File size: 5,286 Bytes
8395863 |
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 |
import os
import torch
import numpy as np
import pandas as pd
from torch.utils.data import Dataset, DataLoader
from skimage import io, transform
from PIL import Image
import torch.nn as nn
from torchvision import transforms, utils, models
import torch.nn.functional as F
import utils.resnet as resnet
from utils.TransformerEncoder import Encoder
cfg1 = {
"hidden_size" : 768,
"mlp_dim" : 768*4,
"num_heads" : 12,
"num_layers" : 2,
"attention_dropout_rate" : 0,
"dropout_rate" : 0.0,
}
cfg2 = {
"hidden_size" : 768,
"mlp_dim" : 768*4,
"num_heads" : 12,
"num_layers" : 2,
"attention_dropout_rate" : 0,
"dropout_rate" : 0.0,
}
cfg3 = {
"hidden_size" : 512,
"mlp_dim" : 512*4,
"num_heads" : 8,
"num_layers" : 2,
"attention_dropout_rate" : 0,
"dropout_rate" : 0.0,
}
class TranSalNet(nn.Module):
def __init__(self):
super(TranSalNet, self).__init__()
self.encoder = _Encoder()
self.decoder = _Decoder()
def forward(self, x):
x = self.encoder(x)
x = self.decoder(x)
return x
class _Encoder(nn.Module):
def __init__(self):
super(_Encoder, self).__init__()
base_model = resnet.resnet50(pretrained=True)
base_layers = list(base_model.children())[:8]
self.encoder = nn.ModuleList(base_layers).eval()
def forward(self, x):
outputs = []
for ii,layer in enumerate(self.encoder):
x = layer(x)
if ii in {5,6,7}:
outputs.append(x)
return outputs
class _Decoder(nn.Module):
def __init__(self):
super(_Decoder, self).__init__()
self.conv1 = nn.Conv2d(768, 768, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.conv2 = nn.Conv2d(768, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.conv3 = nn.Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.conv4 = nn.Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.conv5 = nn.Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.conv6 = nn.Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.conv7 = nn.Conv2d(32, 1, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.batchnorm1 = nn.BatchNorm2d(768, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.batchnorm2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.batchnorm3 = nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.batchnorm4 = nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.batchnorm5 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.batchnorm6 = nn.BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.TransEncoder1 = TransEncoder(in_channels=2048, spatial_size=9*12, cfg=cfg1)
self.TransEncoder2 = TransEncoder(in_channels=1024, spatial_size=18*24, cfg=cfg2)
self.TransEncoder3 = TransEncoder(in_channels=512, spatial_size=36*48, cfg=cfg3)
self.add = torch.add
self.relu = nn.ReLU(True)
self.upsample = nn.Upsample(scale_factor=2, mode='nearest')
self.sigmoid = nn.Sigmoid()
def forward(self, x):
x3, x4, x5 = x
x5 = self.TransEncoder1(x5)
x5 = self.conv1(x5)
x5 = self.batchnorm1(x5)
x5 = self.relu(x5)
x5 = self.upsample(x5)
x4_a = self.TransEncoder2(x4)
x4 = x5 * x4_a
x4 = self.relu(x4)
x4 = self.conv2(x4)
x4 = self.batchnorm2(x4)
x4 = self.relu(x4)
x4 = self.upsample(x4)
x3_a = self.TransEncoder3(x3)
x3 = x4 * x3_a
x3 = self.relu(x3)
x3 = self.conv3(x3)
x3 = self.batchnorm3(x3)
x3 = self.relu(x3)
x3 = self.upsample(x3)
x2 = self.conv4(x3)
x2 = self.batchnorm4(x2)
x2 = self.relu(x2)
x2 = self.upsample(x2)
x2 = self.conv5(x2)
x2 = self.batchnorm5(x2)
x2 = self.relu(x2)
x1 = self.upsample(x2)
x1 = self.conv6(x1)
x1 = self.batchnorm6(x1)
x1 = self.relu(x1)
x1 = self.conv7(x1)
x = self.sigmoid(x1)
return x
class TransEncoder(nn.Module):
def __init__(self, in_channels, spatial_size, cfg):
super(TransEncoder, self).__init__()
self.patch_embeddings = nn.Conv2d(in_channels=in_channels,
out_channels=cfg['hidden_size'],
kernel_size=1,
stride=1)
self.position_embeddings = nn.Parameter(torch.zeros(1, spatial_size, cfg['hidden_size']))
self.transformer_encoder = Encoder(cfg)
def forward(self, x):
a, b = x.shape[2], x.shape[3]
x = self.patch_embeddings(x)
x = x.flatten(2)
x = x.transpose(-1, -2)
embeddings = x + self.position_embeddings
x = self.transformer_encoder(embeddings)
B, n_patch, hidden = x.shape
x = x.permute(0, 2, 1)
x = x.contiguous().view(B, hidden, a, b)
return x
|