APISR / architecture /grl_common /mixed_attn_block.py
HikariDawn's picture
feat: initial push
561c629
raw
history blame
39.2 kB
import math
from abc import ABC
from math import prod
import torch
import torch.nn as nn
import torch.nn.functional as F
from architecture.grl_common.ops import (
bchw_to_bhwc,
bchw_to_blc,
blc_to_bchw,
blc_to_bhwc,
calculate_mask,
calculate_mask_all,
get_relative_coords_table_all,
get_relative_position_index_simple,
window_partition,
window_reverse,
)
from architecture.grl_common.swin_v1_block import Mlp
from timm.models.layers import DropPath
class CPB_MLP(nn.Sequential):
def __init__(self, in_channels, out_channels, channels=512):
m = [
nn.Linear(in_channels, channels, bias=True),
nn.ReLU(inplace=True),
nn.Linear(channels, out_channels, bias=False),
]
super(CPB_MLP, self).__init__(*m)
class AffineTransformWindow(nn.Module):
r"""Affine transformation of the attention map.
The window is a square window.
Supports attention between different window sizes
"""
def __init__(
self,
num_heads,
input_resolution,
window_size,
pretrained_window_size=[0, 0],
shift_size=0,
anchor_window_down_factor=1,
args=None,
):
super(AffineTransformWindow, self).__init__()
# print("AffineTransformWindow", args)
self.num_heads = num_heads
self.input_resolution = input_resolution
self.window_size = window_size
self.pretrained_window_size = pretrained_window_size
self.shift_size = shift_size
self.anchor_window_down_factor = anchor_window_down_factor
self.use_buffer = args.use_buffer
logit_scale = torch.log(10 * torch.ones((num_heads, 1, 1)))
self.logit_scale = nn.Parameter(logit_scale, requires_grad=True)
# mlp to generate continuous relative position bias
self.cpb_mlp = CPB_MLP(2, num_heads)
if self.use_buffer:
table = get_relative_coords_table_all(
window_size, pretrained_window_size, anchor_window_down_factor
)
index = get_relative_position_index_simple(
window_size, anchor_window_down_factor
)
self.register_buffer("relative_coords_table", table)
self.register_buffer("relative_position_index", index)
if self.shift_size > 0:
attn_mask = calculate_mask(
input_resolution, self.window_size, self.shift_size
)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def forward(self, attn, x_size):
B_, H, N, _ = attn.shape
device = attn.device
# logit scale
attn = attn * torch.clamp(self.logit_scale, max=math.log(1.0 / 0.01)).exp()
# relative position bias
if self.use_buffer:
table = self.relative_coords_table
index = self.relative_position_index
else:
table = get_relative_coords_table_all(
self.window_size,
self.pretrained_window_size,
self.anchor_window_down_factor,
).to(device)
index = get_relative_position_index_simple(
self.window_size, self.anchor_window_down_factor
).to(device)
bias_table = self.cpb_mlp(table) # 2*Wh-1, 2*Ww-1, num_heads
bias_table = bias_table.view(-1, self.num_heads)
win_dim = prod(self.window_size)
bias = bias_table[index.view(-1)]
bias = bias.view(win_dim, win_dim, -1).permute(2, 0, 1).contiguous()
# nH, Wh*Ww, Wh*Ww
bias = 16 * torch.sigmoid(bias)
attn = attn + bias.unsqueeze(0)
# W-MSA/SW-MSA
if self.use_buffer:
mask = self.attn_mask
# during test and window shift, recalculate the mask
if self.input_resolution != x_size and self.shift_size > 0:
mask = calculate_mask(x_size, self.window_size, self.shift_size)
mask = mask.to(attn.device)
else:
if self.shift_size > 0:
mask = calculate_mask(x_size, self.window_size, self.shift_size)
mask = mask.to(attn.device)
else:
mask = None
# shift attention mask
if mask is not None:
nW = mask.shape[0]
mask = mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask
attn = attn.view(-1, self.num_heads, N, N)
return attn
class AffineTransformStripe(nn.Module):
r"""Affine transformation of the attention map.
The window is a stripe window. Supports attention between different window sizes
"""
def __init__(
self,
num_heads,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
pretrained_stripe_size=[0, 0],
anchor_window_down_factor=1,
window_to_anchor=True,
args=None,
):
super(AffineTransformStripe, self).__init__()
self.num_heads = num_heads
self.input_resolution = input_resolution
self.stripe_size = stripe_size
self.stripe_groups = stripe_groups
self.pretrained_stripe_size = pretrained_stripe_size
# TODO: be careful when determining the pretrained_stripe_size
self.stripe_shift = stripe_shift
stripe_size, shift_size = self._get_stripe_info(input_resolution)
self.anchor_window_down_factor = anchor_window_down_factor
self.window_to_anchor = window_to_anchor
self.use_buffer = args.use_buffer
logit_scale = torch.log(10 * torch.ones((num_heads, 1, 1)))
self.logit_scale = nn.Parameter(logit_scale, requires_grad=True)
# mlp to generate continuous relative position bias
self.cpb_mlp = CPB_MLP(2, num_heads)
if self.use_buffer:
table = get_relative_coords_table_all(
stripe_size, pretrained_stripe_size, anchor_window_down_factor
)
index = get_relative_position_index_simple(
stripe_size, anchor_window_down_factor, window_to_anchor
)
self.register_buffer("relative_coords_table", table)
self.register_buffer("relative_position_index", index)
if self.stripe_shift:
attn_mask = calculate_mask_all(
input_resolution,
stripe_size,
shift_size,
anchor_window_down_factor,
window_to_anchor,
)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def forward(self, attn, x_size):
B_, H, N1, N2 = attn.shape
device = attn.device
# logit scale
attn = attn * torch.clamp(self.logit_scale, max=math.log(1.0 / 0.01)).exp()
# relative position bias
stripe_size, shift_size = self._get_stripe_info(x_size)
fixed_stripe_size = (
self.stripe_groups[0] is None and self.stripe_groups[1] is None
)
if not self.use_buffer or (
self.use_buffer
and self.input_resolution != x_size
and not fixed_stripe_size
):
# during test and stripe size is not fixed.
pretrained_stripe_size = (
self.pretrained_stripe_size
) # or stripe_size; Needs further pondering
table = get_relative_coords_table_all(
stripe_size, pretrained_stripe_size, self.anchor_window_down_factor
)
table = table.to(device)
index = get_relative_position_index_simple(
stripe_size, self.anchor_window_down_factor, self.window_to_anchor
).to(device)
else:
table = self.relative_coords_table
index = self.relative_position_index
# The same table size-> 1, Wh+AWh-1, Ww+AWw-1, 2
# But different index size -> # Wh*Ww, AWh*AWw
# if N1 < N2:
# index = index.transpose(0, 1)
bias_table = self.cpb_mlp(table).view(-1, self.num_heads)
# if not self.training:
# print(bias_table.shape, index.max(), index.min())
bias = bias_table[index.view(-1)]
bias = bias.view(N1, N2, -1).permute(2, 0, 1).contiguous()
# nH, Wh*Ww, Wh*Ww
bias = 16 * torch.sigmoid(bias)
# print(N1, N2, attn.shape, bias.unsqueeze(0).shape)
attn = attn + bias.unsqueeze(0)
# W-MSA/SW-MSA
if self.use_buffer:
mask = self.attn_mask
# during test and window shift, recalculate the mask
if self.input_resolution != x_size and self.stripe_shift > 0:
mask = calculate_mask_all(
x_size,
stripe_size,
shift_size,
self.anchor_window_down_factor,
self.window_to_anchor,
)
mask = mask.to(device)
else:
if self.stripe_shift > 0:
mask = calculate_mask_all(
x_size,
stripe_size,
shift_size,
self.anchor_window_down_factor,
self.window_to_anchor,
)
mask = mask.to(attn.device)
else:
mask = None
# shift attention mask
if mask is not None:
nW = mask.shape[0]
mask = mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(B_ // nW, nW, self.num_heads, N1, N2) + mask
attn = attn.view(-1, self.num_heads, N1, N2)
return attn
def _get_stripe_info(self, input_resolution):
stripe_size, shift_size = [], []
for s, g, d in zip(self.stripe_size, self.stripe_groups, input_resolution):
if g is None:
stripe_size.append(s)
shift_size.append(s // 2 if self.stripe_shift else 0)
else:
stripe_size.append(d // g)
shift_size.append(0 if g == 1 else d // (g * 2))
return stripe_size, shift_size
class Attention(ABC, nn.Module):
def __init__(self):
super(Attention, self).__init__()
def attn(self, q, k, v, attn_transform, x_size, reshape=True):
# cosine attention map
B_, _, H, head_dim = q.shape
if self.euclidean_dist:
attn = torch.norm(q.unsqueeze(-2) - k.unsqueeze(-3), dim=-1)
else:
attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
attn = attn_transform(attn, x_size)
# attention
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = attn @ v # B_, H, N1, head_dim
if reshape:
x = x.transpose(1, 2).reshape(B_, -1, H * head_dim)
# B_, N, C
return x
class WindowAttention(Attention):
r"""Window attention. QKV is the input to the forward method.
Args:
num_heads (int): Number of attention heads.
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
"""
def __init__(
self,
input_resolution,
window_size,
num_heads,
window_shift=False,
attn_drop=0.0,
pretrained_window_size=[0, 0],
args=None,
):
super(WindowAttention, self).__init__()
self.input_resolution = input_resolution
self.window_size = window_size
self.pretrained_window_size = pretrained_window_size
self.num_heads = num_heads
self.shift_size = window_size[0] // 2 if window_shift else 0
self.euclidean_dist = args.euclidean_dist
self.attn_transform = AffineTransformWindow(
num_heads,
input_resolution,
window_size,
pretrained_window_size,
self.shift_size,
args=args,
)
self.attn_drop = nn.Dropout(attn_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(self, qkv, x_size):
"""
Args:
qkv: input QKV features with shape of (B, L, 3C)
x_size: use x_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
H, W = x_size
B, L, C = qkv.shape
qkv = qkv.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
qkv = torch.roll(
qkv, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
)
# partition windows
qkv = window_partition(qkv, self.window_size) # nW*B, wh, ww, C
qkv = qkv.view(-1, prod(self.window_size), C) # nW*B, wh*ww, C
B_, N, _ = qkv.shape
qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
# attention
x = self.attn(q, k, v, self.attn_transform, x_size)
# merge windows
x = x.view(-1, *self.window_size, C // 3)
x = window_reverse(x, self.window_size, x_size) # B, H, W, C/3
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
x = x.view(B, L, C // 3)
return x
def extra_repr(self) -> str:
return (
f"window_size={self.window_size}, shift_size={self.shift_size}, "
f"pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}"
)
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class StripeAttention(Attention):
r"""Stripe attention
Args:
stripe_size (tuple[int]): The height and width of the stripe.
num_heads (int): Number of attention heads.
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
pretrained_stripe_size (tuple[int]): The height and width of the stripe in pre-training.
"""
def __init__(
self,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
num_heads,
attn_drop=0.0,
pretrained_stripe_size=[0, 0],
args=None,
):
super(StripeAttention, self).__init__()
self.input_resolution = input_resolution
self.stripe_size = stripe_size # Wh, Ww
self.stripe_groups = stripe_groups
self.stripe_shift = stripe_shift
self.num_heads = num_heads
self.pretrained_stripe_size = pretrained_stripe_size
self.euclidean_dist = args.euclidean_dist
self.attn_transform = AffineTransformStripe(
num_heads,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
pretrained_stripe_size,
anchor_window_down_factor=1,
args=args,
)
self.attn_drop = nn.Dropout(attn_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(self, qkv, x_size):
"""
Args:
x: input features with shape of (B, L, C)
stripe_size: use stripe_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
H, W = x_size
B, L, C = qkv.shape
qkv = qkv.view(B, H, W, C)
running_stripe_size, running_shift_size = self.attn_transform._get_stripe_info(
x_size
)
# cyclic shift
if self.stripe_shift:
qkv = torch.roll(
qkv,
shifts=(-running_shift_size[0], -running_shift_size[1]),
dims=(1, 2),
)
# partition windows
qkv = window_partition(qkv, running_stripe_size) # nW*B, wh, ww, C
qkv = qkv.view(-1, prod(running_stripe_size), C) # nW*B, wh*ww, C
B_, N, _ = qkv.shape
qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
# attention
x = self.attn(q, k, v, self.attn_transform, x_size)
# merge windows
x = x.view(-1, *running_stripe_size, C // 3)
x = window_reverse(x, running_stripe_size, x_size) # B H W C/3
# reverse the shift
if self.stripe_shift:
x = torch.roll(x, shifts=running_shift_size, dims=(1, 2))
x = x.view(B, L, C // 3)
return x
def extra_repr(self) -> str:
return (
f"stripe_size={self.stripe_size}, stripe_groups={self.stripe_groups}, stripe_shift={self.stripe_shift}, "
f"pretrained_stripe_size={self.pretrained_stripe_size}, num_heads={self.num_heads}"
)
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class AnchorStripeAttention(Attention):
r"""Stripe attention
Args:
stripe_size (tuple[int]): The height and width of the stripe.
num_heads (int): Number of attention heads.
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
pretrained_stripe_size (tuple[int]): The height and width of the stripe in pre-training.
"""
def __init__(
self,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
num_heads,
attn_drop=0.0,
pretrained_stripe_size=[0, 0],
anchor_window_down_factor=1,
args=None,
):
super(AnchorStripeAttention, self).__init__()
self.input_resolution = input_resolution
self.stripe_size = stripe_size # Wh, Ww
self.stripe_groups = stripe_groups
self.stripe_shift = stripe_shift
self.num_heads = num_heads
self.pretrained_stripe_size = pretrained_stripe_size
self.anchor_window_down_factor = anchor_window_down_factor
self.euclidean_dist = args.euclidean_dist
self.attn_transform1 = AffineTransformStripe(
num_heads,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
pretrained_stripe_size,
anchor_window_down_factor,
window_to_anchor=False,
args=args,
)
self.attn_transform2 = AffineTransformStripe(
num_heads,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
pretrained_stripe_size,
anchor_window_down_factor,
window_to_anchor=True,
args=args,
)
self.attn_drop = nn.Dropout(attn_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(self, qkv, anchor, x_size):
"""
Args:
qkv: input features with shape of (B, L, C)
anchor:
x_size: use stripe_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
H, W = x_size
B, L, C = qkv.shape
qkv = qkv.view(B, H, W, C)
stripe_size, shift_size = self.attn_transform1._get_stripe_info(x_size)
anchor_stripe_size = [s // self.anchor_window_down_factor for s in stripe_size]
anchor_shift_size = [s // self.anchor_window_down_factor for s in shift_size]
# cyclic shift
if self.stripe_shift:
qkv = torch.roll(qkv, shifts=(-shift_size[0], -shift_size[1]), dims=(1, 2))
anchor = torch.roll(
anchor,
shifts=(-anchor_shift_size[0], -anchor_shift_size[1]),
dims=(1, 2),
)
# partition windows
qkv = window_partition(qkv, stripe_size) # nW*B, wh, ww, C
qkv = qkv.view(-1, prod(stripe_size), C) # nW*B, wh*ww, C
anchor = window_partition(anchor, anchor_stripe_size)
anchor = anchor.view(-1, prod(anchor_stripe_size), C // 3)
B_, N1, _ = qkv.shape
N2 = anchor.shape[1]
qkv = qkv.reshape(B_, N1, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
anchor = anchor.reshape(B_, N2, self.num_heads, -1).permute(0, 2, 1, 3)
# attention
x = self.attn(anchor, k, v, self.attn_transform1, x_size, False)
x = self.attn(q, anchor, x, self.attn_transform2, x_size)
# merge windows
x = x.view(B_, *stripe_size, C // 3)
x = window_reverse(x, stripe_size, x_size) # B H' W' C
# reverse the shift
if self.stripe_shift:
x = torch.roll(x, shifts=shift_size, dims=(1, 2))
x = x.view(B, H * W, C // 3)
return x
def extra_repr(self) -> str:
return (
f"stripe_size={self.stripe_size}, stripe_groups={self.stripe_groups}, stripe_shift={self.stripe_shift}, "
f"pretrained_stripe_size={self.pretrained_stripe_size}, num_heads={self.num_heads}, anchor_window_down_factor={self.anchor_window_down_factor}"
)
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class SeparableConv(nn.Sequential):
def __init__(self, in_channels, out_channels, kernel_size, stride, bias, args):
m = [
nn.Conv2d(
in_channels,
in_channels,
kernel_size,
stride,
kernel_size // 2,
groups=in_channels,
bias=bias,
)
]
if args.separable_conv_act:
m.append(nn.GELU())
m.append(nn.Conv2d(in_channels, out_channels, 1, 1, 0, bias=bias))
super(SeparableConv, self).__init__(*m)
class QKVProjection(nn.Module):
def __init__(self, dim, qkv_bias, proj_type, args):
super(QKVProjection, self).__init__()
self.proj_type = proj_type
if proj_type == "linear":
self.body = nn.Linear(dim, dim * 3, bias=qkv_bias)
else:
self.body = SeparableConv(dim, dim * 3, 3, 1, qkv_bias, args)
def forward(self, x, x_size):
if self.proj_type == "separable_conv":
x = blc_to_bchw(x, x_size)
x = self.body(x)
if self.proj_type == "separable_conv":
x = bchw_to_blc(x)
return x
class PatchMerging(nn.Module):
r"""Patch Merging Layer.
Args:
dim (int): Number of input channels.
"""
def __init__(self, in_dim, out_dim):
super().__init__()
self.in_dim = in_dim
self.out_dim = out_dim
self.reduction = nn.Linear(4 * in_dim, out_dim, bias=False)
def forward(self, x, x_size):
"""
x: B, H*W, C
"""
H, W = x_size
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
x = x.view(B, H, W, C)
x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
x = self.reduction(x)
return x
class AnchorLinear(nn.Module):
r"""Linear anchor projection layer
Args:
dim (int): Number of input channels.
"""
def __init__(self, in_channels, out_channels, down_factor, pooling_mode, bias):
super().__init__()
self.down_factor = down_factor
if pooling_mode == "maxpool":
self.pooling = nn.MaxPool2d(down_factor, down_factor)
elif pooling_mode == "avgpool":
self.pooling = nn.AvgPool2d(down_factor, down_factor)
self.reduction = nn.Linear(in_channels, out_channels, bias=bias)
def forward(self, x, x_size):
"""
x: B, H*W, C
"""
x = blc_to_bchw(x, x_size)
x = bchw_to_blc(self.pooling(x))
x = blc_to_bhwc(self.reduction(x), [s // self.down_factor for s in x_size])
return x
class AnchorProjection(nn.Module):
def __init__(self, dim, proj_type, one_stage, anchor_window_down_factor, args):
super(AnchorProjection, self).__init__()
self.proj_type = proj_type
self.body = nn.ModuleList([])
if one_stage:
if proj_type == "patchmerging":
m = PatchMerging(dim, dim // 2)
elif proj_type == "conv2d":
kernel_size = anchor_window_down_factor + 1
stride = anchor_window_down_factor
padding = kernel_size // 2
m = nn.Conv2d(dim, dim // 2, kernel_size, stride, padding)
elif proj_type == "separable_conv":
kernel_size = anchor_window_down_factor + 1
stride = anchor_window_down_factor
m = SeparableConv(dim, dim // 2, kernel_size, stride, True, args)
elif proj_type.find("pool") >= 0:
m = AnchorLinear(
dim, dim // 2, anchor_window_down_factor, proj_type, True
)
self.body.append(m)
else:
for i in range(int(math.log2(anchor_window_down_factor))):
cin = dim if i == 0 else dim // 2
if proj_type == "patchmerging":
m = PatchMerging(cin, dim // 2)
elif proj_type == "conv2d":
m = nn.Conv2d(cin, dim // 2, 3, 2, 1)
elif proj_type == "separable_conv":
m = SeparableConv(cin, dim // 2, 3, 2, True, args)
self.body.append(m)
def forward(self, x, x_size):
if self.proj_type.find("conv") >= 0:
x = blc_to_bchw(x, x_size)
for m in self.body:
x = m(x)
x = bchw_to_bhwc(x)
elif self.proj_type.find("pool") >= 0:
for m in self.body:
x = m(x, x_size)
else:
for i, m in enumerate(self.body):
x = m(x, [s // 2**i for s in x_size])
x = blc_to_bhwc(x, [s // 2 ** (i + 1) for s in x_size])
return x
class MixedAttention(nn.Module):
r"""Mixed window attention and stripe attention
Args:
dim (int): Number of input channels.
stripe_size (tuple[int]): The height and width of the stripe.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
pretrained_stripe_size (tuple[int]): The height and width of the stripe in pre-training.
"""
def __init__(
self,
dim,
input_resolution,
num_heads_w,
num_heads_s,
window_size,
window_shift,
stripe_size,
stripe_groups,
stripe_shift,
qkv_bias=True,
qkv_proj_type="linear",
anchor_proj_type="separable_conv",
anchor_one_stage=True,
anchor_window_down_factor=1,
attn_drop=0.0,
proj_drop=0.0,
pretrained_window_size=[0, 0],
pretrained_stripe_size=[0, 0],
args=None,
):
super(MixedAttention, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.use_anchor = anchor_window_down_factor > 1
self.args = args
# print(args)
self.qkv = QKVProjection(dim, qkv_bias, qkv_proj_type, args)
if self.use_anchor:
# anchor is only used for stripe attention
self.anchor = AnchorProjection(
dim, anchor_proj_type, anchor_one_stage, anchor_window_down_factor, args
)
self.window_attn = WindowAttention(
input_resolution,
window_size,
num_heads_w,
window_shift,
attn_drop,
pretrained_window_size,
args,
)
if self.args.double_window:
self.stripe_attn = WindowAttention(
input_resolution,
window_size,
num_heads_w,
window_shift,
attn_drop,
pretrained_window_size,
args,
)
else:
if self.use_anchor:
self.stripe_attn = AnchorStripeAttention(
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
num_heads_s,
attn_drop,
pretrained_stripe_size,
anchor_window_down_factor,
args,
)
else:
if self.args.stripe_square:
self.stripe_attn = StripeAttention(
input_resolution,
window_size,
[None, None],
window_shift,
num_heads_s,
attn_drop,
pretrained_stripe_size,
args,
)
else:
self.stripe_attn = StripeAttention(
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
num_heads_s,
attn_drop,
pretrained_stripe_size,
args,
)
if self.args.out_proj_type == "linear":
self.proj = nn.Linear(dim, dim)
else:
self.proj = nn.Conv2d(dim, dim, 3, 1, 1)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x, x_size):
"""
Args:
x: input features with shape of (B, L, C)
stripe_size: use stripe_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
B, L, C = x.shape
# qkv projection
qkv = self.qkv(x, x_size)
qkv_window, qkv_stripe = torch.split(qkv, C * 3 // 2, dim=-1)
# anchor projection
if self.use_anchor:
anchor = self.anchor(x, x_size)
# attention
x_window = self.window_attn(qkv_window, x_size)
if self.use_anchor:
x_stripe = self.stripe_attn(qkv_stripe, anchor, x_size)
else:
x_stripe = self.stripe_attn(qkv_stripe, x_size)
x = torch.cat([x_window, x_stripe], dim=-1)
# output projection
if self.args.out_proj_type == "linear":
x = self.proj(x)
else:
x = blc_to_bchw(x, x_size)
x = bchw_to_blc(self.proj(x))
x = self.proj_drop(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}"
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class ChannelAttention(nn.Module):
"""Channel attention used in RCAN.
Args:
num_feat (int): Channel number of intermediate features.
reduction (int): Channel reduction factor. Default: 16.
"""
def __init__(self, num_feat, reduction=16):
super(ChannelAttention, self).__init__()
self.attention = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(num_feat, num_feat // reduction, 1, padding=0),
nn.ReLU(inplace=True),
nn.Conv2d(num_feat // reduction, num_feat, 1, padding=0),
nn.Sigmoid(),
)
def forward(self, x):
y = self.attention(x)
return x * y
class CAB(nn.Module):
def __init__(self, num_feat, compress_ratio=4, reduction=18):
super(CAB, self).__init__()
self.cab = nn.Sequential(
nn.Conv2d(num_feat, num_feat // compress_ratio, 3, 1, 1),
nn.GELU(),
nn.Conv2d(num_feat // compress_ratio, num_feat, 3, 1, 1),
ChannelAttention(num_feat, reduction),
)
def forward(self, x, x_size):
x = self.cab(blc_to_bchw(x, x_size).contiguous())
return bchw_to_blc(x)
class MixAttnTransformerBlock(nn.Module):
r"""Mix attention transformer block with shared QKV projection and output projection for mixed attention modules.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resulotion.
num_heads (int): Number of attention heads.
window_size (int): Window size.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float, optional): Stochastic depth rate. Default: 0.0
act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
pretrained_stripe_size (int): Window size in pre-training.
attn_type (str, optional): Attention type. Default: cwhv.
c: residual blocks
w: window attention
h: horizontal stripe attention
v: vertical stripe attention
"""
def __init__(
self,
dim,
input_resolution,
num_heads_w,
num_heads_s,
window_size=7,
window_shift=False,
stripe_size=[8, 8],
stripe_groups=[None, None],
stripe_shift=False,
stripe_type="H",
mlp_ratio=4.0,
qkv_bias=True,
qkv_proj_type="linear",
anchor_proj_type="separable_conv",
anchor_one_stage=True,
anchor_window_down_factor=1,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
pretrained_window_size=[0, 0],
pretrained_stripe_size=[0, 0],
res_scale=1.0,
args=None,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads_w = num_heads_w
self.num_heads_s = num_heads_s
self.window_size = window_size
self.window_shift = window_shift
self.stripe_shift = stripe_shift
self.stripe_type = stripe_type
self.args = args
if self.stripe_type == "W":
self.stripe_size = stripe_size[::-1]
self.stripe_groups = stripe_groups[::-1]
else:
self.stripe_size = stripe_size
self.stripe_groups = stripe_groups
self.mlp_ratio = mlp_ratio
self.res_scale = res_scale
self.attn = MixedAttention(
dim,
input_resolution,
num_heads_w,
num_heads_s,
window_size,
window_shift,
self.stripe_size,
self.stripe_groups,
stripe_shift,
qkv_bias,
qkv_proj_type,
anchor_proj_type,
anchor_one_stage,
anchor_window_down_factor,
attn_drop,
drop,
pretrained_window_size,
pretrained_stripe_size,
args,
)
self.norm1 = norm_layer(dim)
if self.args.local_connection:
self.conv = CAB(dim)
# self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
# self.mlp = Mlp(
# in_features=dim,
# hidden_features=int(dim * mlp_ratio),
# act_layer=act_layer,
# drop=drop,
# )
# self.norm2 = norm_layer(dim)
def forward(self, x, x_size):
# Mixed attention
if self.args.local_connection:
x = (
x
+ self.res_scale * self.drop_path(self.norm1(self.attn(x, x_size)))
+ self.conv(x, x_size)
)
else:
x = x + self.res_scale * self.drop_path(self.norm1(self.attn(x, x_size)))
# FFN
x = x + self.res_scale * self.drop_path(self.norm2(self.mlp(x)))
# return x
def extra_repr(self) -> str:
return (
f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads=({self.num_heads_w}, {self.num_heads_s}), "
f"window_size={self.window_size}, window_shift={self.window_shift}, "
f"stripe_size={self.stripe_size}, stripe_groups={self.stripe_groups}, stripe_shift={self.stripe_shift}, self.stripe_type={self.stripe_type}, "
f"mlp_ratio={self.mlp_ratio}, res_scale={self.res_scale}"
)
# def flops(self):
# flops = 0
# H, W = self.input_resolution
# # norm1
# flops += self.dim * H * W
# # W-MSA/SW-MSA
# nW = H * W / self.stripe_size[0] / self.stripe_size[1]
# flops += nW * self.attn.flops(self.stripe_size[0] * self.stripe_size[1])
# # mlp
# flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
# # norm2
# flops += self.dim * H * W
# return flops