File size: 10,584 Bytes
feaac02 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 |
import kornia.filters
import scipy.ndimage
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import random
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=dilation, groups=groups, bias=False, dilation=dilation)
def conv1x1(in_planes, out_planes, stride=1):
"""1x1 convolution"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
class DoubleConv(nn.Module):
"""(convolution => [BN] => ReLU) * 2"""
def __init__(self, in_channels, out_channels, mid_channels=None):
super().__init__()
if not mid_channels:
mid_channels = out_channels
norm_layer = nn.BatchNorm2d
self.conv1 = nn.Conv2d(in_channels, mid_channels, kernel_size=3, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(mid_channels)
self.inst1 = nn.InstanceNorm2d(mid_channels)
# self.gn1 = nn.GroupNorm(4, mid_channels)
self.relu = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(mid_channels, out_channels, kernel_size=3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_channels)
self.inst2 = nn.InstanceNorm2d(out_channels)
# self.gn2 = nn.GroupNorm(4, out_channels)
self.downsample = None
if in_channels != out_channels:
self.downsample = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False),
nn.BatchNorm2d(out_channels),
)
def forward(self, x):
identity = x
out = self.conv1(x)
# out = self.bn1(out)
out = self.inst1(out)
# out = self.gn1(out)
out = self.relu(out)
out = self.conv2(out)
# out = self.bn2(out)
out = self.inst2(out)
# out = self.gn2(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class Down(nn.Module):
"""Downscaling with maxpool then double conv"""
def __init__(self, in_channels, out_channels):
super().__init__()
self.maxpool_conv = nn.Sequential(
nn.MaxPool2d(2),
DoubleConv(in_channels, out_channels)
)
def forward(self, x):
return self.maxpool_conv(x)
class Up(nn.Module):
"""Upscaling then double conv"""
def __init__(self, in_channels, out_channels, bilinear=True):
super().__init__()
# if bilinear, use the normal convolutions to reduce the number of channels
if bilinear:
self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.conv = DoubleConv(in_channels, out_channels, in_channels // 2)
else:
if in_channels == out_channels:
self.up = nn.Identity()
else:
self.up = nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size=2, stride=2)
self.conv = DoubleConv(in_channels, out_channels)
def forward(self, x1, x2):
x1 = self.up(x1)
# input is CHW
diffY = x2.size()[2] - x1.size()[2]
diffX = x2.size()[3] - x1.size()[3]
x1 = F.pad(x1, [diffX // 2, diffX - diffX // 2,
diffY // 2, diffY - diffY // 2])
# if you have padding issues, see
# https://github.com/HaiyongJiang/U-Net-Pytorch-Unstructured-Buggy/commit/0e854509c2cea854e247a9c615f175f76fbb2e3a
# https://github.com/xiaopeng-liao/Pytorch-UNet/commit/8ebac70e633bac59fc22bb5195e513d5832fb3bd
x = torch.cat([x2, x1], dim=1)
return self.conv(x)
class OutConv(nn.Module):
def __init__(self, in_channels, out_channels):
super(OutConv, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)
def forward(self, x):
return self.conv(x)
class GaussianLayer(nn.Module):
def __init__(self):
super(GaussianLayer, self).__init__()
self.seq = nn.Sequential(
# nn.ReflectionPad2d(10),
nn.Conv2d(1, 1, 5, stride=1, padding=2, bias=False)
)
self.weights_init()
def forward(self, x):
return self.seq(x)
def weights_init(self):
n= np.zeros((5,5))
n[3,3] = 1
k = scipy.ndimage.gaussian_filter(n,sigma=1)
for name, f in self.named_parameters():
f.data.copy_(torch.from_numpy(k))
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=dilation, groups=groups, bias=False, dilation=dilation)
class Decoder(nn.Module):
def __init__(self):
super(Decoder, self).__init__()
self.up1 = Up(2048, 1024 // 1, False)
self.up2 = Up(1024, 512 // 1, False)
self.up3 = Up(512, 256 // 1, False)
self.conv2d_2_1 = conv3x3(256, 128)
self.gn1 = nn.GroupNorm(4, 128)
self.instance1 = nn.InstanceNorm2d(128)
self.up4 = Up(128, 64 // 1, False)
self.upsample4 = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
# self.upsample4 = nn.ConvTranspose2d(64, 64, 2, stride=2)
self.upsample4_conv = DoubleConv(64, 64, 64 // 2)
self.up_ = Up(128, 128 // 1, False)
self.conv2d_2_2 = conv3x3(128, 6)
self.instance2 = nn.InstanceNorm2d(6)
self.gn2 = nn.GroupNorm(3, 6)
self.gaussian_blur = GaussianLayer()
self.up5 = Up(6, 3, False)
self.conv2d_2_3 = conv3x3(3, 1)
self.instance3 = nn.InstanceNorm2d(1)
self.gaussian_blur = GaussianLayer()
self.kernel = nn.Parameter(torch.tensor(
[[[0.0, 0.0, 0.0], [0.0, 1.0, random.uniform(-1.0, 0.0)], [0.0, 0.0, 0.0]],
[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, random.uniform(-1.0, 0.0)]],
[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, random.uniform(random.uniform(-1.0, 0.0), -0.0), 0.0]],
[[0.0, 0.0, 0.0], [0.0, 1.0, 0.0], [random.uniform(-1.0, 0.0), 0.0, 0.0]],
[[0.0, 0.0, 0.0], [random.uniform(-1.0, 0.0), 1.0, 0.0], [0.0, 0.0, 0.0]],
[[random.uniform(-1.0, 0.0), 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]],
[[0.0, random.uniform(-1.0, 0.0), 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]],
[[0.0, 0.0, random.uniform(-1.0, 0.0)], [0.0, 1.0, 0.0], [0.0, 0.0, 0.0]], ],
).unsqueeze(1))
self.nms_conv = nn.Conv2d(1, 1, kernel_size=3, stride=1, padding=1, bias=False, groups=1)
with torch.no_grad():
self.nms_conv.weight = self.kernel.float()
class Resnet_with_skip(nn.Module):
def __init__(self, model):
super(Resnet_with_skip, self).__init__()
self.model = model
self.decoder = Decoder()
def forward_pred(self, image):
pred_net = self.model(image)
return pred_net
def forward_decode(self, image):
identity = image
image = self.model.conv1(image)
image = self.model.bn1(image)
image = self.model.relu(image)
image1 = self.model.maxpool(image)
image2 = self.model.layer1(image1)
image3 = self.model.layer2(image2)
image4 = self.model.layer3(image3)
image5 = self.model.layer4(image4)
reconst1 = self.decoder.up1(image5, image4)
reconst2 = self.decoder.up2(reconst1, image3)
reconst3 = self.decoder.up3(reconst2, image2)
reconst = self.decoder.conv2d_2_1(reconst3)
# reconst = self.decoder.instance1(reconst)
reconst = self.decoder.gn1(reconst)
reconst = F.relu(reconst)
reconst4 = self.decoder.up4(reconst, image1)
# reconst5 = self.decoder.upsample4(reconst4)
reconst5 = self.decoder.upsample4(reconst4)
# reconst5 = self.decoder.upsample4_conv(reconst4)
reconst5 = self.decoder.up_(reconst5, image)
# reconst5 = reconst5 + image
reconst5 = self.decoder.conv2d_2_2(reconst5)
reconst5 = self.decoder.instance2(reconst5)
# reconst5 = self.decoder.gn2(reconst5)
reconst5 = F.relu(reconst5)
reconst = self.decoder.up5(reconst5, identity)
reconst = self.decoder.conv2d_2_3(reconst)
# reconst = self.decoder.instance3(reconst)
reconst = F.relu(reconst)
# return reconst
blurred = self.decoder.gaussian_blur(reconst)
gradients = kornia.filters.spatial_gradient(blurred, normalized=False)
# Unpack the edges
gx = gradients[:, :, 0]
gy = gradients[:, :, 1]
angle = torch.atan2(gy, gx)
# Radians to Degrees
import math
angle = 180.0 * angle / math.pi
# Round angle to the nearest 45 degree
angle = torch.round(angle / 45) * 45
nms_magnitude = self.decoder.nms_conv(blurred)
# nms_magnitude = F.conv2d(blurred, kernel.unsqueeze(1), padding=kernel.shape[-1]//2)
# Non-maximal suppression
# Get the indices for both directions
positive_idx = (angle / 45) % 8
positive_idx = positive_idx.long()
negative_idx = ((angle / 45) + 4) % 8
negative_idx = negative_idx.long()
# Apply the non-maximum suppression to the different directions
channel_select_filtered_positive = torch.gather(nms_magnitude, 1, positive_idx)
channel_select_filtered_negative = torch.gather(nms_magnitude, 1, negative_idx)
channel_select_filtered = torch.stack(
[channel_select_filtered_positive, channel_select_filtered_negative], 1
)
# is_max = channel_select_filtered.min(dim=1)[0] > 0.0
# magnitude = reconst * is_max
thresh = nn.Threshold(0.01, 0.01)
max_matrix = channel_select_filtered.min(dim=1)[0]
max_matrix = thresh(max_matrix)
magnitude = torch.mul(reconst, max_matrix)
# magnitude = torchvision.transforms.functional.invert(magnitude)
# magnitude = self.decoder.sharpen(magnitude)
# magnitude = self.decoder.threshold(magnitude)
magnitude = kornia.enhance.adjust_gamma(magnitude, 2.0)
# magnitude = F.leaky_relu(magnitude)
return magnitude
def forward(self, image):
reconst = self.forward_decode(image)
pred = self.forward_pred(image)
return pred, reconst
|