import torch import torch.nn as nn import torch.nn.functional as F from einops import rearrange, repeat from timm.models.layers import DropPath, to_2tuple, trunc_normal_ import timm import torch import torch.nn as nn import torch.nn.functional as F from einops import rearrange from torch.nn.init import trunc_normal_ class ConvBNReLU(nn.Sequential): def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, norm_layer=nn.BatchNorm2d, bias=False): super(ConvBNReLU, self).__init__( nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias, dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2), norm_layer(out_channels), nn.ReLU6() ) class ConvBN(nn.Sequential): def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, norm_layer=nn.BatchNorm2d, bias=False): super(ConvBN, self).__init__( nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias, dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2), norm_layer(out_channels) ) class Conv(nn.Sequential): def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, bias=False): super(Conv, self).__init__( nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias, dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2) ) class SelfAttention(nn.Module): def __init__(self, dim, num_heads): super(SelfAttention, self).__init__() self.num_heads = num_heads head_dim = dim // self.num_heads self.scale = head_dim ** -0.5 self.qkv = nn.Linear(dim, 3 * dim) self.o_proj = nn.Linear(dim, dim) def forward(self, x): B, C, H, W = x.shape qkv = self.qkv(x).view(B, -1, self.num_heads, 3, H * W).permute(3, 0, 2, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] dots = torch.matmul(q.transpose(-2, -1), k) * self.scale attn = dots.softmax(dim=-1) out = torch.matmul(attn, v).transpose(1, 2).reshape(B, C, H, W) return self.o_proj(out) class SeparableConvBNReLU(nn.Sequential): def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1, norm_layer=nn.BatchNorm2d): super(SeparableConvBNReLU, self).__init__( nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2, groups=in_channels, bias=False), norm_layer(out_channels), nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False), nn.ReLU6() ) class SeparableConvBN(nn.Sequential): def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1, norm_layer=nn.BatchNorm2d): super(SeparableConvBN, self).__init__( nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2, groups=in_channels, bias=False), norm_layer(out_channels), nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) ) class SeparableConv(nn.Sequential): def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1): super(SeparableConv, self).__init__( nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2, groups=in_channels, bias=False), nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) ) class SEBlock(nn.Module): def __init__(self, in_channels, reduction=16): super(SEBlock, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(in_channels, in_channels // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(in_channels // reduction, in_channels, bias=False), nn.Sigmoid() ) def forward(self, x): b, c, _, _ = x.size() y = self.avg_pool(x).view(b, c) y = self.fc(y).view(b, c, 1, 1) return x * y.expand_as(x) class ImprovedLocalAttention(nn.Module): def __init__(self, dim): super(ImprovedLocalAttention, self).__init__() self.conv1x1 = nn.Conv2d(dim, dim, kernel_size=1, bias=False) self.conv3x3 = nn.Conv2d(dim, dim, kernel_size=3, padding=1, bias=False) self.conv5x5 = nn.Conv2d(dim, dim, kernel_size=5, padding=2, bias=False) self.bn = nn.BatchNorm2d(dim) self.se = SEBlock(dim) def forward(self, x): # Applying different convolutions and combining results out1 = self.conv1x1(x) out2 = self.conv3x3(x) out3 = self.conv5x5(x) out = out1 + out2 + out3 out = self.bn(out) out = self.se(out) out = out + out1 + out3 return out # class MultiHeadGlobalAttention(nn.Module): # def __init__(self, # dim=256, # num_heads=16, # qkv_bias=False, # window_size=8, # relative_pos_embedding=True): # super().__init__() # self.num_heads = num_heads # head_dim = dim // self.num_heads # self.scale = head_dim ** -0.5 # self.ws = window_size # # self.qkv = nn.Conv2d(dim, 3 * dim, kernel_size=1, bias=qkv_bias) # self.proj = nn.Conv2d(dim, dim, kernel_size=1, bias=False) # # self.relative_pos_embedding = relative_pos_embedding # # if self.relative_pos_embedding: # self.relative_position_bias_table = nn.Parameter( # torch.zeros((2 * window_size - 1) * (2 * window_size - 1), num_heads)) # # coords_h = torch.arange(self.ws) # coords_w = torch.arange(self.ws) # coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # coords_flatten = torch.flatten(coords, 1) # relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # relative_coords = relative_coords.permute(1, 2, 0).contiguous() # relative_coords[:, :, 0] += self.ws - 1 # relative_coords[:, :, 1] += self.ws - 1 # relative_coords[:, :, 0] *= 2 * self.ws - 1 # relative_position_index = relative_coords.sum(-1) # self.register_buffer("relative_position_index", relative_position_index) # # nn.init.trunc_normal_(self.relative_position_bias_table, std=.02) # # def pad(self, x, ps): # _, _, H, W = x.size() # if W % ps != 0: # x = F.pad(x, (0, ps - W % ps), mode='reflect') # if H % ps != 0: # x = F.pad(x, (0, 0, 0, ps - H % ps), mode='reflect') # return x # # def forward(self, x): # B, C, H, W = x.shape # # x = self.pad(x, self.ws) # B, C, Hp, Wp = x.shape # qkv = self.qkv(x) # # q, k, v = rearrange(qkv, 'b (qkv h d) (hh ws1) (ww ws2) -> qkv (b hh ww) h (ws1 ws2) d', h=self.num_heads, # d=C // self.num_heads, hh=Hp // self.ws, ww=Wp // self.ws, qkv=3, ws1=self.ws, ws2=self.ws) # # dots = (q @ k.transpose(-2, -1)) * self.scale # # if self.relative_pos_embedding: # relative_position_bias = self.relative_position_bias_table[ # self.relative_position_index.view(-1)].view( # self.ws * self.ws, self.ws * self.ws, -1) # relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # dots += relative_position_bias.unsqueeze(0) # # attn = dots.softmax(dim=-1) # attn = attn @ v # # attn = rearrange(attn, '(b hh ww) h (ws1 ws2) d -> b (h d) (hh ws1) (ww ws2)', h=self.num_heads, # d=C // self.num_heads, hh=Hp // self.ws, ww=Wp // self.ws, ws1=self.ws, ws2=self.ws) # # attn = attn[:, :, :H, :W] # # out = self.proj(attn) # out = out[:, :, :H, :W] # # return out # class EffcirntMutilSelfAttention(nn.Module): # def __init__(self, # dim, # num_heads=8, # sr_ratio=1): # super().__init__() # self.num_heads = num_heads # head_dim = dim // num_heads # self.scale = head_dim ** -0.5 # self.dim = dim # # self.q = nn.Conv2d(dim, dim, kernel_size=1, bias=True) # self.kv = nn.Conv2d(dim, dim * 2, kernel_size=1, bias=True) # # self.sr_ratio = sr_ratio # if sr_ratio > 1: # self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio + 1, stride=sr_ratio, padding=sr_ratio // 2, groups=dim) # self.sr_norm = nn.LayerNorm(dim, eps=1e-6) # # self.up = nn.Sequential( # nn.Conv2d(dim, sr_ratio * sr_ratio * dim, kernel_size=3, stride=1, padding=1, groups=dim), # nn.PixelShuffle(upscale_factor=sr_ratio) # ) # self.up_norm = nn.LayerNorm(dim, eps=1e-6) # # self.proj = nn.Conv2d(dim, dim, kernel_size=1) # # def forward(self, x): # B, C, H, W = x.shape # N = H * W # # q = self.q(x).reshape(B, self.num_heads, C // self.num_heads, N).permute(0, 1, 3, 2) # # if self.sr_ratio > 1: # x = self.sr(x).reshape(B, C, -1).permute(0, 2, 1) # x = self.sr_norm(x) # x = x.permute(0, 2, 1).reshape(B, C, H // self.sr_ratio, W // self.sr_ratio) # else: # x = x.reshape(B, C, N).permute(0, 2, 1) # # kv = self.kv(x).reshape(B, 2, self.num_heads, C // self.num_heads, -1).permute(1, 0, 2, 4, 3) # k, v = kv[0], kv[1] # # attn = (q @ k.transpose(-2, -1)) * self.scale # attn = attn.softmax(dim=-1) # # x = (attn @ v).transpose(1, 2).reshape(B, C, H, W) # # identity = v.transpose(-1, -2).reshape(B, C, H // self.sr_ratio, W // self.sr_ratio) # identity = self.up(identity) # identity = identity.flatten(2).transpose(1, 2).reshape(B, C, H, W) # x = self.proj(x + identity) # return x class EffcirntMutilSelfAttention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1, apply_transform=False): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.q = nn.Conv2d(dim, dim, kernel_size=1, bias=qkv_bias) self.kv = nn.Conv2d(dim, dim * 2, kernel_size=1, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Conv2d(dim, dim, kernel_size=1) self.proj_drop = nn.Dropout(proj_drop) self.sr_ratio = sr_ratio if sr_ratio > 1: self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio + 1, stride=sr_ratio, padding=sr_ratio // 2, groups=dim) self.sr_norm = nn.LayerNorm(dim) self.apply_transform = apply_transform and num_heads > 1 if self.apply_transform: self.transform_conv = nn.Conv2d(self.num_heads, self.num_heads, kernel_size=1, stride=1) self.transform_norm = nn.InstanceNorm2d(self.num_heads) def forward(self, x): B, C, H, W = x.shape q = self.q(x).reshape(B, self.num_heads, C // self.num_heads, H * W).permute(0, 1, 3, 2) if self.sr_ratio > 1: x_ = self.sr(x).reshape(B, C, -1).permute(0, 2, 1) x_ = self.sr_norm(x_) kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) else: kv = self.kv(x).reshape(B, 2, self.num_heads, C // self.num_heads, H * W).permute(1, 0, 2, 4, 3) k, v = kv[0], kv[1] attn = (q @ k.transpose(-2, -1)) * self.scale if self.apply_transform: attn = self.transform_conv(attn) attn = attn.softmax(dim=-1) attn = self.transform_norm(attn) else: attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(2, 3).reshape(B, self.num_heads * (C // self.num_heads), H, W) x = self.proj(x) x = self.proj_drop(x) return x class GlobalLocalAttention(nn.Module): def __init__(self, dim=256, num_heads=16, qkv_bias=False, window_size=8, relative_pos_embedding=True): super(GlobalLocalAttention, self).__init__() self.local1 = ImprovedLocalAttention(dim) # self.local2 = ImprovedLocalAttention(dim) self.global_attention = EffcirntMutilSelfAttention(dim, num_heads, qkv_bias, window_size, relative_pos_embedding) def forward(self, x): # Combining local and global attention local = self.local1(x) global_attn = self.global_attention(x) return local + global_attn # Testing the model with a random tensor gl_attention = GlobalLocalAttention(dim=256, num_heads=16, qkv_bias=False, window_size=8, relative_pos_embedding=True) x = torch.randn(1, 256, 64, 64) output = gl_attention(x) print(output.shape) # Output should be (1, 256, 64, 64)