| import torch
|
| import torch.nn as nn
|
| import torch.nn.functional as F
|
|
|
| from lib.models.backbones.backbone_selector import BackboneSelector
|
| from lib.models.tools.module_helper import ModuleHelper
|
|
|
|
|
| class ISANet(nn.Module):
|
| """
|
| Interlaced Sparse Self-Attention for Semantic Segmentation
|
| """
|
| def __init__(self, configer):
|
| self.inplanes = 128
|
| super(ISANet, self).__init__()
|
| self.configer = configer
|
| self.num_classes = self.configer.get('data', 'num_classes')
|
| self.backbone = BackboneSelector(configer).get_backbone()
|
|
|
|
|
| bn_type = self.configer.get('network', 'bn_type')
|
| factors = self.configer.get('network', 'factors')
|
| from lib.models.modules.isa_block import ISA_Module
|
| self.isa_head = nn.Sequential(
|
| nn.Conv2d(2048, 512, kernel_size=3, stride=1, padding=1, bias=False),
|
| ModuleHelper.BNReLU(512, bn_type=bn_type),
|
| ISA_Module(in_channels=512, key_channels=256, value_channels=512,
|
| out_channels=512, down_factors=factors, dropout=0.05, bn_type=bn_type),
|
| )
|
| self.cls_head = nn.Conv2d(512, self.num_classes, kernel_size=1, stride=1, padding=0, bias=True)
|
| self.dsn_head = nn.Sequential(
|
| nn.Conv2d(1024, 512, kernel_size=3, stride=1, padding=1, bias=False),
|
| ModuleHelper.BNReLU(512, bn_type=bn_type),
|
| nn.Dropout2d(0.05),
|
| nn.Conv2d(512, self.num_classes, kernel_size=1, stride=1, padding=0, bias=True),
|
| )
|
|
|
| def forward(self, x_):
|
| x = self.backbone(x_)
|
| x_dsn = self.dsn_head(x[-2])
|
| x = self.isa_head(x[-1])
|
| x = self.cls_head(x)
|
| x_dsn = F.interpolate(x_dsn, size=(x_.size(2), x_.size(3)), mode="bilinear", align_corners=True)
|
| x = F.interpolate(x, size=(x_.size(2), x_.size(3)), mode="bilinear", align_corners=True)
|
| return x_dsn, x |
