File size: 5,912 Bytes
9dce458 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models import resnet101
import DBHead
import einops
class ImageMultiheadSelfAttention(nn.Module):
def __init__(self, planes):
super(ImageMultiheadSelfAttention, self).__init__()
self.attn = nn.MultiheadAttention(planes, 8)
def forward(self, x):
res = x
n, c, h, w = x.shape
x = einops.rearrange(x, 'n c h w -> (h w) n c')
x = self.attn(x, x, x)[0]
x = einops.rearrange(x, '(h w) n c -> n c h w', n = n, c = c, h = h, w = w)
return res + x
class double_conv(nn.Module):
def __init__(self, in_ch, mid_ch, out_ch, stride = 1, planes = 256):
super(double_conv, self).__init__()
self.planes = planes
# down = None
# if stride > 1:
# down = nn.Sequential(
# nn.AvgPool2d(2, 2),
# nn.Conv2d(in_ch + mid_ch, self.planes * Bottleneck.expansion, kernel_size=1, stride=1, bias=False),nn.BatchNorm2d(self.planes * Bottleneck.expansion)
# )
self.down = None
if stride > 1:
self.down = nn.AvgPool2d(2,stride=2)
self.conv = nn.Sequential(
nn.Conv2d(in_ch + mid_ch, mid_ch, kernel_size=3, padding=1, stride = 1, bias=False),
nn.BatchNorm2d(mid_ch),
nn.ReLU(inplace=True),
#Bottleneck(mid_ch, self.planes, stride, down, 2, 1, avd = True, norm_layer = nn.BatchNorm2d),
nn.Conv2d(mid_ch, out_ch, kernel_size=3, stride = 1, padding=1, bias=False),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True),
)
def forward(self, x):
if self.down is not None:
x = self.down(x)
x = self.conv(x)
return x
class double_conv_up(nn.Module):
def __init__(self, in_ch, mid_ch, out_ch, stride = 1, planes = 256):
super(double_conv_up, self).__init__()
self.planes = planes
# down = None
# if stride > 1:
# down = nn.Sequential(
# nn.AvgPool2d(2, 2),
# nn.Conv2d(in_ch + mid_ch, self.planes * Bottleneck.expansion, kernel_size=1, stride=1, bias=False),nn.BatchNorm2d(self.planes * Bottleneck.expansion)
# )
self.down = None
if stride > 1:
self.down = nn.AvgPool2d(2,stride=2)
self.conv = nn.Sequential(
nn.Conv2d(in_ch + mid_ch, mid_ch, kernel_size=3, padding=1, stride = 1, bias=False),
nn.BatchNorm2d(mid_ch),
nn.ReLU(inplace=True),
#Bottleneck(mid_ch, self.planes, stride, down, 2, 1, avd = True, norm_layer = nn.BatchNorm2d),
nn.Conv2d(mid_ch, mid_ch, kernel_size=3, stride = 1, padding=1, bias=False),
nn.BatchNorm2d(mid_ch),
nn.ReLU(inplace=True),
nn.ConvTranspose2d(mid_ch, out_ch, kernel_size=4, stride = 2, padding=1, bias=False),
nn.BatchNorm2d(out_ch),
nn.ReLU(inplace=True),
)
def forward(self, x):
if self.down is not None:
x = self.down(x)
x = self.conv(x)
return x
class TextDetection(nn.Module):
def __init__(self, pretrained=None):
super(TextDetection, self).__init__()
self.backbone = resnet101(pretrained=True if pretrained else False)
self.conv_db = DBHead.DBHead(64, 0)
self.conv_mask = nn.Sequential(
nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True),
nn.Conv2d(64, 64, kernel_size=3, padding=1), nn.ReLU(inplace=True),
nn.Conv2d(64, 32, kernel_size=3, padding=1), nn.ReLU(inplace=True),
nn.Conv2d(32, 1, kernel_size=1),
nn.Sigmoid()
)
self.down_conv1 = double_conv(0, 512, 512, 2)
self.down_conv2 = double_conv(0, 512, 512, 2)
self.down_conv3 = double_conv(0, 512, 512, 2)
self.upconv1 = double_conv_up(0, 512, 256)
self.upconv2 = double_conv_up(256, 512, 256)
self.upconv3 = double_conv_up(256, 512, 256)
self.upconv4 = double_conv_up(256, 512, 256, planes = 128)
self.upconv5 = double_conv_up(256, 256, 128, planes = 64)
self.upconv6 = double_conv_up(128, 128, 64, planes = 32)
self.upconv7 = double_conv_up(64, 64, 64, planes = 16)
self.proj_h4 = nn.Conv2d(64 * 4, 64, 1)
self.proj_h8 = nn.Conv2d(128 * 4, 128, 1)
self.proj_h16 = nn.Conv2d(256 * 4, 256, 1)
self.proj_h32 = nn.Conv2d(512 * 4, 512, 1)
def forward(self, x):
x = self.backbone.conv1(x)
x = self.backbone.bn1(x)
x = self.backbone.relu(x)
x = self.backbone.maxpool(x) # 64@384
h4 = self.backbone.layer1(x) # 64@384
h8 = self.backbone.layer2(h4) # 128@192
h16 = self.backbone.layer3(h8) # 256@96
h32 = self.backbone.layer4(h16) # 512@48
h4 = self.proj_h4(h4)
h8 = self.proj_h8(h8)
h16 = self.proj_h16(h16)
h32 = self.proj_h32(h32)
h64 = self.down_conv1(h32) # 512@24
h128 = self.down_conv2(h64) # 512@12
h256 = self.down_conv3(h128) # 512@6
up256 = self.upconv1(h256) # 128@12
up128 = self.upconv2(torch.cat([up256, h128], dim = 1)) # 64@24
up64 = self.upconv3(torch.cat([up128, h64], dim = 1)) # 128@48
up32 = self.upconv4(torch.cat([up64, h32], dim = 1)) # 64@96
up16 = self.upconv5(torch.cat([up32, h16], dim = 1)) # 128@192
up8 = self.upconv6(torch.cat([up16, h8], dim = 1)) # 64@384
up4 = self.upconv7(torch.cat([up8, h4], dim = 1)) # 64@768
return self.conv_db(up8), self.conv_mask(up4)
if __name__ == '__main__':
device = torch.device("cuda:0")
net = TextDetection().to(device)
img = torch.randn(2, 3, 1024, 1024).to(device)
db, seg = net(img)
print(db.shape)
print(seg.shape)
|