ComfyUI/comfy_extras/chainner_models/architecture/DAT.py

1183 lines
40 KiB
Python

# pylint: skip-file
import math
import re
import numpy as np
import torch
import torch.nn as nn
import torch.utils.checkpoint as checkpoint
from einops import rearrange
from einops.layers.torch import Rearrange
from torch import Tensor
from torch.nn import functional as F
from .timm.drop import DropPath
from .timm.weight_init import trunc_normal_
def img2windows(img, H_sp, W_sp):
"""
Input: Image (B, C, H, W)
Output: Window Partition (B', N, C)
"""
B, C, H, W = img.shape
img_reshape = img.view(B, C, H // H_sp, H_sp, W // W_sp, W_sp)
img_perm = (
img_reshape.permute(0, 2, 4, 3, 5, 1).contiguous().reshape(-1, H_sp * W_sp, C)
)
return img_perm
def windows2img(img_splits_hw, H_sp, W_sp, H, W):
"""
Input: Window Partition (B', N, C)
Output: Image (B, H, W, C)
"""
B = int(img_splits_hw.shape[0] / (H * W / H_sp / W_sp))
img = img_splits_hw.view(B, H // H_sp, W // W_sp, H_sp, W_sp, -1)
img = img.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return img
class SpatialGate(nn.Module):
"""Spatial-Gate.
Args:
dim (int): Half of input channels.
"""
def __init__(self, dim):
super().__init__()
self.norm = nn.LayerNorm(dim)
self.conv = nn.Conv2d(
dim, dim, kernel_size=3, stride=1, padding=1, groups=dim
) # DW Conv
def forward(self, x, H, W):
# Split
x1, x2 = x.chunk(2, dim=-1)
B, N, C = x.shape
x2 = (
self.conv(self.norm(x2).transpose(1, 2).contiguous().view(B, C // 2, H, W))
.flatten(2)
.transpose(-1, -2)
.contiguous()
)
return x1 * x2
class SGFN(nn.Module):
"""Spatial-Gate Feed-Forward Network.
Args:
in_features (int): Number of input channels.
hidden_features (int | None): Number of hidden channels. Default: None
out_features (int | None): Number of output channels. Default: None
act_layer (nn.Module): Activation layer. Default: nn.GELU
drop (float): Dropout rate. Default: 0.0
"""
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer=nn.GELU,
drop=0.0,
):
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.sg = SpatialGate(hidden_features // 2)
self.fc2 = nn.Linear(hidden_features // 2, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x, H, W):
"""
Input: x: (B, H*W, C), H, W
Output: x: (B, H*W, C)
"""
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.sg(x, H, W)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class DynamicPosBias(nn.Module):
# The implementation builds on Crossformer code https://github.com/cheerss/CrossFormer/blob/main/models/crossformer.py
"""Dynamic Relative Position Bias.
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads.
residual (bool): If True, use residual strage to connect conv.
"""
def __init__(self, dim, num_heads, residual):
super().__init__()
self.residual = residual
self.num_heads = num_heads
self.pos_dim = dim // 4
self.pos_proj = nn.Linear(2, self.pos_dim)
self.pos1 = nn.Sequential(
nn.LayerNorm(self.pos_dim),
nn.ReLU(inplace=True),
nn.Linear(self.pos_dim, self.pos_dim),
)
self.pos2 = nn.Sequential(
nn.LayerNorm(self.pos_dim),
nn.ReLU(inplace=True),
nn.Linear(self.pos_dim, self.pos_dim),
)
self.pos3 = nn.Sequential(
nn.LayerNorm(self.pos_dim),
nn.ReLU(inplace=True),
nn.Linear(self.pos_dim, self.num_heads),
)
def forward(self, biases):
if self.residual:
pos = self.pos_proj(biases) # 2Gh-1 * 2Gw-1, heads
pos = pos + self.pos1(pos)
pos = pos + self.pos2(pos)
pos = self.pos3(pos)
else:
pos = self.pos3(self.pos2(self.pos1(self.pos_proj(biases))))
return pos
class Spatial_Attention(nn.Module):
"""Spatial Window Self-Attention.
It supports rectangle window (containing square window).
Args:
dim (int): Number of input channels.
idx (int): The indentix of window. (0/1)
split_size (tuple(int)): Height and Width of spatial window.
dim_out (int | None): The dimension of the attention output. Default: None
num_heads (int): Number of attention heads. Default: 6
attn_drop (float): Dropout ratio of attention weight. Default: 0.0
proj_drop (float): Dropout ratio of output. Default: 0.0
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set
position_bias (bool): The dynamic relative position bias. Default: True
"""
def __init__(
self,
dim,
idx,
split_size=[8, 8],
dim_out=None,
num_heads=6,
attn_drop=0.0,
proj_drop=0.0,
qk_scale=None,
position_bias=True,
):
super().__init__()
self.dim = dim
self.dim_out = dim_out or dim
self.split_size = split_size
self.num_heads = num_heads
self.idx = idx
self.position_bias = position_bias
head_dim = dim // num_heads
self.scale = qk_scale or head_dim**-0.5
if idx == 0:
H_sp, W_sp = self.split_size[0], self.split_size[1]
elif idx == 1:
W_sp, H_sp = self.split_size[0], self.split_size[1]
else:
print("ERROR MODE", idx)
exit(0)
self.H_sp = H_sp
self.W_sp = W_sp
if self.position_bias:
self.pos = DynamicPosBias(self.dim // 4, self.num_heads, residual=False)
# generate mother-set
position_bias_h = torch.arange(1 - self.H_sp, self.H_sp)
position_bias_w = torch.arange(1 - self.W_sp, self.W_sp)
biases = torch.stack(torch.meshgrid([position_bias_h, position_bias_w]))
biases = biases.flatten(1).transpose(0, 1).contiguous().float()
self.register_buffer("rpe_biases", biases)
# get pair-wise relative position index for each token inside the window
coords_h = torch.arange(self.H_sp)
coords_w = torch.arange(self.W_sp)
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.H_sp - 1
relative_coords[:, :, 1] += self.W_sp - 1
relative_coords[:, :, 0] *= 2 * self.W_sp - 1
relative_position_index = relative_coords.sum(-1)
self.register_buffer("relative_position_index", relative_position_index)
self.attn_drop = nn.Dropout(attn_drop)
def im2win(self, x, H, W):
B, N, C = x.shape
x = x.transpose(-2, -1).contiguous().view(B, C, H, W)
x = img2windows(x, self.H_sp, self.W_sp)
x = (
x.reshape(-1, self.H_sp * self.W_sp, self.num_heads, C // self.num_heads)
.permute(0, 2, 1, 3)
.contiguous()
)
return x
def forward(self, qkv, H, W, mask=None):
"""
Input: qkv: (B, 3*L, C), H, W, mask: (B, N, N), N is the window size
Output: x (B, H, W, C)
"""
q, k, v = qkv[0], qkv[1], qkv[2]
B, L, C = q.shape
assert L == H * W, "flatten img_tokens has wrong size"
# partition the q,k,v, image to window
q = self.im2win(q, H, W)
k = self.im2win(k, H, W)
v = self.im2win(v, H, W)
q = q * self.scale
attn = q @ k.transpose(-2, -1) # B head N C @ B head C N --> B head N N
# calculate drpe
if self.position_bias:
pos = self.pos(self.rpe_biases)
# select position bias
relative_position_bias = pos[self.relative_position_index.view(-1)].view(
self.H_sp * self.W_sp, self.H_sp * self.W_sp, -1
)
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous()
attn = attn + relative_position_bias.unsqueeze(0)
N = attn.shape[3]
# use mask for shift window
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(
0
)
attn = attn.view(-1, self.num_heads, N, N)
attn = nn.functional.softmax(attn, dim=-1, dtype=attn.dtype)
attn = self.attn_drop(attn)
x = attn @ v
x = x.transpose(1, 2).reshape(
-1, self.H_sp * self.W_sp, C
) # B head N N @ B head N C
# merge the window, window to image
x = windows2img(x, self.H_sp, self.W_sp, H, W) # B H' W' C
return x
class Adaptive_Spatial_Attention(nn.Module):
# The implementation builds on CAT code https://github.com/Zhengchen1999/CAT
"""Adaptive Spatial Self-Attention
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads. Default: 6
split_size (tuple(int)): Height and Width of spatial window.
shift_size (tuple(int)): Shift size for spatial window.
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set.
drop (float): Dropout rate. Default: 0.0
attn_drop (float): Attention dropout rate. Default: 0.0
rg_idx (int): The indentix of Residual Group (RG)
b_idx (int): The indentix of Block in each RG
"""
def __init__(
self,
dim,
num_heads,
reso=64,
split_size=[8, 8],
shift_size=[1, 2],
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
rg_idx=0,
b_idx=0,
):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.split_size = split_size
self.shift_size = shift_size
self.b_idx = b_idx
self.rg_idx = rg_idx
self.patches_resolution = reso
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
assert (
0 <= self.shift_size[0] < self.split_size[0]
), "shift_size must in 0-split_size0"
assert (
0 <= self.shift_size[1] < self.split_size[1]
), "shift_size must in 0-split_size1"
self.branch_num = 2
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(drop)
self.attns = nn.ModuleList(
[
Spatial_Attention(
dim // 2,
idx=i,
split_size=split_size,
num_heads=num_heads // 2,
dim_out=dim // 2,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
position_bias=True,
)
for i in range(self.branch_num)
]
)
if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or (
self.rg_idx % 2 != 0 and self.b_idx % 4 == 0
):
attn_mask = self.calculate_mask(
self.patches_resolution, self.patches_resolution
)
self.register_buffer("attn_mask_0", attn_mask[0])
self.register_buffer("attn_mask_1", attn_mask[1])
else:
attn_mask = None
self.register_buffer("attn_mask_0", None)
self.register_buffer("attn_mask_1", None)
self.dwconv = nn.Sequential(
nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim),
nn.BatchNorm2d(dim),
nn.GELU(),
)
self.channel_interaction = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(dim, dim // 8, kernel_size=1),
nn.BatchNorm2d(dim // 8),
nn.GELU(),
nn.Conv2d(dim // 8, dim, kernel_size=1),
)
self.spatial_interaction = nn.Sequential(
nn.Conv2d(dim, dim // 16, kernel_size=1),
nn.BatchNorm2d(dim // 16),
nn.GELU(),
nn.Conv2d(dim // 16, 1, kernel_size=1),
)
def calculate_mask(self, H, W):
# The implementation builds on Swin Transformer code https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
# calculate attention mask for shift window
img_mask_0 = torch.zeros((1, H, W, 1)) # 1 H W 1 idx=0
img_mask_1 = torch.zeros((1, H, W, 1)) # 1 H W 1 idx=1
h_slices_0 = (
slice(0, -self.split_size[0]),
slice(-self.split_size[0], -self.shift_size[0]),
slice(-self.shift_size[0], None),
)
w_slices_0 = (
slice(0, -self.split_size[1]),
slice(-self.split_size[1], -self.shift_size[1]),
slice(-self.shift_size[1], None),
)
h_slices_1 = (
slice(0, -self.split_size[1]),
slice(-self.split_size[1], -self.shift_size[1]),
slice(-self.shift_size[1], None),
)
w_slices_1 = (
slice(0, -self.split_size[0]),
slice(-self.split_size[0], -self.shift_size[0]),
slice(-self.shift_size[0], None),
)
cnt = 0
for h in h_slices_0:
for w in w_slices_0:
img_mask_0[:, h, w, :] = cnt
cnt += 1
cnt = 0
for h in h_slices_1:
for w in w_slices_1:
img_mask_1[:, h, w, :] = cnt
cnt += 1
# calculate mask for window-0
img_mask_0 = img_mask_0.view(
1,
H // self.split_size[0],
self.split_size[0],
W // self.split_size[1],
self.split_size[1],
1,
)
img_mask_0 = (
img_mask_0.permute(0, 1, 3, 2, 4, 5)
.contiguous()
.view(-1, self.split_size[0], self.split_size[1], 1)
) # nW, sw[0], sw[1], 1
mask_windows_0 = img_mask_0.view(-1, self.split_size[0] * self.split_size[1])
attn_mask_0 = mask_windows_0.unsqueeze(1) - mask_windows_0.unsqueeze(2)
attn_mask_0 = attn_mask_0.masked_fill(
attn_mask_0 != 0, float(-100.0)
).masked_fill(attn_mask_0 == 0, float(0.0))
# calculate mask for window-1
img_mask_1 = img_mask_1.view(
1,
H // self.split_size[1],
self.split_size[1],
W // self.split_size[0],
self.split_size[0],
1,
)
img_mask_1 = (
img_mask_1.permute(0, 1, 3, 2, 4, 5)
.contiguous()
.view(-1, self.split_size[1], self.split_size[0], 1)
) # nW, sw[1], sw[0], 1
mask_windows_1 = img_mask_1.view(-1, self.split_size[1] * self.split_size[0])
attn_mask_1 = mask_windows_1.unsqueeze(1) - mask_windows_1.unsqueeze(2)
attn_mask_1 = attn_mask_1.masked_fill(
attn_mask_1 != 0, float(-100.0)
).masked_fill(attn_mask_1 == 0, float(0.0))
return attn_mask_0, attn_mask_1
def forward(self, x, H, W):
"""
Input: x: (B, H*W, C), H, W
Output: x: (B, H*W, C)
"""
B, L, C = x.shape
assert L == H * W, "flatten img_tokens has wrong size"
qkv = self.qkv(x).reshape(B, -1, 3, C).permute(2, 0, 1, 3) # 3, B, HW, C
# V without partition
v = qkv[2].transpose(-2, -1).contiguous().view(B, C, H, W)
# image padding
max_split_size = max(self.split_size[0], self.split_size[1])
pad_l = pad_t = 0
pad_r = (max_split_size - W % max_split_size) % max_split_size
pad_b = (max_split_size - H % max_split_size) % max_split_size
qkv = qkv.reshape(3 * B, H, W, C).permute(0, 3, 1, 2) # 3B C H W
qkv = (
F.pad(qkv, (pad_l, pad_r, pad_t, pad_b))
.reshape(3, B, C, -1)
.transpose(-2, -1)
) # l r t b
_H = pad_b + H
_W = pad_r + W
_L = _H * _W
# window-0 and window-1 on split channels [C/2, C/2]; for square windows (e.g., 8x8), window-0 and window-1 can be merged
# shift in block: (0, 4, 8, ...), (2, 6, 10, ...), (0, 4, 8, ...), (2, 6, 10, ...), ...
if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or (
self.rg_idx % 2 != 0 and self.b_idx % 4 == 0
):
qkv = qkv.view(3, B, _H, _W, C)
qkv_0 = torch.roll(
qkv[:, :, :, :, : C // 2],
shifts=(-self.shift_size[0], -self.shift_size[1]),
dims=(2, 3),
)
qkv_0 = qkv_0.view(3, B, _L, C // 2)
qkv_1 = torch.roll(
qkv[:, :, :, :, C // 2 :],
shifts=(-self.shift_size[1], -self.shift_size[0]),
dims=(2, 3),
)
qkv_1 = qkv_1.view(3, B, _L, C // 2)
if self.patches_resolution != _H or self.patches_resolution != _W:
mask_tmp = self.calculate_mask(_H, _W)
x1_shift = self.attns[0](qkv_0, _H, _W, mask=mask_tmp[0].to(x.device))
x2_shift = self.attns[1](qkv_1, _H, _W, mask=mask_tmp[1].to(x.device))
else:
x1_shift = self.attns[0](qkv_0, _H, _W, mask=self.attn_mask_0)
x2_shift = self.attns[1](qkv_1, _H, _W, mask=self.attn_mask_1)
x1 = torch.roll(
x1_shift, shifts=(self.shift_size[0], self.shift_size[1]), dims=(1, 2)
)
x2 = torch.roll(
x2_shift, shifts=(self.shift_size[1], self.shift_size[0]), dims=(1, 2)
)
x1 = x1[:, :H, :W, :].reshape(B, L, C // 2)
x2 = x2[:, :H, :W, :].reshape(B, L, C // 2)
# attention output
attened_x = torch.cat([x1, x2], dim=2)
else:
x1 = self.attns[0](qkv[:, :, :, : C // 2], _H, _W)[:, :H, :W, :].reshape(
B, L, C // 2
)
x2 = self.attns[1](qkv[:, :, :, C // 2 :], _H, _W)[:, :H, :W, :].reshape(
B, L, C // 2
)
# attention output
attened_x = torch.cat([x1, x2], dim=2)
# convolution output
conv_x = self.dwconv(v)
# Adaptive Interaction Module (AIM)
# C-Map (before sigmoid)
channel_map = (
self.channel_interaction(conv_x)
.permute(0, 2, 3, 1)
.contiguous()
.view(B, 1, C)
)
# S-Map (before sigmoid)
attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W)
spatial_map = self.spatial_interaction(attention_reshape)
# C-I
attened_x = attened_x * torch.sigmoid(channel_map)
# S-I
conv_x = torch.sigmoid(spatial_map) * conv_x
conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, L, C)
x = attened_x + conv_x
x = self.proj(x)
x = self.proj_drop(x)
return x
class Adaptive_Channel_Attention(nn.Module):
# The implementation builds on XCiT code https://github.com/facebookresearch/xcit
"""Adaptive Channel Self-Attention
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads. Default: 6
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set.
attn_drop (float): Attention dropout rate. Default: 0.0
drop_path (float): Stochastic depth rate. Default: 0.0
"""
def __init__(
self,
dim,
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.num_heads = num_heads
self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1))
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)
self.dwconv = nn.Sequential(
nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim),
nn.BatchNorm2d(dim),
nn.GELU(),
)
self.channel_interaction = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(dim, dim // 8, kernel_size=1),
nn.BatchNorm2d(dim // 8),
nn.GELU(),
nn.Conv2d(dim // 8, dim, kernel_size=1),
)
self.spatial_interaction = nn.Sequential(
nn.Conv2d(dim, dim // 16, kernel_size=1),
nn.BatchNorm2d(dim // 16),
nn.GELU(),
nn.Conv2d(dim // 16, 1, kernel_size=1),
)
def forward(self, x, H, W):
"""
Input: x: (B, H*W, C), H, W
Output: x: (B, H*W, C)
"""
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
qkv = qkv.permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
q = q.transpose(-2, -1)
k = k.transpose(-2, -1)
v = v.transpose(-2, -1)
v_ = v.reshape(B, C, N).contiguous().view(B, C, H, W)
q = torch.nn.functional.normalize(q, dim=-1)
k = torch.nn.functional.normalize(k, dim=-1)
attn = (q @ k.transpose(-2, -1)) * self.temperature
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
# attention output
attened_x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C)
# convolution output
conv_x = self.dwconv(v_)
# Adaptive Interaction Module (AIM)
# C-Map (before sigmoid)
attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W)
channel_map = self.channel_interaction(attention_reshape)
# S-Map (before sigmoid)
spatial_map = (
self.spatial_interaction(conv_x)
.permute(0, 2, 3, 1)
.contiguous()
.view(B, N, 1)
)
# S-I
attened_x = attened_x * torch.sigmoid(spatial_map)
# C-I
conv_x = conv_x * torch.sigmoid(channel_map)
conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, N, C)
x = attened_x + conv_x
x = self.proj(x)
x = self.proj_drop(x)
return x
class DATB(nn.Module):
def __init__(
self,
dim,
num_heads,
reso=64,
split_size=[2, 4],
shift_size=[1, 2],
expansion_factor=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
rg_idx=0,
b_idx=0,
):
super().__init__()
self.norm1 = norm_layer(dim)
if b_idx % 2 == 0:
# DSTB
self.attn = Adaptive_Spatial_Attention(
dim,
num_heads=num_heads,
reso=reso,
split_size=split_size,
shift_size=shift_size,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
rg_idx=rg_idx,
b_idx=b_idx,
)
else:
# DCTB
self.attn = Adaptive_Channel_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 > 0.0 else nn.Identity()
ffn_hidden_dim = int(dim * expansion_factor)
self.ffn = SGFN(
in_features=dim,
hidden_features=ffn_hidden_dim,
out_features=dim,
act_layer=act_layer,
)
self.norm2 = norm_layer(dim)
def forward(self, x, x_size):
"""
Input: x: (B, H*W, C), x_size: (H, W)
Output: x: (B, H*W, C)
"""
H, W = x_size
x = x + self.drop_path(self.attn(self.norm1(x), H, W))
x = x + self.drop_path(self.ffn(self.norm2(x), H, W))
return x
class ResidualGroup(nn.Module):
"""ResidualGroup
Args:
dim (int): Number of input channels.
reso (int): Input resolution.
num_heads (int): Number of attention heads.
split_size (tuple(int)): Height and Width of spatial window.
expansion_factor (float): Ratio of ffn hidden dim to embedding dim.
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None
drop (float): Dropout rate. Default: 0
attn_drop(float): Attention dropout rate. Default: 0
drop_paths (float | None): Stochastic depth rate.
act_layer (nn.Module): Activation layer. Default: nn.GELU
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm
depth (int): Number of dual aggregation Transformer blocks in residual group.
use_chk (bool): Whether to use checkpointing to save memory.
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(
self,
dim,
reso,
num_heads,
split_size=[2, 4],
expansion_factor=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_paths=None,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
depth=2,
use_chk=False,
resi_connection="1conv",
rg_idx=0,
):
super().__init__()
self.use_chk = use_chk
self.reso = reso
self.blocks = nn.ModuleList(
[
DATB(
dim=dim,
num_heads=num_heads,
reso=reso,
split_size=split_size,
shift_size=[split_size[0] // 2, split_size[1] // 2],
expansion_factor=expansion_factor,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_paths[i],
act_layer=act_layer,
norm_layer=norm_layer,
rg_idx=rg_idx,
b_idx=i,
)
for i in range(depth)
]
)
if resi_connection == "1conv":
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == "3conv":
self.conv = nn.Sequential(
nn.Conv2d(dim, dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim, 3, 1, 1),
)
def forward(self, x, x_size):
"""
Input: x: (B, H*W, C), x_size: (H, W)
Output: x: (B, H*W, C)
"""
H, W = x_size
res = x
for blk in self.blocks:
if self.use_chk:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W)
x = self.conv(x)
x = rearrange(x, "b c h w -> b (h w) c")
x = res + x
return x
class Upsample(nn.Sequential):
"""Upsample module.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat):
m = []
if (scale & (scale - 1)) == 0: # scale = 2^n
for _ in range(int(math.log(scale, 2))):
m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(2))
elif scale == 3:
m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(3))
else:
raise ValueError(
f"scale {scale} is not supported. " "Supported scales: 2^n and 3."
)
super(Upsample, self).__init__(*m)
class UpsampleOneStep(nn.Sequential):
"""UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
Used in lightweight SR to save parameters.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
self.num_feat = num_feat
self.input_resolution = input_resolution
m = []
m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1))
m.append(nn.PixelShuffle(scale))
super(UpsampleOneStep, self).__init__(*m)
def flops(self):
h, w = self.input_resolution
flops = h * w * self.num_feat * 3 * 9
return flops
class DAT(nn.Module):
"""Dual Aggregation Transformer
Args:
img_size (int): Input image size. Default: 64
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 180
depths (tuple(int)): Depth of each residual group (number of DATB in each RG).
split_size (tuple(int)): Height and Width of spatial window.
num_heads (tuple(int)): Number of attention heads in different residual groups.
expansion_factor (float): Ratio of ffn hidden dim to embedding dim. Default: 4
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None
drop_rate (float): Dropout rate. Default: 0
attn_drop_rate (float): Attention dropout rate. Default: 0
drop_path_rate (float): Stochastic depth rate. Default: 0.1
act_layer (nn.Module): Activation layer. Default: nn.GELU
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm
use_chk (bool): Whether to use checkpointing to save memory.
upscale: Upscale factor. 2/3/4 for image SR
img_range: Image range. 1. or 255.
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(self, state_dict):
super().__init__()
# defaults
img_size = 64
in_chans = 3
embed_dim = 180
split_size = [2, 4]
depth = [2, 2, 2, 2]
num_heads = [2, 2, 2, 2]
expansion_factor = 4.0
qkv_bias = True
qk_scale = None
drop_rate = 0.0
attn_drop_rate = 0.0
drop_path_rate = 0.1
act_layer = nn.GELU
norm_layer = nn.LayerNorm
use_chk = False
upscale = 2
img_range = 1.0
resi_connection = "1conv"
upsampler = "pixelshuffle"
self.model_arch = "DAT"
self.sub_type = "SR"
self.state = state_dict
state_keys = state_dict.keys()
if "conv_before_upsample.0.weight" in state_keys:
if "conv_up1.weight" in state_keys:
upsampler = "nearest+conv"
else:
upsampler = "pixelshuffle"
supports_fp16 = False
elif "upsample.0.weight" in state_keys:
upsampler = "pixelshuffledirect"
else:
upsampler = ""
num_feat = (
state_dict.get("conv_before_upsample.0.weight", None).shape[1]
if state_dict.get("conv_before_upsample.weight", None)
else 64
)
num_in_ch = state_dict["conv_first.weight"].shape[1]
in_chans = num_in_ch
if "conv_last.weight" in state_keys:
num_out_ch = state_dict["conv_last.weight"].shape[0]
else:
num_out_ch = num_in_ch
upscale = 1
if upsampler == "nearest+conv":
upsample_keys = [
x for x in state_keys if "conv_up" in x and "bias" not in x
]
for upsample_key in upsample_keys:
upscale *= 2
elif upsampler == "pixelshuffle":
upsample_keys = [
x
for x in state_keys
if "upsample" in x and "conv" not in x and "bias" not in x
]
for upsample_key in upsample_keys:
shape = state_dict[upsample_key].shape[0]
upscale *= math.sqrt(shape // num_feat)
upscale = int(upscale)
elif upsampler == "pixelshuffledirect":
upscale = int(
math.sqrt(state_dict["upsample.0.bias"].shape[0] // num_out_ch)
)
max_layer_num = 0
max_block_num = 0
for key in state_keys:
result = re.match(r"layers.(\d*).blocks.(\d*).norm1.weight", key)
if result:
layer_num, block_num = result.groups()
max_layer_num = max(max_layer_num, int(layer_num))
max_block_num = max(max_block_num, int(block_num))
depth = [max_block_num + 1 for _ in range(max_layer_num + 1)]
if "layers.0.blocks.1.attn.temperature" in state_keys:
num_heads_num = state_dict["layers.0.blocks.1.attn.temperature"].shape[0]
num_heads = [num_heads_num for _ in range(max_layer_num + 1)]
else:
num_heads = depth
embed_dim = state_dict["conv_first.weight"].shape[0]
expansion_factor = float(
state_dict["layers.0.blocks.0.ffn.fc1.weight"].shape[0] / embed_dim
)
# TODO: could actually count the layers, but this should do
if "layers.0.conv.4.weight" in state_keys:
resi_connection = "3conv"
else:
resi_connection = "1conv"
if "layers.0.blocks.2.attn.attn_mask_0" in state_keys:
attn_mask_0_x, attn_mask_0_y, attn_mask_0_z = state_dict[
"layers.0.blocks.2.attn.attn_mask_0"
].shape
img_size = int(math.sqrt(attn_mask_0_x * attn_mask_0_y))
if "layers.0.blocks.0.attn.attns.0.rpe_biases" in state_keys:
split_sizes = (
state_dict["layers.0.blocks.0.attn.attns.0.rpe_biases"][-1] + 1
)
split_size = [int(x) for x in split_sizes]
self.in_nc = num_in_ch
self.out_nc = num_out_ch
self.num_feat = num_feat
self.embed_dim = embed_dim
self.num_heads = num_heads
self.depth = depth
self.scale = upscale
self.upsampler = upsampler
self.img_size = img_size
self.img_range = img_range
self.expansion_factor = expansion_factor
self.resi_connection = resi_connection
self.split_size = split_size
self.supports_fp16 = False # Too much weirdness to support this at the moment
self.supports_bfp16 = True
self.min_size_restriction = 16
num_in_ch = in_chans
num_out_ch = in_chans
num_feat = 64
self.img_range = img_range
if in_chans == 3:
rgb_mean = (0.4488, 0.4371, 0.4040)
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
else:
self.mean = torch.zeros(1, 1, 1, 1)
self.upscale = upscale
self.upsampler = upsampler
# ------------------------- 1, Shallow Feature Extraction ------------------------- #
self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
# ------------------------- 2, Deep Feature Extraction ------------------------- #
self.num_layers = len(depth)
self.use_chk = use_chk
self.num_features = (
self.embed_dim
) = embed_dim # num_features for consistency with other models
heads = num_heads
self.before_RG = nn.Sequential(
Rearrange("b c h w -> b (h w) c"), nn.LayerNorm(embed_dim)
)
curr_dim = embed_dim
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, np.sum(depth))
] # stochastic depth decay rule
self.layers = nn.ModuleList()
for i in range(self.num_layers):
layer = ResidualGroup(
dim=embed_dim,
num_heads=heads[i],
reso=img_size,
split_size=split_size,
expansion_factor=expansion_factor,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_paths=dpr[sum(depth[:i]) : sum(depth[: i + 1])],
act_layer=act_layer,
norm_layer=norm_layer,
depth=depth[i],
use_chk=use_chk,
resi_connection=resi_connection,
rg_idx=i,
)
self.layers.append(layer)
self.norm = norm_layer(curr_dim)
# build the last conv layer in deep feature extraction
if resi_connection == "1conv":
self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
elif resi_connection == "3conv":
# to save parameters and memory
self.conv_after_body = nn.Sequential(
nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1),
)
# ------------------------- 3, Reconstruction ------------------------- #
if self.upsampler == "pixelshuffle":
# for classical SR
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.upsample = Upsample(upscale, num_feat)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
elif self.upsampler == "pixelshuffledirect":
# for lightweight SR (to save parameters)
self.upsample = UpsampleOneStep(
upscale, embed_dim, num_out_ch, (img_size, img_size)
)
self.apply(self._init_weights)
self.load_state_dict(state_dict, strict=True)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(
m, (nn.LayerNorm, nn.BatchNorm2d, nn.GroupNorm, nn.InstanceNorm2d)
):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward_features(self, x):
_, _, H, W = x.shape
x_size = [H, W]
x = self.before_RG(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x)
x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W)
return x
def forward(self, x):
"""
Input: x: (B, C, H, W)
"""
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == "pixelshuffle":
# for image SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.conv_last(self.upsample(x))
elif self.upsampler == "pixelshuffledirect":
# for lightweight SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.upsample(x)
x = x / self.img_range + self.mean
return x