-
Notifications
You must be signed in to change notification settings - Fork 7
/
refinenet.py
189 lines (143 loc) · 6.17 KB
/
refinenet.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
from torch import optim, nn
from model_utils import pre_bgr_image, speedy_bargmax2d
from metrics import Refinenet_Metrics
import torch
import numpy as np
import pytorch_lightning as pl
class RefineNet(torch.nn.Module):
def __init__(self):
super(RefineNet, self).__init__()
self.relu = torch.nn.ReLU(inplace=True)
self.pool = torch.nn.MaxPool2d(kernel_size=2, stride=2, return_indices=False)
c1, c2, c3, c4, c5 = 64, 128, 128, 128, 64 # Small architecture
self.last_c = 64
self.up_sample = torch.nn.UpsamplingNearest2d(scale_factor=2)
self.conv1a = torch.nn.Conv2d(1, c1, kernel_size=3, stride=1, padding=0)
self.bn1a = nn.BatchNorm2d(c1)
self.conv1b = torch.nn.Conv2d(c1, c1, kernel_size=3, stride=1, padding=0)
self.bn1b = nn.BatchNorm2d(c1)
self.conv2a = torch.nn.Conv2d(c1, c2, kernel_size=3, stride=1, padding=0)
self.bn2a = nn.BatchNorm2d(c2)
self.conv2b = torch.nn.Conv2d(c2, c2, kernel_size=3, stride=1, padding=0)
self.bn2b = nn.BatchNorm2d(c2)
self.conv3a = torch.nn.Conv2d(c2, c3, kernel_size=3, stride=1, padding=1)
self.bn3a = nn.BatchNorm2d(c3)
self.conv3b = torch.nn.Conv2d(c3, c3, kernel_size=3, stride=1, padding=1)
self.bn3b = nn.BatchNorm2d(c3)
self.conv4a = torch.nn.Conv2d(c3, c4, kernel_size=3, stride=1, padding=1)
self.bn4a = nn.BatchNorm2d(c4)
self.conv4b = torch.nn.Conv2d(c4, c4, kernel_size=3, stride=1, padding=1)
self.bn4b = nn.BatchNorm2d(c4)
self.conv5a = torch.nn.Conv2d(c4, c5, kernel_size=3, stride=1, padding=1)
self.bn5a = nn.BatchNorm2d(c5)
self.conv5b = torch.nn.Conv2d(c5, c5, kernel_size=3, stride=1, padding=1)
self.bn5b = nn.BatchNorm2d(c5)
# Detector Head.
self.convPa = torch.nn.Conv2d(c5, self.last_c, kernel_size=3, stride=1, padding=1)
self.bnPa = nn.BatchNorm2d(self.last_c)
self.convPb = torch.nn.Conv2d(self.last_c, 1, kernel_size=1, stride=1, padding=0)
def forward(self, x):
"""
Input
x: Image pytorch tensor shaped N x 1 x 24 x 24.
"""
# Let's stick to this version: first BN, then relu
x = self.relu(self.bn1a(self.conv1a(x)))
x = self.relu(self.bn1b(self.conv1b(x)))
x = self.relu(self.bn2a(self.conv2a(x)))
x = self.relu(self.bn2b(self.conv2b(x)))
x = self.pool(x)
x = self.relu(self.bn3a(self.conv3a(x)))
x = self.relu(self.bn3b(self.conv3b(x)))
x = self.up_sample(x)
x = self.relu(self.bn4a(self.conv4a(x)))
x = self.relu(self.bn4b(self.conv4b(x)))
x = self.up_sample(x)
x = self.relu(self.bn5a(self.conv5a(x)))
x = self.relu(self.bn5b(self.conv5b(x)))
x = self.up_sample(x)
# Head
cPa = self.relu(self.bnPa(self.convPa(x)))
loc = self.convPb(cPa)
return loc
def infer_patches(self, patches: torch.Tensor,
keypoints: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
"""
Inference on 24x24 patches, assuming pre processing
Parameters
----------
patches : np.ndarray
The patches (N, 24, 24) or (N, 1, 24, 24)
keypoints : np.ndarray
The keypoints corresponding to the patches in (x, y) format
Returns
-------
tuple(np.ndarray, np.ndarray)
corners in 8x resolution (in original image)
and corners in 64x64 window
"""
assert patches.shape[-2:] == (24, 24)
device = next(self.parameters()).device
with torch.no_grad():
if patches.ndim == 3:
patches = torch.unsqueeze(patches, dim=1)
loc_hat = self(patches)
loc_hat = loc_hat[:, 0, ...] # TODO: REMOVE ELLIPSIS
# loc_hat: (N, H/8, W/8)
corners = speedy_bargmax2d(loc_hat)
# Add keypoints to center on keypoints, divide by 8 to account for 8x resolution
corners_og = (corners - 32) / 8 + keypoints # Is 32 right? :)
return corners_og, corners
def conv(in_planes, out_planes, kernel_size=3):
return nn.Sequential(
nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size,
padding=(kernel_size - 1) // 2, stride=2),
nn.ReLU(inplace=True)
)
def upconv(in_planes, out_planes):
return nn.Sequential(
nn.ConvTranspose2d(in_planes, out_planes, kernel_size=4, stride=2, padding=1),
nn.ReLU(inplace=True)
)
# define the LightningModule
class lRefineNet(pl.LightningModule):
def __init__(self, refinenet):
super().__init__()
self.model = refinenet
self.rn_metrics = Refinenet_Metrics()
def forward(self, x):
return self.model(x)
def infer_patches(self, patches: torch.Tensor,
keypoints: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
return self.model.infer_patches(patches, keypoints)
def validation_step(self, batch, batch_idx):
x, loc = batch
x = torch.stack(x, dim=0)
loc = torch.stack(loc, dim=0)
x = x.view(-1, *x.size()[-3:])
loc = loc.view(-1, *loc.size()[-3:])
loc_hat = self.model(x)
loss_loc = nn.functional.mse_loss(loc_hat, loc)
dist = self.rn_metrics(loc_hat, loc)
self.log("val_dist_refinenet_pixels", dist)
self.log("val_refinenet_loss", loss_loc)
return loss_loc
def training_step(self, batch, batch_idx):
x, loc = batch
x = torch.stack(x, dim=0)
loc = torch.stack(loc, dim=0)
x = x.view(-1, *x.size()[-3:])
loc = loc.view(-1, *loc.size()[-3:])
loc_hat = self.model(x)
loss_loc = nn.functional.mse_loss(loc_hat, loc)
self.log("train_refinenet_loss", loss_loc)
return loss_loc
def configure_optimizers(self):
optimizer = optim.Adam(self.parameters(), lr=1e-4)
return optimizer
if __name__ == '__main__':
model = RefineNet()
# from torchinfo import summary
from torchinfo import summary
summary(model, input_size=(1, 1, 24, 24))
# print(model.infer_patches(np.random.randn(16, 24, 24).astype(np.float32)))