313 lines
12 KiB
Python
313 lines
12 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from .swin import window_partition, window_reverse, WindowAttention
|
|
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
|
|
|
|
|
|
class Mlp(nn.Module):
|
|
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
|
|
super().__init__()
|
|
out_features = out_features or in_features
|
|
hidden_features = hidden_features or in_features
|
|
self.fc1 = nn.Linear(in_features, hidden_features)
|
|
self.act = act_layer()
|
|
self.fc2 = nn.Linear(hidden_features, out_features)
|
|
self.drop = nn.Dropout(drop)
|
|
|
|
def forward(self, x):
|
|
x = self.fc1(x)
|
|
x = self.act(x)
|
|
x = self.drop(x)
|
|
x = self.fc2(x)
|
|
x = self.drop(x)
|
|
return x
|
|
|
|
def flops(self, N):
|
|
return N * (self.in_features + self.out_features) * self.hidden_features
|
|
|
|
|
|
class Attention(nn.Module):
|
|
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.num_heads = num_heads
|
|
head_dim = dim // num_heads
|
|
|
|
self.scale = qk_scale or head_dim ** -0.5
|
|
|
|
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
|
self.attn_drop = nn.Dropout(attn_drop)
|
|
self.proj = nn.Linear(dim, dim)
|
|
self.proj_drop = nn.Dropout(proj_drop)
|
|
|
|
def forward(self, x):
|
|
B, N, C = x.shape
|
|
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
|
q, k, v = qkv[0], qkv[1], qkv[2]
|
|
|
|
attn = (q @ k.transpose(-2, -1)) * self.scale
|
|
attn = attn.softmax(dim=-1)
|
|
attn = self.attn_drop(attn)
|
|
|
|
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
|
|
x = self.proj(x)
|
|
x = self.proj_drop(x)
|
|
return x
|
|
|
|
def flops(self, N):
|
|
flops = 0
|
|
# q
|
|
flops += N * self.dim * self.dim * 3
|
|
# qk
|
|
flops += self.num_heads * N * self.dim // self.num_heads * N
|
|
# att v
|
|
flops += self.num_heads * N * self.dim // self.num_heads * N
|
|
# proj
|
|
flops += N * self.dim * self.dim
|
|
return flops
|
|
|
|
|
|
class CrossAttention(nn.Module):
|
|
def __init__(self, dim1, dim2, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
|
|
super().__init__()
|
|
self.num_heads = num_heads
|
|
head_dim = dim // num_heads
|
|
self.dim = dim
|
|
self.dim1 = dim1
|
|
self.dim2 = dim2
|
|
self.scale = qk_scale or head_dim ** -0.5
|
|
|
|
self.q1 = nn.Linear(dim1, dim, bias=qkv_bias)
|
|
self.proj = nn.Linear(dim, dim1)
|
|
|
|
self.k2 = nn.Linear(dim2, dim, bias=qkv_bias)
|
|
self.v2 = nn.Linear(dim2, dim, bias=qkv_bias)
|
|
|
|
self.attn_drop = nn.Dropout(attn_drop)
|
|
self.proj_drop = nn.Dropout(proj_drop)
|
|
|
|
def forward(self, fea, depth_fea):
|
|
_, N1, _ = fea.shape
|
|
B, N, _ = depth_fea.shape
|
|
C = self.dim
|
|
q1 = self.q1(fea).reshape(B, N1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
|
# q [B, nhead, N, C//nhead]
|
|
|
|
k2 = self.k2(depth_fea).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
v2 = self.v2(depth_fea).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
|
attn = (q1 @ k2.transpose(-2, -1)) * self.scale
|
|
attn = attn.softmax(dim=-1)
|
|
attn = self.attn_drop(attn)
|
|
|
|
fea = (attn @ v2).transpose(1, 2).reshape(B, N1, C)
|
|
fea = self.proj(fea)
|
|
fea = self.proj_drop(fea)
|
|
|
|
return fea
|
|
|
|
def flops(self, N1, N2):
|
|
flops = 0
|
|
# q
|
|
flops += N1 * self.dim1 * self.dim
|
|
# kv
|
|
flops += N2 * self.dim2 * self.dim * 2
|
|
# qk
|
|
flops += self.num_heads * N1 * self.dim // self.num_heads * N2
|
|
# att v
|
|
flops += self.num_heads * N1 * self.dim // self.num_heads * N2
|
|
# proj
|
|
flops += N1 * self.dim * self.dim1
|
|
return flops
|
|
|
|
|
|
class Block(nn.Module):
|
|
# Remove FFN
|
|
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
|
|
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
|
|
super().__init__()
|
|
self.norm1 = norm_layer(dim)
|
|
self.dim = dim
|
|
self.attn = Attention(
|
|
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
|
|
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
|
# self.norm2 = norm_layer(dim)
|
|
# mlp_hidden_dim = int(dim * mlp_ratio)
|
|
# self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
|
|
|
|
def forward(self, x):
|
|
x = self.drop_path(self.attn(self.norm1(x)))
|
|
# x = x + self.drop_path(self.mlp(self.norm2(x)))
|
|
return x
|
|
|
|
def flops(self, N):
|
|
flops = 0
|
|
# att
|
|
flops += self.attn.flops(N)
|
|
# norm
|
|
flops += self.dim * N
|
|
return flops
|
|
|
|
|
|
class WindowAttentionBlock(nn.Module):
|
|
r""" Based on Swin Transformer Block, We remove FFN.
|
|
"""
|
|
|
|
def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
|
|
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
|
|
act_layer=nn.GELU, norm_layer=nn.LayerNorm,
|
|
fused_window_process=False):
|
|
super().__init__()
|
|
self.dim = dim
|
|
self.input_resolution = input_resolution
|
|
self.num_heads = num_heads
|
|
self.window_size = window_size
|
|
self.shift_size = shift_size
|
|
self.mlp_ratio = mlp_ratio
|
|
if min(self.input_resolution) <= self.window_size:
|
|
# if window size is larger than input resolution, we don't partition windows
|
|
self.shift_size = 0
|
|
self.window_size = min(self.input_resolution)
|
|
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
|
|
|
|
self.norm1 = norm_layer(dim)
|
|
self.attn = WindowAttention(
|
|
dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
|
|
qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
|
|
|
|
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
|
# self.norm2 = norm_layer(dim)
|
|
# mlp_hidden_dim = int(dim * mlp_ratio)
|
|
# self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
|
|
|
|
if self.shift_size > 0:
|
|
# calculate attention mask for SW-MSA
|
|
H, W = self.input_resolution
|
|
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
|
|
h_slices = (slice(0, -self.window_size),
|
|
slice(-self.window_size, -self.shift_size),
|
|
slice(-self.shift_size, None))
|
|
w_slices = (slice(0, -self.window_size),
|
|
slice(-self.window_size, -self.shift_size),
|
|
slice(-self.shift_size, None))
|
|
cnt = 0
|
|
for h in h_slices:
|
|
for w in w_slices:
|
|
img_mask[:, h, w, :] = cnt
|
|
cnt += 1
|
|
|
|
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
|
|
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
|
|
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
|
|
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
|
|
else:
|
|
attn_mask = None
|
|
|
|
self.register_buffer("attn_mask", attn_mask)
|
|
self.fused_window_process = fused_window_process
|
|
|
|
def forward(self, x):
|
|
H, W = self.input_resolution
|
|
B, L, C = x.shape
|
|
assert L == H * W, "input feature has wrong size"
|
|
|
|
shortcut = x
|
|
x = self.norm1(x)
|
|
x = x.view(B, H, W, C)
|
|
|
|
# cyclic shift
|
|
if self.shift_size > 0:
|
|
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
|
|
# partition windows
|
|
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
|
|
else:
|
|
shifted_x = x
|
|
# partition windows
|
|
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
|
|
|
|
x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
|
|
|
|
# W-MSA/SW-MSA
|
|
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
|
|
|
|
# merge windows
|
|
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
|
|
|
|
# reverse cyclic shift
|
|
if self.shift_size > 0:
|
|
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
|
|
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
|
|
else:
|
|
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
|
|
x = shifted_x
|
|
x = x.view(B, H * W, C)
|
|
x = self.drop_path(x)
|
|
|
|
# FFN
|
|
# x = x + self.drop_path(self.mlp(self.norm2(x)))
|
|
|
|
return x
|
|
|
|
def extra_repr(self) -> str:
|
|
return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
|
|
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
|
|
|
|
def flops(self):
|
|
flops = 0
|
|
H, W = self.input_resolution
|
|
# norm1
|
|
flops += self.dim * H * W
|
|
# W-MSA/SW-MSA
|
|
nW = H * W / self.window_size / self.window_size
|
|
flops += nW * self.attn.flops(self.window_size * self.window_size)
|
|
# mlp
|
|
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
|
|
# norm2
|
|
flops += self.dim * H * W
|
|
return flops
|
|
|
|
|
|
class MixedAttentionBlock(nn.Module):
|
|
def __init__(self, dim, img_size, window_size, num_heads=1, mlp_ratio=3, drop_path=0.):
|
|
super(MixedAttentionBlock, self).__init__()
|
|
|
|
self.img_size = img_size
|
|
self.dim = dim
|
|
self.mlp_ratio = mlp_ratio
|
|
|
|
self.windowatt = WindowAttentionBlock(dim=dim, input_resolution=img_size, num_heads=num_heads,
|
|
window_size=window_size, shift_size=0,
|
|
mlp_ratio=mlp_ratio, qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
|
|
drop_path=0.,
|
|
act_layer=nn.GELU, norm_layer=nn.LayerNorm,
|
|
fused_window_process=False)
|
|
self.globalatt = Block(dim=dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=False, qk_scale=None,
|
|
drop=0., attn_drop=0.,
|
|
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm)
|
|
|
|
self.norm = nn.LayerNorm(dim)
|
|
self.mlp = nn.Sequential(
|
|
nn.Linear(dim, dim * mlp_ratio),
|
|
nn.GELU(),
|
|
nn.Linear(dim * mlp_ratio, dim),
|
|
)
|
|
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
|
|
|
def forward(self, x):
|
|
att1 = self.windowatt(x)
|
|
att2 = self.globalatt(x)
|
|
x = x + att1 + att2
|
|
x = x + self.drop_path(self.mlp(self.norm(x)))
|
|
return x
|
|
|
|
def flops(self):
|
|
N = self.img_size[0] * self.img_size[1]
|
|
flops = 0
|
|
flops += self.windowatt.flops()
|
|
flops += self.globalatt.flops(N)
|
|
flops += self.dim * N
|
|
flops += 2 * N * self.dim * self.dim * self.mlp_ratio
|
|
return flops
|