down_channels,
up_channels=up_channels,
time_emb_dim=64)
def forward(self, I: torch.Tensor, J: torch.Tensor, xyz: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
"""
Args:
I (torch.Tensor): fixed image with size [B, 1, D, H, W]
J (torch.Tensor): moving image with size [B, 1, D, H, W]
xyz (torch.Tensor): identity grid with size [B, D, H, W, 3]
t (torch.Tensor): sampled time with size [B]
Returns:
torch.Tensor: the deformation grid at time t with size [B, D, H, W, 3]
"""
flow = t * self.velocity(I, J, t)
phi_t = self.make_grid(flow, xyz)
return phi_t
def velocity(self, I: torch.Tensor, J: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
"""
Args:
I (torch.Tensor): fixed image with size [B, 1, D, H, W]
J (torch.Tensor): moving image with size [B, 1, D, H, W]
t (torch.Tensor): sampled time with size [1]
Returns:
torch.Tensor: the vector field at time t with size [B, 3, D, H, W]
"""
u_in = torch.cat([I, J], dim=1)
velocity = self.net(u_in, t)
return velocity
def loss_flow(self, I: torch.Tensor, J: torch.Tensor, xyz: torch.Tensor, res: float) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
I (torch.Tensor): fixed image with size [B, 1, D, H, W]
J (torch.Tensor): moving image with size [B, 1, D, H, W]
xyz (torch.Tensor): identity grid with size [B, D, H, W, 3]
res (float): the resolution at which the ncc loss is computed
Returns:
torch.Tensor, torch.Tensor: ncc loss, semigroup loss
"""
t = torch.rand(1, device=I.device)
flow_J = t * self.velocity(I, J, t)
Jw = self.warp(J, flow_J, xyz)
flow_I = (t - 1.) * self.velocity(I, J, t - 1.)
Iw = self.warp(I, flow_I, xyz)
if res != 1:
Iw = F.interpolate(Iw, scale_factor=res, mode='trilinear')
Jw = F.interpolate(Jw, scale_factor=res, mode='trilinear')
if self.loss_type == 'mse':
image_loss = res * F.mse_loss(Jw, Iw)
else:
image_loss = res * self.ncc_loss(Jw, Iw)
flow_I_J = self.compose(flow_I, flow_J, xyz)
grid_I_J = self.make_grid(flow_I_J, xyz)
flow_J_I = self.compose(flow_J, flow_I, xyz)
grid_J_I = self.make_grid(flow_J_I, xyz)
flow = (2. * t - 1.) * self.velocity(I, J, 2. * t - 1.)
grid = self.make_grid(flow, xyz)
flow_loss = 0.5 * (torch.mean((grid - grid_I_J) ** 2) + torch.mean((grid - grid_J_I) ** 2))
return image_loss, flow_loss
def make_grid(self, flow: torch.Tensor, grid: torch.Tensor) -> torch.Tensor:
phi = grid + flow.permute(0, 2, 3, 4, 1)
phi = self.grid_normalizer(phi)
return phi
def warp(self, image: torch.Tensor, flow: torch.Tensor, grid: torch.Tensor) -> torch.Tensor:
grid = grid + flow.permute(0, 2, 3, 4, 1)
grid = self.grid_normalizer(grid)
warped = F.grid_sample(image, grid, padding_mode='reflection', align_corners=True)
return warped
def compose(self, flow1: torch.Tensor, flow2: torch.Tensor, grid: torch.Tensor):
grid = grid + flow2.permute(0, 2, 3, 4, 1)
grid = self.grid_normalizer(grid)
composed_flow = F.grid_sample(flow1, grid, padding_mode='reflection', align_corners=True) + flow2
return composed_flow
def grid_normalizer(self, grid: torch.Tensor) -> torch.Tensor:
_, d, h, w, _ = grid.size()
grid[:, :, :, :, 0] = (grid[:, :, :, :, 0] - ((w - 1) / 2)) / (w - 1) * 2
grid[:, :, :, :, 1] = (grid[:, :, :, :, 1] - ((h - 1) / 2)) / (h - 1) * 2
grid[:, :, :, :, 2] = (grid[:, :, :, :, 2] - ((d - 1) / 2)) / (d - 1) * 2
return grid