Skip to content

GitLab

Commit 905857ed authored by comfyanonymous's avatar comfyanonymous
Browse files

Take some code from chainner to implement ESRGAN and other upscale models.

parent 8c4ccb55
Hide whitespace changes
Inline Side-by-side
Showing with 8469 additions and 0 deletions
comfy_extras/chainner_models/architecture/HAT.py 0 → 100644
View file @ 905857ed
# pylint: skip-file
# HAT from https://github.com/XPixelGroup/HAT/blob/main/hat/archs/hat_arch.py
import math
import re
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from .timm.helpers import to_2tuple
from .timm.weight_init import trunc_normal_
def drop_path(x, drop_prob: float = 0.0, training: bool = False):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py
"""
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (
x.ndim - 1
) # work with diff dim tensors, not just 2D ConvNets
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
random_tensor.floor_() # binarize
output = x.div(keep_prob) * random_tensor
return output
class DropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py
"""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training) # type: ignore
class ChannelAttention(nn.Module):
"""Channel attention used in RCAN.
Args:
num_feat (int): Channel number of intermediate features.
squeeze_factor (int): Channel squeeze factor. Default: 16.
"""
def __init__(self, num_feat, squeeze_factor=16):
super(ChannelAttention, self).__init__()
self.attention = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(num_feat, num_feat // squeeze_factor, 1, padding=0),
nn.ReLU(inplace=True),
nn.Conv2d(num_feat // squeeze_factor, 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=3, squeeze_factor=30):
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, squeeze_factor),
)
def forward(self, x):
return self.cab(x)
class Mlp(nn.Module):
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.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 window_partition(x, window_size):
"""
Args:
x: (b, h, w, c)
window_size (int): window size
Returns:
windows: (num_windows*b, window_size, window_size, c)
"""
b, h, w, c = x.shape
x = x.view(b, h // window_size, window_size, w // window_size, window_size, c)
windows = (
x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, c)
)
return windows
def window_reverse(windows, window_size, h, w):
"""
Args:
windows: (num_windows*b, window_size, window_size, c)
window_size (int): Window size
h (int): Height of image
w (int): Width of image
Returns:
x: (b, h, w, c)
"""
b = int(windows.shape[0] / (h * w / window_size / window_size))
x = windows.view(
b, h // window_size, w // window_size, window_size, window_size, -1
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(b, h, w, -1)
return x
class WindowAttention(nn.Module):
r"""Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
"""
def __init__(
self,
dim,
window_size,
num_heads,
qkv_bias=True,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim**-0.5
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter( # type: ignore
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
) # 2*Wh-1 * 2*Ww-1, nH
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)
trunc_normal_(self.relative_position_bias_table, std=0.02)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, rpi, mask=None):
"""
Args:
x: input features with shape of (num_windows*b, n, c)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
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],
) # make torchscript happy (cannot use tensor as tuple)
q = q * self.scale
attn = q @ k.transpose(-2, -1)
relative_position_bias = self.relative_position_bias_table[rpi.view(-1)].view(
self.window_size[0] * self.window_size[1],
self.window_size[0] * self.window_size[1],
-1,
) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous() # nH, Wh*Ww, Wh*Ww
attn = attn + relative_position_bias.unsqueeze(0)
if mask is not None:
nw = mask.shape[0]
attn = attn.view(b_ // nw, nw, self.num_heads, n, n) + mask.unsqueeze(
1
).unsqueeze(0)
attn = attn.view(-1, self.num_heads, n, n)
attn = self.softmax(attn)
else:
attn = self.softmax(attn)
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
class HAB(nn.Module):
r"""Hybrid Attention Block.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
num_heads (int): Number of attention heads.
window_size (int): Window size.
shift_size (int): Shift size for SW-MSA.
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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
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
"""
def __init__(
self,
dim,
input_resolution,
num_heads,
window_size=7,
shift_size=0,
compress_ratio=3,
squeeze_factor=30,
conv_scale=0.01,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
):
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.conv_scale = conv_scale
self.conv_block = CAB(
num_feat=dim, compress_ratio=compress_ratio, squeeze_factor=squeeze_factor
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 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_size, rpi_sa, attn_mask):
h, w = x_size
b, _, c = x.shape
# assert seq_len == h * w, "input feature has wrong size"
shortcut = x
x = self.norm1(x)
x = x.view(b, h, w, c)
# Conv_X
conv_x = self.conv_block(x.permute(0, 3, 1, 2))
conv_x = conv_x.permute(0, 2, 3, 1).contiguous().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)
)
attn_mask = attn_mask
else:
shifted_x = x
attn_mask = None
# 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 (to be compatible for testing on images whose shapes are the multiple of window size
attn_windows = self.attn(x_windows, rpi=rpi_sa, mask=attn_mask)
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, c)
shifted_x = window_reverse(attn_windows, self.window_size, h, w) # b h' w' c
# reverse cyclic shift
if self.shift_size > 0:
attn_x = torch.roll(
shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
)
else:
attn_x = shifted_x
attn_x = attn_x.view(b, h * w, c)
# FFN
x = shortcut + self.drop_path(attn_x) + conv_x * self.conv_scale
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
class PatchMerging(nn.Module):
r"""Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: b, h*w, c
"""
h, w = self.input_resolution
b, seq_len, c = x.shape
assert seq_len == 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.norm(x)
x = self.reduction(x)
return x
class OCAB(nn.Module):
# overlapping cross-attention block
def __init__(
self,
dim,
input_resolution,
window_size,
overlap_ratio,
num_heads,
qkv_bias=True,
qk_scale=None,
mlp_ratio=2,
norm_layer=nn.LayerNorm,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.window_size = window_size
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim**-0.5
self.overlap_win_size = int(window_size * overlap_ratio) + window_size
self.norm1 = norm_layer(dim)
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.unfold = nn.Unfold(
kernel_size=(self.overlap_win_size, self.overlap_win_size),
stride=window_size,
padding=(self.overlap_win_size - window_size) // 2,
)
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter( # type: ignore
torch.zeros(
(window_size + self.overlap_win_size - 1)
* (window_size + self.overlap_win_size - 1),
num_heads,
)
) # 2*Wh-1 * 2*Ww-1, nH
trunc_normal_(self.relative_position_bias_table, std=0.02)
self.softmax = nn.Softmax(dim=-1)
self.proj = nn.Linear(dim, dim)
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=nn.GELU
)
def forward(self, x, x_size, rpi):
h, w = x_size
b, _, c = x.shape
shortcut = x
x = self.norm1(x)
x = x.view(b, h, w, c)
qkv = self.qkv(x).reshape(b, h, w, 3, c).permute(3, 0, 4, 1, 2) # 3, b, c, h, w
q = qkv[0].permute(0, 2, 3, 1) # b, h, w, c
kv = torch.cat((qkv[1], qkv[2]), dim=1) # b, 2*c, h, w
# partition windows
q_windows = window_partition(
q, self.window_size
) # nw*b, window_size, window_size, c
q_windows = q_windows.view(
-1, self.window_size * self.window_size, c
) # nw*b, window_size*window_size, c
kv_windows = self.unfold(kv) # b, c*w*w, nw
kv_windows = rearrange(
kv_windows,
"b (nc ch owh oww) nw -> nc (b nw) (owh oww) ch",
nc=2,
ch=c,
owh=self.overlap_win_size,
oww=self.overlap_win_size,
).contiguous() # 2, nw*b, ow*ow, c
# Do the above rearrangement without the rearrange function
# kv_windows = kv_windows.view(
# 2, b, self.overlap_win_size, self.overlap_win_size, c, -1
# )
# kv_windows = kv_windows.permute(0, 5, 1, 2, 3, 4).contiguous()
# kv_windows = kv_windows.view(
# 2, -1, self.overlap_win_size * self.overlap_win_size, c
# )
k_windows, v_windows = kv_windows[0], kv_windows[1] # nw*b, ow*ow, c
b_, nq, _ = q_windows.shape
_, n, _ = k_windows.shape
d = self.dim // self.num_heads
q = q_windows.reshape(b_, nq, self.num_heads, d).permute(
0, 2, 1, 3
) # nw*b, nH, nq, d
k = k_windows.reshape(b_, n, self.num_heads, d).permute(
0, 2, 1, 3
) # nw*b, nH, n, d
v = v_windows.reshape(b_, n, self.num_heads, d).permute(
0, 2, 1, 3
) # nw*b, nH, n, d
q = q * self.scale
attn = q @ k.transpose(-2, -1)
relative_position_bias = self.relative_position_bias_table[rpi.view(-1)].view(
self.window_size * self.window_size,
self.overlap_win_size * self.overlap_win_size,
-1,
) # ws*ws, wse*wse, nH
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous() # nH, ws*ws, wse*wse
attn = attn + relative_position_bias.unsqueeze(0)
attn = self.softmax(attn)
attn_windows = (attn @ v).transpose(1, 2).reshape(b_, nq, self.dim)
# merge windows
attn_windows = attn_windows.view(
-1, self.window_size, self.window_size, self.dim
)
x = window_reverse(attn_windows, self.window_size, h, w) # b h w c
x = x.view(b, h * w, self.dim)
x = self.proj(x) + shortcut
x = x + self.mlp(self.norm2(x))
return x
class AttenBlocks(nn.Module):
"""A series of attention blocks for one RHAG.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
compress_ratio,
squeeze_factor,
conv_scale,
overlap_ratio,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList(
[
HAB(
dim=dim,
input_resolution=input_resolution,
num_heads=num_heads,
window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
compress_ratio=compress_ratio,
squeeze_factor=squeeze_factor,
conv_scale=conv_scale,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i]
if isinstance(drop_path, list)
else drop_path,
norm_layer=norm_layer,
)
for i in range(depth)
]
)
# OCAB
self.overlap_attn = OCAB(
dim=dim,
input_resolution=input_resolution,
window_size=window_size,
overlap_ratio=overlap_ratio,
num_heads=num_heads,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
mlp_ratio=mlp_ratio, # type: ignore
norm_layer=norm_layer,
)
# patch merging layer
if downsample is not None:
self.downsample = downsample(
input_resolution, dim=dim, norm_layer=norm_layer
)
else:
self.downsample = None
def forward(self, x, x_size, params):
for blk in self.blocks:
x = blk(x, x_size, params["rpi_sa"], params["attn_mask"])
x = self.overlap_attn(x, x_size, params["rpi_oca"])
if self.downsample is not None:
x = self.downsample(x)
return x
class RHAG(nn.Module):
"""Residual Hybrid Attention Group (RHAG).
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
img_size: Input image size.
patch_size: Patch size.
resi_connection: The convolutional block before residual connection.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
compress_ratio,
squeeze_factor,
conv_scale,
overlap_ratio,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
img_size=224,
patch_size=4,
resi_connection="1conv",
):
super(RHAG, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.residual_group = AttenBlocks(
dim=dim,
input_resolution=input_resolution,
depth=depth,
num_heads=num_heads,
window_size=window_size,
compress_ratio=compress_ratio,
squeeze_factor=squeeze_factor,
conv_scale=conv_scale,
overlap_ratio=overlap_ratio,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path,
norm_layer=norm_layer,
downsample=downsample,
use_checkpoint=use_checkpoint,
)
if resi_connection == "1conv":
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == "identity":
self.conv = nn.Identity()
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=0,
embed_dim=dim,
norm_layer=None,
)
self.patch_unembed = PatchUnEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=0,
embed_dim=dim,
norm_layer=None,
)
def forward(self, x, x_size, params):
return (
self.patch_embed(
self.conv(
self.patch_unembed(self.residual_group(x, x_size, params), x_size)
)
)
+ x
)
class PatchEmbed(nn.Module):
r"""Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [
img_size[0] // patch_size[0], # type: ignore
img_size[1] // patch_size[1], # type: ignore
]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
x = x.flatten(2).transpose(1, 2) # b Ph*Pw c
if self.norm is not None:
x = self.norm(x)
return x
class PatchUnEmbed(nn.Module):
r"""Image to Patch Unembedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [
img_size[0] // patch_size[0], # type: ignore
img_size[1] // patch_size[1], # type: ignore
]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
x = (
x.transpose(1, 2)
.contiguous()
.view(x.shape[0], self.embed_dim, x_size[0], x_size[1])
) # b Ph*Pw c
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 HAT(nn.Module):
r"""Hybrid Attention Transformer
A PyTorch implementation of : `Activating More Pixels in Image Super-Resolution Transformer`.
Some codes are based on SwinIR.
Args:
img_size (int | tuple(int)): Input image size. Default 64
patch_size (int | tuple(int)): Patch size. Default: 1
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
window_size (int): Window size. Default: 7
mlp_ratio (float): Ratio of mlp 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): 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
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
patch_norm (bool): If True, add normalization after patch embedding. Default: True
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
img_range: Image range. 1. or 255.
upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(
self,
state_dict,
**kwargs,
):
super(HAT, self).__init__()
# Defaults
img_size = 64
patch_size = 1
in_chans = 3
embed_dim = 96
depths = (6, 6, 6, 6)
num_heads = (6, 6, 6, 6)
window_size = 7
compress_ratio = 3
squeeze_factor = 30
conv_scale = 0.01
overlap_ratio = 0.5
mlp_ratio = 4.0
qkv_bias = True
qk_scale = None
drop_rate = 0.0
attn_drop_rate = 0.0
drop_path_rate = 0.1
norm_layer = nn.LayerNorm
ape = False
patch_norm = True
use_checkpoint = False
upscale = 2
img_range = 1.0
upsampler = ""
resi_connection = "1conv"
self.state = state_dict
self.model_arch = "HAT"
self.sub_type = "SR"
self.supports_fp16 = False
self.support_bf16 = True
self.min_size_restriction = 16
state_keys = list(state_dict.keys())
num_feat = state_dict["conv_last.weight"].shape[1]
in_chans = state_dict["conv_first.weight"].shape[1]
num_out_ch = state_dict["conv_last.weight"].shape[0]
embed_dim = state_dict["conv_first.weight"].shape[0]
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 = ""
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 = self.state[upsample_key].shape[0]
upscale *= math.sqrt(shape // num_feat)
upscale = int(upscale)
elif upsampler == "pixelshuffledirect":
upscale = int(
math.sqrt(self.state["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*).residual_group.blocks.(\d*).conv_block.cab.0.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))
depths = [max_block_num + 1 for _ in range(max_layer_num + 1)]
if (
"layers.0.residual_group.blocks.0.attn.relative_position_bias_table"
in state_keys
):
num_heads_num = self.state[
"layers.0.residual_group.blocks.0.attn.relative_position_bias_table"
].shape[-1]
num_heads = [num_heads_num for _ in range(max_layer_num + 1)]
else:
num_heads = depths
mlp_ratio = float(
self.state["layers.0.residual_group.blocks.0.mlp.fc1.bias"].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"
window_size = int(math.sqrt(self.state["relative_position_index_SA"].shape[0]))
# Not sure if this is needed or used at all anywhere in HAT's config
if "layers.0.residual_group.blocks.1.attn_mask" in state_keys:
img_size = int(
math.sqrt(
self.state["layers.0.residual_group.blocks.1.attn_mask"].shape[0]
)
* window_size
)
self.window_size = window_size
self.shift_size = window_size // 2
self.overlap_ratio = overlap_ratio
self.in_nc = in_chans
self.out_nc = num_out_ch
self.num_feat = num_feat
self.embed_dim = embed_dim
self.num_heads = num_heads
self.depths = depths
self.window_size = window_size
self.mlp_ratio = mlp_ratio
self.scale = upscale
self.upsampler = upsampler
self.img_size = img_size
self.img_range = img_range
self.resi_connection = resi_connection
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
# relative position index
relative_position_index_SA = self.calculate_rpi_sa()
relative_position_index_OCA = self.calculate_rpi_oca()
self.register_buffer("relative_position_index_SA", relative_position_index_SA)
self.register_buffer("relative_position_index_OCA", relative_position_index_OCA)
# ------------------------- 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(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = embed_dim
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=embed_dim,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None,
)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# merge non-overlapping patches into image
self.patch_unembed = PatchUnEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=embed_dim,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None,
)
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter( # type: ignore[arg-type]
torch.zeros(1, num_patches, embed_dim)
)
trunc_normal_(self.absolute_pos_embed, std=0.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
] # stochastic depth decay rule
# build Residual Hybrid Attention Groups (RHAG)
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RHAG(
dim=embed_dim,
input_resolution=(patches_resolution[0], patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
compress_ratio=compress_ratio,
squeeze_factor=squeeze_factor,
conv_scale=conv_scale,
overlap_ratio=overlap_ratio,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[
sum(depths[:i_layer]) : sum(depths[: i_layer + 1]) # type: ignore
], # no impact on SR results
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection,
)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
# 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 == "identity":
self.conv_after_body = nn.Identity()
# ------------------------- 3, high quality image 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)
self.apply(self._init_weights)
self.load_state_dict(self.state, strict=False)
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.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def calculate_rpi_sa(self):
# calculate relative position index for SA
coords_h = torch.arange(self.window_size)
coords_w = torch.arange(self.window_size)
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
relative_coords = (
coords_flatten[:, :, None] - coords_flatten[:, None, :]
) # 2, Wh*Ww, Wh*Ww
relative_coords = relative_coords.permute(
1, 2, 0
).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size - 1
relative_coords[:, :, 0] *= 2 * self.window_size - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
return relative_position_index
def calculate_rpi_oca(self):
# calculate relative position index for OCA
window_size_ori = self.window_size
window_size_ext = self.window_size + int(self.overlap_ratio * self.window_size)
coords_h = torch.arange(window_size_ori)
coords_w = torch.arange(window_size_ori)
coords_ori = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, ws, ws
coords_ori_flatten = torch.flatten(coords_ori, 1) # 2, ws*ws
coords_h = torch.arange(window_size_ext)
coords_w = torch.arange(window_size_ext)
coords_ext = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, wse, wse
coords_ext_flatten = torch.flatten(coords_ext, 1) # 2, wse*wse
relative_coords = (
coords_ext_flatten[:, None, :] - coords_ori_flatten[:, :, None]
) # 2, ws*ws, wse*wse
relative_coords = relative_coords.permute(
1, 2, 0
).contiguous() # ws*ws, wse*wse, 2
relative_coords[:, :, 0] += (
window_size_ori - window_size_ext + 1
) # shift to start from 0
relative_coords[:, :, 1] += window_size_ori - window_size_ext + 1
relative_coords[:, :, 0] *= window_size_ori + window_size_ext - 1
relative_position_index = relative_coords.sum(-1)
return relative_position_index
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
h, w = x_size
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)
)
return attn_mask
@torch.jit.ignore # type: ignore
def no_weight_decay(self):
return {"absolute_pos_embed"}
@torch.jit.ignore # type: ignore
def no_weight_decay_keywords(self):
return {"relative_position_bias_table"}
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
return x
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
# Calculate attention mask and relative position index in advance to speed up inference.
# The original code is very time-cosuming for large window size.
attn_mask = self.calculate_mask(x_size).to(x.device)
params = {
"attn_mask": attn_mask,
"rpi_sa": self.relative_position_index_SA,
"rpi_oca": self.relative_position_index_OCA,
}
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x, x_size, params)
x = self.norm(x) # b seq_len c
x = self.patch_unembed(x, x_size)
return x
def forward(self, x):
H, W = x.shape[2:]
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
x = self.check_image_size(x)
if self.upsampler == "pixelshuffle":
# for classical 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))
x = x / self.img_range + self.mean
return x[:, :, : H * self.upscale, : W * self.upscale]
comfy_extras/chainner_models/architecture/LICENSE-ESRGAN 0 → 100644
View file @ 905857ed
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
comfy_extras/chainner_models/architecture/LICENSE-HAT 0 → 100644
View file @ 905857ed
MIT License
Copyright (c) 2022 Xiangyu Chen
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
comfy_extras/chainner_models/architecture/LICENSE-RealESRGAN 0 → 100644
View file @ 905857ed
BSD 3-Clause License
Copyright (c) 2021, Xintao Wang
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
1. Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
3. Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
comfy_extras/chainner_models/architecture/LICENSE-SPSR 0 → 100644
View file @ 905857ed
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright 2018-2022 BasicSR Authors
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
comfy_extras/chainner_models/architecture/LICENSE-SwiftSRGAN 0 → 100644
View file @ 905857ed
Creative Commons Legal Code
CC0 1.0 Universal
CREATIVE COMMONS CORPORATION IS NOT A LAW FIRM AND DOES NOT PROVIDE
LEGAL SERVICES. DISTRIBUTION OF THIS DOCUMENT DOES NOT CREATE AN
ATTORNEY-CLIENT RELATIONSHIP. CREATIVE COMMONS PROVIDES THIS
INFORMATION ON AN "AS-IS" BASIS. CREATIVE COMMONS MAKES NO WARRANTIES
REGARDING THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS
PROVIDED HEREUNDER, AND DISCLAIMS LIABILITY FOR DAMAGES RESULTING FROM
THE USE OF THIS DOCUMENT OR THE INFORMATION OR WORKS PROVIDED
HEREUNDER.
Statement of Purpose
The laws of most jurisdictions throughout the world automatically confer
exclusive Copyright and Related Rights (defined below) upon the creator
and subsequent owner(s) (each and all, an "owner") of an original work of
authorship and/or a database (each, a "Work").
Certain owners wish to permanently relinquish those rights to a Work for
the purpose of contributing to a commons of creative, cultural and
scientific works ("Commons") that the public can reliably and without fear
of later claims of infringement build upon, modify, incorporate in other
works, reuse and redistribute as freely as possible in any form whatsoever
and for any purposes, including without limitation commercial purposes.
These owners may contribute to the Commons to promote the ideal of a free
culture and the further production of creative, cultural and scientific
works, or to gain reputation or greater distribution for their Work in
part through the use and efforts of others.
For these and/or other purposes and motivations, and without any
expectation of additional consideration or compensation, the person
associating CC0 with a Work (the "Affirmer"), to the extent that he or she
is an owner of Copyright and Related Rights in the Work, voluntarily
elects to apply CC0 to the Work and publicly distribute the Work under its
terms, with knowledge of his or her Copyright and Related Rights in the
Work and the meaning and intended legal effect of CC0 on those rights.
1. Copyright and Related Rights. A Work made available under CC0 may be
protected by copyright and related or neighboring rights ("Copyright and
Related Rights"). Copyright and Related Rights include, but are not
limited to, the following:
i. the right to reproduce, adapt, distribute, perform, display,
communicate, and translate a Work;
ii. moral rights retained by the original author(s) and/or performer(s);
iii. publicity and privacy rights pertaining to a person's image or
likeness depicted in a Work;
iv. rights protecting against unfair competition in regards to a Work,
subject to the limitations in paragraph 4(a), below;
v. rights protecting the extraction, dissemination, use and reuse of data
in a Work;
vi. database rights (such as those arising under Directive 96/9/EC of the
European Parliament and of the Council of 11 March 1996 on the legal
protection of databases, and under any national implementation
thereof, including any amended or successor version of such
directive); and
vii. other similar, equivalent or corresponding rights throughout the
world based on applicable law or treaty, and any national
implementations thereof.
2. Waiver. To the greatest extent permitted by, but not in contravention
of, applicable law, Affirmer hereby overtly, fully, permanently,
irrevocably and unconditionally waives, abandons, and surrenders all of
Affirmer's Copyright and Related Rights and associated claims and causes
of action, whether now known or unknown (including existing as well as
future claims and causes of action), in the Work (i) in all territories
worldwide, (ii) for the maximum duration provided by applicable law or
treaty (including future time extensions), (iii) in any current or future
medium and for any number of copies, and (iv) for any purpose whatsoever,
including without limitation commercial, advertising or promotional
purposes (the "Waiver"). Affirmer makes the Waiver for the benefit of each
member of the public at large and to the detriment of Affirmer's heirs and
successors, fully intending that such Waiver shall not be subject to
revocation, rescission, cancellation, termination, or any other legal or
equitable action to disrupt the quiet enjoyment of the Work by the public
as contemplated by Affirmer's express Statement of Purpose.
3. Public License Fallback. Should any part of the Waiver for any reason
be judged legally invalid or ineffective under applicable law, then the
Waiver shall be preserved to the maximum extent permitted taking into
account Affirmer's express Statement of Purpose. In addition, to the
extent the Waiver is so judged Affirmer hereby grants to each affected
person a royalty-free, non transferable, non sublicensable, non exclusive,
irrevocable and unconditional license to exercise Affirmer's Copyright and
Related Rights in the Work (i) in all territories worldwide, (ii) for the
maximum duration provided by applicable law or treaty (including future
time extensions), (iii) in any current or future medium and for any number
of copies, and (iv) for any purpose whatsoever, including without
limitation commercial, advertising or promotional purposes (the
"License"). The License shall be deemed effective as of the date CC0 was
applied by Affirmer to the Work. Should any part of the License for any
reason be judged legally invalid or ineffective under applicable law, such
partial invalidity or ineffectiveness shall not invalidate the remainder
of the License, and in such case Affirmer hereby affirms that he or she
will not (i) exercise any of his or her remaining Copyright and Related
Rights in the Work or (ii) assert any associated claims and causes of
action with respect to the Work, in either case contrary to Affirmer's
express Statement of Purpose.
4. Limitations and Disclaimers.
a. No trademark or patent rights held by Affirmer are waived, abandoned,
surrendered, licensed or otherwise affected by this document.
b. Affirmer offers the Work as-is and makes no representations or
warranties of any kind concerning the Work, express, implied,
statutory or otherwise, including without limitation warranties of
title, merchantability, fitness for a particular purpose, non
infringement, or the absence of latent or other defects, accuracy, or
the present or absence of errors, whether or not discoverable, all to
the greatest extent permissible under applicable law.
c. Affirmer disclaims responsibility for clearing rights of other persons
that may apply to the Work or any use thereof, including without
limitation any person's Copyright and Related Rights in the Work.
Further, Affirmer disclaims responsibility for obtaining any necessary
consents, permissions or other rights required for any use of the
Work.
d. Affirmer understands and acknowledges that Creative Commons is not a
party to this document and has no duty or obligation with respect to
this CC0 or use of the Work.
comfy_extras/chainner_models/architecture/LICENSE-Swin2SR 0 → 100644
View file @ 905857ed
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [2021] [SwinIR Authors]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
comfy_extras/chainner_models/architecture/LICENSE-SwinIR 0 → 100644
View file @ 905857ed
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [2021] [SwinIR Authors]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
comfy_extras/chainner_models/architecture/LICENSE-lama 0 → 100644
View file @ 905857ed
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [2021] Samsung Research
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
comfy_extras/chainner_models/architecture/LICENSE-mat 0 → 100644
View file @ 905857ed
## creative commons
# Attribution-NonCommercial 4.0 International
Creative Commons Corporation (“Creative Commons”) is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an “as-is” basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible.
### Using Creative Commons Public Licenses
Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses.
* __Considerations for licensors:__ Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC-licensed material, or material used under an exception or limitation to copyright. [More considerations for licensors](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensors).
* __Considerations for the public:__ By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor’s permission is not necessary for any reason–for example, because of any applicable exception or limitation to copyright–then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. [More considerations for the public](http://wiki.creativecommons.org/Considerations_for_licensors_and_licensees#Considerations_for_licensees).
## Creative Commons Attribution-NonCommercial 4.0 International Public License
By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions.
### Section 1 – Definitions.
a. __Adapted Material__ means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image.
b. __Adapter's License__ means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License.
c. __Copyright and Similar Rights__ means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights.
d. __Effective Technological Measures__ means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements.
e. __Exceptions and Limitations__ means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material.
f. __Licensed Material__ means the artistic or literary work, database, or other material to which the Licensor applied this Public License.
g. __Licensed Rights__ means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license.
h. __Licensor__ means the individual(s) or entity(ies) granting rights under this Public License.
i. __NonCommercial__ means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange.
j. __Share__ means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them.
k. __Sui Generis Database Rights__ means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world.
l. __You__ means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning.
### Section 2 – Scope.
a. ___License grant.___
1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to:
A. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and
B. produce, reproduce, and Share Adapted Material for NonCommercial purposes only.
2. __Exceptions and Limitations.__ For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions.
3. __Term.__ The term of this Public License is specified in Section 6(a).
4. __Media and formats; technical modifications allowed.__ The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a)(4) never produces Adapted Material.
5. __Downstream recipients.__
A. __Offer from the Licensor – Licensed Material.__ Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License.
B. __No downstream restrictions.__ You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material.
6. __No endorsement.__ Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i).
b. ___Other rights.___
1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise.
2. Patent and trademark rights are not licensed under this Public License.
3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes.
### Section 3 – License Conditions.
Your exercise of the Licensed Rights is expressly made subject to the following conditions.
a. ___Attribution.___
1. If You Share the Licensed Material (including in modified form), You must:
A. retain the following if it is supplied by the Licensor with the Licensed Material:
i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated);
ii. a copyright notice;
iii. a notice that refers to this Public License;
iv. a notice that refers to the disclaimer of warranties;
v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable;
B. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and
C. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License.
2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information.
3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable.
4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License.
### Section 4 – Sui Generis Database Rights.
Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material:
a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only;
b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and
c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database.
For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights.
### Section 5 – Disclaimer of Warranties and Limitation of Liability.
a. __Unless otherwise separately undertaken by the Licensor, to the extent possible, the Licensor offers the Licensed Material as-is and as-available, and makes no representations or warranties of any kind concerning the Licensed Material, whether express, implied, statutory, or other. This includes, without limitation, warranties of title, merchantability, fitness for a particular purpose, non-infringement, absence of latent or other defects, accuracy, or the presence or absence of errors, whether or not known or discoverable. Where disclaimers of warranties are not allowed in full or in part, this disclaimer may not apply to You.__
b. __To the extent possible, in no event will the Licensor be liable to You on any legal theory (including, without limitation, negligence) or otherwise for any direct, special, indirect, incidental, consequential, punitive, exemplary, or other losses, costs, expenses, or damages arising out of this Public License or use of the Licensed Material, even if the Licensor has been advised of the possibility of such losses, costs, expenses, or damages. Where a limitation of liability is not allowed in full or in part, this limitation may not apply to You.__
c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability.
### Section 6 – Term and Termination.
a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically.
b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates:
1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or
2. upon express reinstatement by the Licensor.
For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License.
c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License.
d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License.
### Section 7 – Other Terms and Conditions.
a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed.
b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License.
### Section 8 – Interpretation.
a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License.
b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions.
c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor.
d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority.
> Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at [creativecommons.org/policies](http://creativecommons.org/policies), Creative Commons does not authorize the use of the trademark “Creative Commons” or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses.
>
> Creative Commons may be contacted at creativecommons.org
comfy_extras/chainner_models/architecture/LaMa.py 0 → 100644
View file @ 905857ed
# pylint: skip-file
"""
Model adapted from advimman's lama project: https://github.com/advimman/lama
"""
# Fast Fourier Convolution NeurIPS 2020
# original implementation https://github.com/pkumivision/FFC/blob/main/model_zoo/ffc.py
# paper https://proceedings.neurips.cc/paper/2020/file/2fd5d41ec6cfab47e32164d5624269b1-Paper.pdf
from typing import List
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.transforms.functional import InterpolationMode, rotate
class LearnableSpatialTransformWrapper(nn.Module):
def __init__(self, impl, pad_coef=0.5, angle_init_range=80, train_angle=True):
super().__init__()
self.impl = impl
self.angle = torch.rand(1) * angle_init_range
if train_angle:
self.angle = nn.Parameter(self.angle, requires_grad=True)
self.pad_coef = pad_coef
def forward(self, x):
if torch.is_tensor(x):
return self.inverse_transform(self.impl(self.transform(x)), x)
elif isinstance(x, tuple):
x_trans = tuple(self.transform(elem) for elem in x)
y_trans = self.impl(x_trans)
return tuple(
self.inverse_transform(elem, orig_x) for elem, orig_x in zip(y_trans, x)
)
else:
raise ValueError(f"Unexpected input type {type(x)}")
def transform(self, x):
height, width = x.shape[2:]
pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef)
x_padded = F.pad(x, [pad_w, pad_w, pad_h, pad_h], mode="reflect")
x_padded_rotated = rotate(
x_padded, self.angle.to(x_padded), InterpolationMode.BILINEAR, fill=0
)
return x_padded_rotated
def inverse_transform(self, y_padded_rotated, orig_x):
height, width = orig_x.shape[2:]
pad_h, pad_w = int(height * self.pad_coef), int(width * self.pad_coef)
y_padded = rotate(
y_padded_rotated,
-self.angle.to(y_padded_rotated),
InterpolationMode.BILINEAR,
fill=0,
)
y_height, y_width = y_padded.shape[2:]
y = y_padded[:, :, pad_h : y_height - pad_h, pad_w : y_width - pad_w]
return y
class SELayer(nn.Module):
def __init__(self, channel, reduction=16):
super(SELayer, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Sequential(
nn.Linear(channel, channel // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(channel // reduction, channel, 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)
res = x * y.expand_as(x)
return res
class FourierUnit(nn.Module):
def __init__(
self,
in_channels,
out_channels,
groups=1,
spatial_scale_factor=None,
spatial_scale_mode="bilinear",
spectral_pos_encoding=False,
use_se=False,
se_kwargs=None,
ffc3d=False,
fft_norm="ortho",
):
# bn_layer not used
super(FourierUnit, self).__init__()
self.groups = groups
self.conv_layer = torch.nn.Conv2d(
in_channels=in_channels * 2 + (2 if spectral_pos_encoding else 0),
out_channels=out_channels * 2,
kernel_size=1,
stride=1,
padding=0,
groups=self.groups,
bias=False,
)
self.bn = torch.nn.BatchNorm2d(out_channels * 2)
self.relu = torch.nn.ReLU(inplace=True)
# squeeze and excitation block
self.use_se = use_se
if use_se:
if se_kwargs is None:
se_kwargs = {}
self.se = SELayer(self.conv_layer.in_channels, **se_kwargs)
self.spatial_scale_factor = spatial_scale_factor
self.spatial_scale_mode = spatial_scale_mode
self.spectral_pos_encoding = spectral_pos_encoding
self.ffc3d = ffc3d
self.fft_norm = fft_norm
def forward(self, x):
half_check = False
if x.type() == "torch.cuda.HalfTensor":
# half only works on gpu anyway
half_check = True
batch = x.shape[0]
if self.spatial_scale_factor is not None:
orig_size = x.shape[-2:]
x = F.interpolate(
x,
scale_factor=self.spatial_scale_factor,
mode=self.spatial_scale_mode,
align_corners=False,
)
# (batch, c, h, w/2+1, 2)
fft_dim = (-3, -2, -1) if self.ffc3d else (-2, -1)
if half_check == True:
ffted = torch.fft.rfftn(
x.float(), dim=fft_dim, norm=self.fft_norm
) # .type(torch.cuda.HalfTensor)
else:
ffted = torch.fft.rfftn(x, dim=fft_dim, norm=self.fft_norm)
ffted = torch.stack((ffted.real, ffted.imag), dim=-1)
ffted = ffted.permute(0, 1, 4, 2, 3).contiguous() # (batch, c, 2, h, w/2+1)
ffted = ffted.view(
(
batch,
-1,
)
+ ffted.size()[3:]
)
if self.spectral_pos_encoding:
height, width = ffted.shape[-2:]
coords_vert = (
torch.linspace(0, 1, height)[None, None, :, None]
.expand(batch, 1, height, width)
.to(ffted)
)
coords_hor = (
torch.linspace(0, 1, width)[None, None, None, :]
.expand(batch, 1, height, width)
.to(ffted)
)
ffted = torch.cat((coords_vert, coords_hor, ffted), dim=1)
if self.use_se:
ffted = self.se(ffted)
if half_check == True:
ffted = self.conv_layer(ffted.half()) # (batch, c*2, h, w/2+1)
else:
ffted = self.conv_layer(
ffted
) # .type(torch.cuda.FloatTensor) # (batch, c*2, h, w/2+1)
ffted = self.relu(self.bn(ffted))
# forcing to be always float
ffted = ffted.float()
ffted = (
ffted.view(
(
batch,
-1,
2,
)
+ ffted.size()[2:]
)
.permute(0, 1, 3, 4, 2)
.contiguous()
) # (batch,c, t, h, w/2+1, 2)
ffted = torch.complex(ffted[..., 0], ffted[..., 1])
ifft_shape_slice = x.shape[-3:] if self.ffc3d else x.shape[-2:]
output = torch.fft.irfftn(
ffted, s=ifft_shape_slice, dim=fft_dim, norm=self.fft_norm
)
if half_check == True:
output = output.half()
if self.spatial_scale_factor is not None:
output = F.interpolate(
output,
size=orig_size,
mode=self.spatial_scale_mode,
align_corners=False,
)
return output
class SpectralTransform(nn.Module):
def __init__(
self,
in_channels,
out_channels,
stride=1,
groups=1,
enable_lfu=True,
separable_fu=False,
**fu_kwargs,
):
# bn_layer not used
super(SpectralTransform, self).__init__()
self.enable_lfu = enable_lfu
if stride == 2:
self.downsample = nn.AvgPool2d(kernel_size=(2, 2), stride=2)
else:
self.downsample = nn.Identity()
self.stride = stride
self.conv1 = nn.Sequential(
nn.Conv2d(
in_channels, out_channels // 2, kernel_size=1, groups=groups, bias=False
),
nn.BatchNorm2d(out_channels // 2),
nn.ReLU(inplace=True),
)
fu_class = FourierUnit
self.fu = fu_class(out_channels // 2, out_channels // 2, groups, **fu_kwargs)
if self.enable_lfu:
self.lfu = fu_class(out_channels // 2, out_channels // 2, groups)
self.conv2 = torch.nn.Conv2d(
out_channels // 2, out_channels, kernel_size=1, groups=groups, bias=False
)
def forward(self, x):
x = self.downsample(x)
x = self.conv1(x)
output = self.fu(x)
if self.enable_lfu:
_, c, h, _ = x.shape
split_no = 2
split_s = h // split_no
xs = torch.cat(
torch.split(x[:, : c // 4], split_s, dim=-2), dim=1
).contiguous()
xs = torch.cat(torch.split(xs, split_s, dim=-1), dim=1).contiguous()
xs = self.lfu(xs)
xs = xs.repeat(1, 1, split_no, split_no).contiguous()
else:
xs = 0
output = self.conv2(x + output + xs)
return output
class FFC(nn.Module):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
ratio_gin,
ratio_gout,
stride=1,
padding=0,
dilation=1,
groups=1,
bias=False,
enable_lfu=True,
padding_type="reflect",
gated=False,
**spectral_kwargs,
):
super(FFC, self).__init__()
assert stride == 1 or stride == 2, "Stride should be 1 or 2."
self.stride = stride
in_cg = int(in_channels * ratio_gin)
in_cl = in_channels - in_cg
out_cg = int(out_channels * ratio_gout)
out_cl = out_channels - out_cg
# groups_g = 1 if groups == 1 else int(groups * ratio_gout)
# groups_l = 1 if groups == 1 else groups - groups_g
self.ratio_gin = ratio_gin
self.ratio_gout = ratio_gout
self.global_in_num = in_cg
module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d
self.convl2l = module(
in_cl,
out_cl,
kernel_size,
stride,
padding,
dilation,
groups,
bias,
padding_mode=padding_type,
)
module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d
self.convl2g = module(
in_cl,
out_cg,
kernel_size,
stride,
padding,
dilation,
groups,
bias,
padding_mode=padding_type,
)
module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d
self.convg2l = module(
in_cg,
out_cl,
kernel_size,
stride,
padding,
dilation,
groups,
bias,
padding_mode=padding_type,
)
module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform
self.convg2g = module(
in_cg,
out_cg,
stride,
1 if groups == 1 else groups // 2,
enable_lfu,
**spectral_kwargs,
)
self.gated = gated
module = (
nn.Identity if in_cg == 0 or out_cl == 0 or not self.gated else nn.Conv2d
)
self.gate = module(in_channels, 2, 1)
def forward(self, x):
x_l, x_g = x if type(x) is tuple else (x, 0)
out_xl, out_xg = 0, 0
if self.gated:
total_input_parts = [x_l]
if torch.is_tensor(x_g):
total_input_parts.append(x_g)
total_input = torch.cat(total_input_parts, dim=1)
gates = torch.sigmoid(self.gate(total_input))
g2l_gate, l2g_gate = gates.chunk(2, dim=1)
else:
g2l_gate, l2g_gate = 1, 1
if self.ratio_gout != 1:
out_xl = self.convl2l(x_l) + self.convg2l(x_g) * g2l_gate
if self.ratio_gout != 0:
out_xg = self.convl2g(x_l) * l2g_gate + self.convg2g(x_g)
return out_xl, out_xg
class FFC_BN_ACT(nn.Module):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
ratio_gin,
ratio_gout,
stride=1,
padding=0,
dilation=1,
groups=1,
bias=False,
norm_layer=nn.BatchNorm2d,
activation_layer=nn.Identity,
padding_type="reflect",
enable_lfu=True,
**kwargs,
):
super(FFC_BN_ACT, self).__init__()
self.ffc = FFC(
in_channels,
out_channels,
kernel_size,
ratio_gin,
ratio_gout,
stride,
padding,
dilation,
groups,
bias,
enable_lfu,
padding_type=padding_type,
**kwargs,
)
lnorm = nn.Identity if ratio_gout == 1 else norm_layer
gnorm = nn.Identity if ratio_gout == 0 else norm_layer
global_channels = int(out_channels * ratio_gout)
self.bn_l = lnorm(out_channels - global_channels)
self.bn_g = gnorm(global_channels)
lact = nn.Identity if ratio_gout == 1 else activation_layer
gact = nn.Identity if ratio_gout == 0 else activation_layer
self.act_l = lact(inplace=True)
self.act_g = gact(inplace=True)
def forward(self, x):
x_l, x_g = self.ffc(x)
x_l = self.act_l(self.bn_l(x_l))
x_g = self.act_g(self.bn_g(x_g))
return x_l, x_g
class FFCResnetBlock(nn.Module):
def __init__(
self,
dim,
padding_type,
norm_layer,
activation_layer=nn.ReLU,
dilation=1,
spatial_transform_kwargs=None,
inline=False,
**conv_kwargs,
):
super().__init__()
self.conv1 = FFC_BN_ACT(
dim,
dim,
kernel_size=3,
padding=dilation,
dilation=dilation,
norm_layer=norm_layer,
activation_layer=activation_layer,
padding_type=padding_type,
**conv_kwargs,
)
self.conv2 = FFC_BN_ACT(
dim,
dim,
kernel_size=3,
padding=dilation,
dilation=dilation,
norm_layer=norm_layer,
activation_layer=activation_layer,
padding_type=padding_type,
**conv_kwargs,
)
if spatial_transform_kwargs is not None:
self.conv1 = LearnableSpatialTransformWrapper(
self.conv1, **spatial_transform_kwargs
)
self.conv2 = LearnableSpatialTransformWrapper(
self.conv2, **spatial_transform_kwargs
)
self.inline = inline
def forward(self, x):
if self.inline:
x_l, x_g = (
x[:, : -self.conv1.ffc.global_in_num],
x[:, -self.conv1.ffc.global_in_num :],
)
else:
x_l, x_g = x if type(x) is tuple else (x, 0)
id_l, id_g = x_l, x_g
x_l, x_g = self.conv1((x_l, x_g))
x_l, x_g = self.conv2((x_l, x_g))
x_l, x_g = id_l + x_l, id_g + x_g
out = x_l, x_g
if self.inline:
out = torch.cat(out, dim=1)
return out
class ConcatTupleLayer(nn.Module):
def forward(self, x):
assert isinstance(x, tuple)
x_l, x_g = x
assert torch.is_tensor(x_l) or torch.is_tensor(x_g)
if not torch.is_tensor(x_g):
return x_l
return torch.cat(x, dim=1)
class FFCResNetGenerator(nn.Module):
def __init__(
self,
input_nc,
output_nc,
ngf=64,
n_downsampling=3,
n_blocks=18,
norm_layer=nn.BatchNorm2d,
padding_type="reflect",
activation_layer=nn.ReLU,
up_norm_layer=nn.BatchNorm2d,
up_activation=nn.ReLU(True),
init_conv_kwargs={},
downsample_conv_kwargs={},
resnet_conv_kwargs={},
spatial_transform_layers=None,
spatial_transform_kwargs={},
max_features=1024,
out_ffc=False,
out_ffc_kwargs={},
):
assert n_blocks >= 0
super().__init__()
"""
init_conv_kwargs = {'ratio_gin': 0, 'ratio_gout': 0, 'enable_lfu': False}
downsample_conv_kwargs = {'ratio_gin': '${generator.init_conv_kwargs.ratio_gout}', 'ratio_gout': '${generator.downsample_conv_kwargs.ratio_gin}', 'enable_lfu': False}
resnet_conv_kwargs = {'ratio_gin': 0.75, 'ratio_gout': '${generator.resnet_conv_kwargs.ratio_gin}', 'enable_lfu': False}
spatial_transform_kwargs = {}
out_ffc_kwargs = {}
"""
"""
print(input_nc, output_nc, ngf, n_downsampling, n_blocks, norm_layer,
padding_type, activation_layer,
up_norm_layer, up_activation,
spatial_transform_layers,
add_out_act, max_features, out_ffc, file=sys.stderr)
4 3 64 3 18 <class 'torch.nn.modules.batchnorm.BatchNorm2d'>
reflect <class 'torch.nn.modules.activation.ReLU'>
<class 'torch.nn.modules.batchnorm.BatchNorm2d'>
ReLU(inplace=True)
None sigmoid 1024 False
"""
init_conv_kwargs = {"ratio_gin": 0, "ratio_gout": 0, "enable_lfu": False}
downsample_conv_kwargs = {"ratio_gin": 0, "ratio_gout": 0, "enable_lfu": False}
resnet_conv_kwargs = {
"ratio_gin": 0.75,
"ratio_gout": 0.75,
"enable_lfu": False,
}
spatial_transform_kwargs = {}
out_ffc_kwargs = {}
model = [
nn.ReflectionPad2d(3),
FFC_BN_ACT(
input_nc,
ngf,
kernel_size=7,
padding=0,
norm_layer=norm_layer,
activation_layer=activation_layer,
**init_conv_kwargs,
),
]
### downsample
for i in range(n_downsampling):
mult = 2**i
if i == n_downsampling - 1:
cur_conv_kwargs = dict(downsample_conv_kwargs)
cur_conv_kwargs["ratio_gout"] = resnet_conv_kwargs.get("ratio_gin", 0)
else:
cur_conv_kwargs = downsample_conv_kwargs
model += [
FFC_BN_ACT(
min(max_features, ngf * mult),
min(max_features, ngf * mult * 2),
kernel_size=3,
stride=2,
padding=1,
norm_layer=norm_layer,
activation_layer=activation_layer,
**cur_conv_kwargs,
)
]
mult = 2**n_downsampling
feats_num_bottleneck = min(max_features, ngf * mult)
### resnet blocks
for i in range(n_blocks):
cur_resblock = FFCResnetBlock(
feats_num_bottleneck,
padding_type=padding_type,
activation_layer=activation_layer,
norm_layer=norm_layer,
**resnet_conv_kwargs,
)
if spatial_transform_layers is not None and i in spatial_transform_layers:
cur_resblock = LearnableSpatialTransformWrapper(
cur_resblock, **spatial_transform_kwargs
)
model += [cur_resblock]
model += [ConcatTupleLayer()]
### upsample
for i in range(n_downsampling):
mult = 2 ** (n_downsampling - i)
model += [
nn.ConvTranspose2d(
min(max_features, ngf * mult),
min(max_features, int(ngf * mult / 2)),
kernel_size=3,
stride=2,
padding=1,
output_padding=1,
),
up_norm_layer(min(max_features, int(ngf * mult / 2))),
up_activation,
]
if out_ffc:
model += [
FFCResnetBlock(
ngf,
padding_type=padding_type,
activation_layer=activation_layer,
norm_layer=norm_layer,
inline=True,
**out_ffc_kwargs,
)
]
model += [
nn.ReflectionPad2d(3),
nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0),
]
model.append(nn.Sigmoid())
self.model = nn.Sequential(*model)
def forward(self, image, mask):
return self.model(torch.cat([image, mask], dim=1))
class LaMa(nn.Module):
def __init__(self, state_dict) -> None:
super(LaMa, self).__init__()
self.model_arch = "LaMa"
self.sub_type = "Inpaint"
self.in_nc = 4
self.out_nc = 3
self.scale = 1
self.min_size = None
self.pad_mod = 8
self.pad_to_square = False
self.model = FFCResNetGenerator(self.in_nc, self.out_nc)
self.state = {
k.replace("generator.model", "model.model"): v
for k, v in state_dict.items()
}
self.supports_fp16 = False
self.support_bf16 = True
self.load_state_dict(self.state, strict=False)
def forward(self, img, mask):
masked_img = img * (1 - mask)
inpainted_mask = mask * self.model.forward(masked_img, mask)
result = inpainted_mask + (1 - mask) * img
return result
comfy_extras/chainner_models/architecture/MAT.py 0 → 100644
View file @ 905857ed
# pylint: skip-file
"""Original MAT project is copyright of fenglingwb: https://github.com/fenglinglwb/MAT
Code used for this implementation of MAT is modified from lama-cleaner,
copyright of Sanster: https://github.com/fenglinglwb/MAT"""
import random
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from .mat.utils import (
Conv2dLayer,
FullyConnectedLayer,
activation_funcs,
bias_act,
conv2d_resample,
normalize_2nd_moment,
setup_filter,
to_2tuple,
upsample2d,
)
class ModulatedConv2d(nn.Module):
def __init__(
self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
kernel_size, # Width and height of the convolution kernel.
style_dim, # dimension of the style code
demodulate=True, # perfrom demodulation
up=1, # Integer upsampling factor.
down=1, # Integer downsampling factor.
resample_filter=[
1,
3,
3,
1,
], # Low-pass filter to apply when resampling activations.
conv_clamp=None, # Clamp the output to +-X, None = disable clamping.
):
super().__init__()
self.demodulate = demodulate
self.weight = torch.nn.Parameter(
torch.randn([1, out_channels, in_channels, kernel_size, kernel_size])
)
self.out_channels = out_channels
self.kernel_size = kernel_size
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size**2))
self.padding = self.kernel_size // 2
self.up = up
self.down = down
self.register_buffer("resample_filter", setup_filter(resample_filter))
self.conv_clamp = conv_clamp
self.affine = FullyConnectedLayer(style_dim, in_channels, bias_init=1)
def forward(self, x, style):
batch, in_channels, height, width = x.shape
style = self.affine(style).view(batch, 1, in_channels, 1, 1).to(x.device)
weight = self.weight.to(x.device) * self.weight_gain * style
if self.demodulate:
decoefs = (weight.pow(2).sum(dim=[2, 3, 4]) + 1e-8).rsqrt()
weight = weight * decoefs.view(batch, self.out_channels, 1, 1, 1)
weight = weight.view(
batch * self.out_channels, in_channels, self.kernel_size, self.kernel_size
)
x = x.view(1, batch * in_channels, height, width)
x = conv2d_resample(
x=x,
w=weight,
f=self.resample_filter,
up=self.up,
down=self.down,
padding=self.padding,
groups=batch,
)
out = x.view(batch, self.out_channels, *x.shape[2:])
return out
class StyleConv(torch.nn.Module):
def __init__(
self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
style_dim, # Intermediate latent (W) dimensionality.
resolution, # Resolution of this layer.
kernel_size=3, # Convolution kernel size.
up=1, # Integer upsampling factor.
use_noise=False, # Enable noise input?
activation="lrelu", # Activation function: 'relu', 'lrelu', etc.
resample_filter=[
1,
3,
3,
1,
], # Low-pass filter to apply when resampling activations.
conv_clamp=None, # Clamp the output of convolution layers to +-X, None = disable clamping.
demodulate=True, # perform demodulation
):
super().__init__()
self.conv = ModulatedConv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
style_dim=style_dim,
demodulate=demodulate,
up=up,
resample_filter=resample_filter,
conv_clamp=conv_clamp,
)
self.use_noise = use_noise
self.resolution = resolution
if use_noise:
self.register_buffer("noise_const", torch.randn([resolution, resolution]))
self.noise_strength = torch.nn.Parameter(torch.zeros([]))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
self.activation = activation
self.act_gain = activation_funcs[activation].def_gain
self.conv_clamp = conv_clamp
def forward(self, x, style, noise_mode="random", gain=1):
x = self.conv(x, style)
assert noise_mode in ["random", "const", "none"]
if self.use_noise:
if noise_mode == "random":
xh, xw = x.size()[-2:]
noise = (
torch.randn([x.shape[0], 1, xh, xw], device=x.device)
* self.noise_strength
)
if noise_mode == "const":
noise = self.noise_const * self.noise_strength
x = x + noise
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
out = bias_act(
x, self.bias, act=self.activation, gain=act_gain, clamp=act_clamp
)
return out
class ToRGB(torch.nn.Module):
def __init__(
self,
in_channels,
out_channels,
style_dim,
kernel_size=1,
resample_filter=[1, 3, 3, 1],
conv_clamp=None,
demodulate=False,
):
super().__init__()
self.conv = ModulatedConv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
style_dim=style_dim,
demodulate=demodulate,
resample_filter=resample_filter,
conv_clamp=conv_clamp,
)
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
self.register_buffer("resample_filter", setup_filter(resample_filter))
self.conv_clamp = conv_clamp
def forward(self, x, style, skip=None):
x = self.conv(x, style)
out = bias_act(x, self.bias, clamp=self.conv_clamp)
if skip is not None:
if skip.shape != out.shape:
skip = upsample2d(skip, self.resample_filter)
out = out + skip
return out
def get_style_code(a, b):
return torch.cat([a, b.to(a.device)], dim=1)
class DecBlockFirst(nn.Module):
def __init__(
self,
in_channels,
out_channels,
activation,
style_dim,
use_noise,
demodulate,
img_channels,
):
super().__init__()
self.fc = FullyConnectedLayer(
in_features=in_channels * 2,
out_features=in_channels * 4**2,
activation=activation,
)
self.conv = StyleConv(
in_channels=in_channels,
out_channels=out_channels,
style_dim=style_dim,
resolution=4,
kernel_size=3,
use_noise=use_noise,
activation=activation,
demodulate=demodulate,
)
self.toRGB = ToRGB(
in_channels=out_channels,
out_channels=img_channels,
style_dim=style_dim,
kernel_size=1,
demodulate=False,
)
def forward(self, x, ws, gs, E_features, noise_mode="random"):
x = self.fc(x).view(x.shape[0], -1, 4, 4)
x = x + E_features[2]
style = get_style_code(ws[:, 0], gs)
x = self.conv(x, style, noise_mode=noise_mode)
style = get_style_code(ws[:, 1], gs)
img = self.toRGB(x, style, skip=None)
return x, img
class MappingNet(torch.nn.Module):
def __init__(
self,
z_dim, # Input latent (Z) dimensionality, 0 = no latent.
c_dim, # Conditioning label (C) dimensionality, 0 = no label.
w_dim, # Intermediate latent (W) dimensionality.
num_ws, # Number of intermediate latents to output, None = do not broadcast.
num_layers=8, # Number of mapping layers.
embed_features=None, # Label embedding dimensionality, None = same as w_dim.
layer_features=None, # Number of intermediate features in the mapping layers, None = same as w_dim.
activation="lrelu", # Activation function: 'relu', 'lrelu', etc.
lr_multiplier=0.01, # Learning rate multiplier for the mapping layers.
w_avg_beta=0.995, # Decay for tracking the moving average of W during training, None = do not track.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.num_ws = num_ws
self.num_layers = num_layers
self.w_avg_beta = w_avg_beta
if embed_features is None:
embed_features = w_dim
if c_dim == 0:
embed_features = 0
if layer_features is None:
layer_features = w_dim
features_list = (
[z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]
)
if c_dim > 0:
self.embed = FullyConnectedLayer(c_dim, embed_features)
for idx in range(num_layers):
in_features = features_list[idx]
out_features = features_list[idx + 1]
layer = FullyConnectedLayer(
in_features,
out_features,
activation=activation,
lr_multiplier=lr_multiplier,
)
setattr(self, f"fc{idx}", layer)
if num_ws is not None and w_avg_beta is not None:
self.register_buffer("w_avg", torch.zeros([w_dim]))
def forward(
self, z, c, truncation_psi=1, truncation_cutoff=None, skip_w_avg_update=False
):
# Embed, normalize, and concat inputs.
x = None
with torch.autograd.profiler.record_function("input"):
if self.z_dim > 0:
x = normalize_2nd_moment(z.to(torch.float32))
if self.c_dim > 0:
y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
x = torch.cat([x, y], dim=1) if x is not None else y
# Main layers.
for idx in range(self.num_layers):
layer = getattr(self, f"fc{idx}")
x = layer(x)
# Update moving average of W.
if self.w_avg_beta is not None and self.training and not skip_w_avg_update:
with torch.autograd.profiler.record_function("update_w_avg"):
self.w_avg.copy_(
x.detach().mean(dim=0).lerp(self.w_avg, self.w_avg_beta)
)
# Broadcast.
if self.num_ws is not None:
with torch.autograd.profiler.record_function("broadcast"):
x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
# Apply truncation.
if truncation_psi != 1:
with torch.autograd.profiler.record_function("truncate"):
assert self.w_avg_beta is not None
if self.num_ws is None or truncation_cutoff is None:
x = self.w_avg.lerp(x, truncation_psi)
else:
x[:, :truncation_cutoff] = self.w_avg.lerp(
x[:, :truncation_cutoff], truncation_psi
)
return x
class DisFromRGB(nn.Module):
def __init__(
self, in_channels, out_channels, activation
): # res = 2, ..., resolution_log2
super().__init__()
self.conv = Conv2dLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
activation=activation,
)
def forward(self, x):
return self.conv(x)
class DisBlock(nn.Module):
def __init__(
self, in_channels, out_channels, activation
): # res = 2, ..., resolution_log2
super().__init__()
self.conv0 = Conv2dLayer(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=3,
activation=activation,
)
self.conv1 = Conv2dLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
down=2,
activation=activation,
)
self.skip = Conv2dLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
down=2,
bias=False,
)
def forward(self, x):
skip = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x)
x = self.conv1(x, gain=np.sqrt(0.5))
out = skip + x
return out
def nf(stage, channel_base=32768, channel_decay=1.0, channel_max=512):
NF = {512: 64, 256: 128, 128: 256, 64: 512, 32: 512, 16: 512, 8: 512, 4: 512}
return NF[2**stage]
class Mlp(nn.Module):
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 = FullyConnectedLayer(
in_features=in_features, out_features=hidden_features, activation="lrelu"
)
self.fc2 = FullyConnectedLayer(
in_features=hidden_features, out_features=out_features
)
def forward(self, x):
x = self.fc1(x)
x = self.fc2(x)
return x
def window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
windows = (
x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
)
return windows
def window_reverse(windows, window_size: int, H: int, W: int):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
B = int(windows.shape[0] / (H * W / window_size / window_size))
# B = windows.shape[0] / (H * W / window_size / window_size)
x = windows.view(
B, H // window_size, W // window_size, window_size, window_size, -1
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
class Conv2dLayerPartial(nn.Module):
def __init__(
self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
kernel_size, # Width and height of the convolution kernel.
bias=True, # Apply additive bias before the activation function?
activation="linear", # Activation function: 'relu', 'lrelu', etc.
up=1, # Integer upsampling factor.
down=1, # Integer downsampling factor.
resample_filter=[
1,
3,
3,
1,
], # Low-pass filter to apply when resampling activations.
conv_clamp=None, # Clamp the output to +-X, None = disable clamping.
trainable=True, # Update the weights of this layer during training?
):
super().__init__()
self.conv = Conv2dLayer(
in_channels,
out_channels,
kernel_size,
bias,
activation,
up,
down,
resample_filter,
conv_clamp,
trainable,
)
self.weight_maskUpdater = torch.ones(1, 1, kernel_size, kernel_size)
self.slide_winsize = kernel_size**2
self.stride = down
self.padding = kernel_size // 2 if kernel_size % 2 == 1 else 0
def forward(self, x, mask=None):
if mask is not None:
with torch.no_grad():
if self.weight_maskUpdater.type() != x.type():
self.weight_maskUpdater = self.weight_maskUpdater.to(x)
update_mask = F.conv2d(
mask,
self.weight_maskUpdater,
bias=None,
stride=self.stride,
padding=self.padding,
)
mask_ratio = self.slide_winsize / (update_mask + 1e-8)
update_mask = torch.clamp(update_mask, 0, 1) # 0 or 1
mask_ratio = torch.mul(mask_ratio, update_mask)
x = self.conv(x)
x = torch.mul(x, mask_ratio)
return x, update_mask
else:
x = self.conv(x)
return x, None
class WindowAttention(nn.Module):
r"""Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
"""
def __init__(
self,
dim,
window_size,
num_heads,
down_ratio=1,
qkv_bias=True,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim**-0.5
self.q = FullyConnectedLayer(in_features=dim, out_features=dim)
self.k = FullyConnectedLayer(in_features=dim, out_features=dim)
self.v = FullyConnectedLayer(in_features=dim, out_features=dim)
self.proj = FullyConnectedLayer(in_features=dim, out_features=dim)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask_windows=None, mask=None):
"""
Args:
x: input features with shape of (num_windows*B, N, C)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
B_, N, C = x.shape
norm_x = F.normalize(x, p=2.0, dim=-1)
q = (
self.q(norm_x)
.reshape(B_, N, self.num_heads, C // self.num_heads)
.permute(0, 2, 1, 3)
)
k = (
self.k(norm_x)
.view(B_, -1, self.num_heads, C // self.num_heads)
.permute(0, 2, 3, 1)
)
v = (
self.v(x)
.view(B_, -1, self.num_heads, C // self.num_heads)
.permute(0, 2, 1, 3)
)
attn = (q @ k) * self.scale
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
1
).unsqueeze(0)
attn = attn.view(-1, self.num_heads, N, N)
if mask_windows is not None:
attn_mask_windows = mask_windows.squeeze(-1).unsqueeze(1).unsqueeze(1)
attn = attn + attn_mask_windows.masked_fill(
attn_mask_windows == 0, float(-100.0)
).masked_fill(attn_mask_windows == 1, float(0.0))
with torch.no_grad():
mask_windows = torch.clamp(
torch.sum(mask_windows, dim=1, keepdim=True), 0, 1
).repeat(1, N, 1)
attn = self.softmax(attn)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
x = self.proj(x)
return x, mask_windows
class SwinTransformerBlock(nn.Module):
r"""Swin Transformer Block.
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.
shift_size (int): Shift size for SW-MSA.
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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
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
"""
def __init__(
self,
dim,
input_resolution,
num_heads,
down_ratio=1,
window_size=7,
shift_size=0,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
):
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"
if self.shift_size > 0:
down_ratio = 1
self.attn = WindowAttention(
dim,
window_size=to_2tuple(self.window_size),
num_heads=num_heads,
down_ratio=down_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
)
self.fuse = FullyConnectedLayer(
in_features=dim * 2, out_features=dim, activation="lrelu"
)
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:
attn_mask = self.calculate_mask(self.input_resolution)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
H, W = x_size
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)
)
return attn_mask
def forward(self, x, x_size, mask=None):
# H, W = self.input_resolution
H, W = x_size
B, _, C = x.shape
# assert L == H * W, "input feature has wrong size"
shortcut = x
x = x.view(B, H, W, C)
if mask is not None:
mask = mask.view(B, H, W, 1)
# cyclic shift
if self.shift_size > 0:
shifted_x = torch.roll(
x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
)
if mask is not None:
shifted_mask = torch.roll(
mask, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
)
else:
shifted_x = x
if mask is not None:
shifted_mask = mask
# 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
if mask is not None:
mask_windows = window_partition(shifted_mask, self.window_size)
mask_windows = mask_windows.view(-1, self.window_size * self.window_size, 1)
else:
mask_windows = None
# W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows, mask_windows = self.attn(
x_windows, mask_windows, mask=self.attn_mask
) # nW*B, window_size*window_size, C
else:
attn_windows, mask_windows = self.attn(
x_windows, mask_windows, mask=self.calculate_mask(x_size).to(x.device)
) # nW*B, window_size*window_size, C
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
if mask is not None:
mask_windows = mask_windows.view(-1, self.window_size, self.window_size, 1)
shifted_mask = window_reverse(mask_windows, self.window_size, H, W)
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(
shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
)
if mask is not None:
mask = torch.roll(
shifted_mask, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
)
else:
x = shifted_x
if mask is not None:
mask = shifted_mask
x = x.view(B, H * W, C)
if mask is not None:
mask = mask.view(B, H * W, 1)
# FFN
x = self.fuse(torch.cat([shortcut, x], dim=-1))
x = self.mlp(x)
return x, mask
class PatchMerging(nn.Module):
def __init__(self, in_channels, out_channels, down=2):
super().__init__()
self.conv = Conv2dLayerPartial(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
activation="lrelu",
down=down,
)
self.down = down
def forward(self, x, x_size, mask=None):
x = token2feature(x, x_size)
if mask is not None:
mask = token2feature(mask, x_size)
x, mask = self.conv(x, mask)
if self.down != 1:
ratio = 1 / self.down
x_size = (int(x_size[0] * ratio), int(x_size[1] * ratio))
x = feature2token(x)
if mask is not None:
mask = feature2token(mask)
return x, x_size, mask
class PatchUpsampling(nn.Module):
def __init__(self, in_channels, out_channels, up=2):
super().__init__()
self.conv = Conv2dLayerPartial(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
activation="lrelu",
up=up,
)
self.up = up
def forward(self, x, x_size, mask=None):
x = token2feature(x, x_size)
if mask is not None:
mask = token2feature(mask, x_size)
x, mask = self.conv(x, mask)
if self.up != 1:
x_size = (int(x_size[0] * self.up), int(x_size[1] * self.up))
x = feature2token(x)
if mask is not None:
mask = feature2token(mask)
return x, x_size, mask
class BasicLayer(nn.Module):
"""A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
down_ratio=1,
mlp_ratio=2.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# patch merging layer
if downsample is not None:
# self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
self.downsample = downsample
else:
self.downsample = None
# build blocks
self.blocks = nn.ModuleList(
[
SwinTransformerBlock(
dim=dim,
input_resolution=input_resolution,
num_heads=num_heads,
down_ratio=down_ratio,
window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i]
if isinstance(drop_path, list)
else drop_path,
norm_layer=norm_layer,
)
for i in range(depth)
]
)
self.conv = Conv2dLayerPartial(
in_channels=dim, out_channels=dim, kernel_size=3, activation="lrelu"
)
def forward(self, x, x_size, mask=None):
if self.downsample is not None:
x, x_size, mask = self.downsample(x, x_size, mask)
identity = x
for blk in self.blocks:
if self.use_checkpoint:
x, mask = checkpoint.checkpoint(blk, x, x_size, mask)
else:
x, mask = blk(x, x_size, mask)
if mask is not None:
mask = token2feature(mask, x_size)
x, mask = self.conv(token2feature(x, x_size), mask)
x = feature2token(x) + identity
if mask is not None:
mask = feature2token(mask)
return x, x_size, mask
class ToToken(nn.Module):
def __init__(self, in_channels=3, dim=128, kernel_size=5, stride=1):
super().__init__()
self.proj = Conv2dLayerPartial(
in_channels=in_channels,
out_channels=dim,
kernel_size=kernel_size,
activation="lrelu",
)
def forward(self, x, mask):
x, mask = self.proj(x, mask)
return x, mask
class EncFromRGB(nn.Module):
def __init__(
self, in_channels, out_channels, activation
): # res = 2, ..., resolution_log2
super().__init__()
self.conv0 = Conv2dLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
activation=activation,
)
self.conv1 = Conv2dLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
activation=activation,
)
def forward(self, x):
x = self.conv0(x)
x = self.conv1(x)
return x
class ConvBlockDown(nn.Module):
def __init__(
self, in_channels, out_channels, activation
): # res = 2, ..., resolution_log
super().__init__()
self.conv0 = Conv2dLayer(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
activation=activation,
down=2,
)
self.conv1 = Conv2dLayer(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=3,
activation=activation,
)
def forward(self, x):
x = self.conv0(x)
x = self.conv1(x)
return x
def token2feature(x, x_size):
B, _, C = x.shape
h, w = x_size
x = x.permute(0, 2, 1).reshape(B, C, h, w)
return x
def feature2token(x):
B, C, _, _ = x.shape
x = x.view(B, C, -1).transpose(1, 2)
return x
class Encoder(nn.Module):
def __init__(
self,
res_log2,
img_channels,
activation,
patch_size=5,
channels=16,
drop_path_rate=0.1,
):
super().__init__()
self.resolution = []
for i in range(res_log2, 3, -1): # from input size to 16x16
res = 2**i
self.resolution.append(res)
if i == res_log2:
block = EncFromRGB(img_channels * 2 + 1, nf(i), activation)
else:
block = ConvBlockDown(nf(i + 1), nf(i), activation)
setattr(self, "EncConv_Block_%dx%d" % (res, res), block)
def forward(self, x):
out = {}
for res in self.resolution:
res_log2 = int(np.log2(res))
x = getattr(self, "EncConv_Block_%dx%d" % (res, res))(x)
out[res_log2] = x
return out
class ToStyle(nn.Module):
def __init__(self, in_channels, out_channels, activation, drop_rate):
super().__init__()
self.conv = nn.Sequential(
Conv2dLayer(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=3,
activation=activation,
down=2,
),
Conv2dLayer(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=3,
activation=activation,
down=2,
),
Conv2dLayer(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=3,
activation=activation,
down=2,
),
)
self.pool = nn.AdaptiveAvgPool2d(1)
self.fc = FullyConnectedLayer(
in_features=in_channels, out_features=out_channels, activation=activation
)
# self.dropout = nn.Dropout(drop_rate)
def forward(self, x):
x = self.conv(x)
x = self.pool(x)
x = self.fc(x.flatten(start_dim=1))
# x = self.dropout(x)
return x
class DecBlockFirstV2(nn.Module):
def __init__(
self,
res,
in_channels,
out_channels,
activation,
style_dim,
use_noise,
demodulate,
img_channels,
):
super().__init__()
self.res = res
self.conv0 = Conv2dLayer(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=3,
activation=activation,
)
self.conv1 = StyleConv(
in_channels=in_channels,
out_channels=out_channels,
style_dim=style_dim,
resolution=2**res,
kernel_size=3,
use_noise=use_noise,
activation=activation,
demodulate=demodulate,
)
self.toRGB = ToRGB(
in_channels=out_channels,
out_channels=img_channels,
style_dim=style_dim,
kernel_size=1,
demodulate=False,
)
def forward(self, x, ws, gs, E_features, noise_mode="random"):
# x = self.fc(x).view(x.shape[0], -1, 4, 4)
x = self.conv0(x)
x = x + E_features[self.res]
style = get_style_code(ws[:, 0], gs)
x = self.conv1(x, style, noise_mode=noise_mode)
style = get_style_code(ws[:, 1], gs)
img = self.toRGB(x, style, skip=None)
return x, img
class DecBlock(nn.Module):
def __init__(
self,
res,
in_channels,
out_channels,
activation,
style_dim,
use_noise,
demodulate,
img_channels,
): # res = 4, ..., resolution_log2
super().__init__()
self.res = res
self.conv0 = StyleConv(
in_channels=in_channels,
out_channels=out_channels,
style_dim=style_dim,
resolution=2**res,
kernel_size=3,
up=2,
use_noise=use_noise,
activation=activation,
demodulate=demodulate,
)
self.conv1 = StyleConv(
in_channels=out_channels,
out_channels=out_channels,
style_dim=style_dim,
resolution=2**res,
kernel_size=3,
use_noise=use_noise,
activation=activation,
demodulate=demodulate,
)
self.toRGB = ToRGB(
in_channels=out_channels,
out_channels=img_channels,
style_dim=style_dim,
kernel_size=1,
demodulate=False,
)
def forward(self, x, img, ws, gs, E_features, noise_mode="random"):
style = get_style_code(ws[:, self.res * 2 - 9], gs)
x = self.conv0(x, style, noise_mode=noise_mode)
x = x + E_features[self.res]
style = get_style_code(ws[:, self.res * 2 - 8], gs)
x = self.conv1(x, style, noise_mode=noise_mode)
style = get_style_code(ws[:, self.res * 2 - 7], gs)
img = self.toRGB(x, style, skip=img)
return x, img
class Decoder(nn.Module):
def __init__(
self, res_log2, activation, style_dim, use_noise, demodulate, img_channels
):
super().__init__()
self.Dec_16x16 = DecBlockFirstV2(
4, nf(4), nf(4), activation, style_dim, use_noise, demodulate, img_channels
)
for res in range(5, res_log2 + 1):
setattr(
self,
"Dec_%dx%d" % (2**res, 2**res),
DecBlock(
res,
nf(res - 1),
nf(res),
activation,
style_dim,
use_noise,
demodulate,
img_channels,
),
)
self.res_log2 = res_log2
def forward(self, x, ws, gs, E_features, noise_mode="random"):
x, img = self.Dec_16x16(x, ws, gs, E_features, noise_mode=noise_mode)
for res in range(5, self.res_log2 + 1):
block = getattr(self, "Dec_%dx%d" % (2**res, 2**res))
x, img = block(x, img, ws, gs, E_features, noise_mode=noise_mode)
return img
class DecStyleBlock(nn.Module):
def __init__(
self,
res,
in_channels,
out_channels,
activation,
style_dim,
use_noise,
demodulate,
img_channels,
):
super().__init__()
self.res = res
self.conv0 = StyleConv(
in_channels=in_channels,
out_channels=out_channels,
style_dim=style_dim,
resolution=2**res,
kernel_size=3,
up=2,
use_noise=use_noise,
activation=activation,
demodulate=demodulate,
)
self.conv1 = StyleConv(
in_channels=out_channels,
out_channels=out_channels,
style_dim=style_dim,
resolution=2**res,
kernel_size=3,
use_noise=use_noise,
activation=activation,
demodulate=demodulate,
)
self.toRGB = ToRGB(
in_channels=out_channels,
out_channels=img_channels,
style_dim=style_dim,
kernel_size=1,
demodulate=False,
)
def forward(self, x, img, style, skip, noise_mode="random"):
x = self.conv0(x, style, noise_mode=noise_mode)
x = x + skip
x = self.conv1(x, style, noise_mode=noise_mode)
img = self.toRGB(x, style, skip=img)
return x, img
class FirstStage(nn.Module):
def __init__(
self,
img_channels,
img_resolution=256,
dim=180,
w_dim=512,
use_noise=False,
demodulate=True,
activation="lrelu",
):
super().__init__()
res = 64
self.conv_first = Conv2dLayerPartial(
in_channels=img_channels + 1,
out_channels=dim,
kernel_size=3,
activation=activation,
)
self.enc_conv = nn.ModuleList()
down_time = int(np.log2(img_resolution // res))
# 根据图片尺寸构建 swim transformer 的层数
for i in range(down_time): # from input size to 64
self.enc_conv.append(
Conv2dLayerPartial(
in_channels=dim,
out_channels=dim,
kernel_size=3,
down=2,
activation=activation,
)
)
# from 64 -> 16 -> 64
depths = [2, 3, 4, 3, 2]
ratios = [1, 1 / 2, 1 / 2, 2, 2]
num_heads = 6
window_sizes = [8, 16, 16, 16, 8]
drop_path_rate = 0.1
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
self.tran = nn.ModuleList()
for i, depth in enumerate(depths):
res = int(res * ratios[i])
if ratios[i] < 1:
merge = PatchMerging(dim, dim, down=int(1 / ratios[i]))
elif ratios[i] > 1:
merge = PatchUpsampling(dim, dim, up=ratios[i])
else:
merge = None
self.tran.append(
BasicLayer(
dim=dim,
input_resolution=[res, res],
depth=depth,
num_heads=num_heads,
window_size=window_sizes[i],
drop_path=dpr[sum(depths[:i]) : sum(depths[: i + 1])],
downsample=merge,
)
)
# global style
down_conv = []
for i in range(int(np.log2(16))):
down_conv.append(
Conv2dLayer(
in_channels=dim,
out_channels=dim,
kernel_size=3,
down=2,
activation=activation,
)
)
down_conv.append(nn.AdaptiveAvgPool2d((1, 1)))
self.down_conv = nn.Sequential(*down_conv)
self.to_style = FullyConnectedLayer(
in_features=dim, out_features=dim * 2, activation=activation
)
self.ws_style = FullyConnectedLayer(
in_features=w_dim, out_features=dim, activation=activation
)
self.to_square = FullyConnectedLayer(
in_features=dim, out_features=16 * 16, activation=activation
)
style_dim = dim * 3
self.dec_conv = nn.ModuleList()
for i in range(down_time): # from 64 to input size
res = res * 2
self.dec_conv.append(
DecStyleBlock(
res,
dim,
dim,
activation,
style_dim,
use_noise,
demodulate,
img_channels,
)
)
def forward(self, images_in, masks_in, ws, noise_mode="random"):
x = torch.cat([masks_in - 0.5, images_in * masks_in], dim=1)
skips = []
x, mask = self.conv_first(x, masks_in) # input size
skips.append(x)
for i, block in enumerate(self.enc_conv): # input size to 64
x, mask = block(x, mask)
if i != len(self.enc_conv) - 1:
skips.append(x)
x_size = x.size()[-2:]
x = feature2token(x)
mask = feature2token(mask)
mid = len(self.tran) // 2
for i, block in enumerate(self.tran): # 64 to 16
if i < mid:
x, x_size, mask = block(x, x_size, mask)
skips.append(x)
elif i > mid:
x, x_size, mask = block(x, x_size, None)
x = x + skips[mid - i]
else:
x, x_size, mask = block(x, x_size, None)
mul_map = torch.ones_like(x) * 0.5
mul_map = F.dropout(mul_map, training=True).to(x.device)
ws = self.ws_style(ws[:, -1]).to(x.device)
add_n = self.to_square(ws).unsqueeze(1).to(x.device)
add_n = (
F.interpolate(
add_n, size=x.size(1), mode="linear", align_corners=False
)
.squeeze(1)
.unsqueeze(-1)
).to(x.device)
x = x * mul_map + add_n * (1 - mul_map)
gs = self.to_style(
self.down_conv(token2feature(x, x_size)).flatten(start_dim=1)
).to(x.device)
style = torch.cat([gs, ws], dim=1)
x = token2feature(x, x_size).contiguous()
img = None
for i, block in enumerate(self.dec_conv):
x, img = block(
x, img, style, skips[len(self.dec_conv) - i - 1], noise_mode=noise_mode
)
# ensemble
img = img * (1 - masks_in) + images_in * masks_in
return img
class SynthesisNet(nn.Module):
def __init__(
self,
w_dim, # Intermediate latent (W) dimensionality.
img_resolution, # Output image resolution.
img_channels=3, # Number of color channels.
channel_base=32768, # Overall multiplier for the number of channels.
channel_decay=1.0,
channel_max=512, # Maximum number of channels in any layer.
activation="lrelu", # Activation function: 'relu', 'lrelu', etc.
drop_rate=0.5,
use_noise=False,
demodulate=True,
):
super().__init__()
resolution_log2 = int(np.log2(img_resolution))
assert img_resolution == 2**resolution_log2 and img_resolution >= 4
self.num_layers = resolution_log2 * 2 - 3 * 2
self.img_resolution = img_resolution
self.resolution_log2 = resolution_log2
# first stage
self.first_stage = FirstStage(
img_channels,
img_resolution=img_resolution,
w_dim=w_dim,
use_noise=False,
demodulate=demodulate,
)
# second stage
self.enc = Encoder(
resolution_log2, img_channels, activation, patch_size=5, channels=16
)
self.to_square = FullyConnectedLayer(
in_features=w_dim, out_features=16 * 16, activation=activation
)
self.to_style = ToStyle(
in_channels=nf(4),
out_channels=nf(2) * 2,
activation=activation,
drop_rate=drop_rate,
)
style_dim = w_dim + nf(2) * 2
self.dec = Decoder(
resolution_log2, activation, style_dim, use_noise, demodulate, img_channels
)
def forward(self, images_in, masks_in, ws, noise_mode="random", return_stg1=False):
out_stg1 = self.first_stage(images_in, masks_in, ws, noise_mode=noise_mode)
# encoder
x = images_in * masks_in + out_stg1 * (1 - masks_in)
x = torch.cat([masks_in - 0.5, x, images_in * masks_in], dim=1)
E_features = self.enc(x)
fea_16 = E_features[4].to(x.device)
mul_map = torch.ones_like(fea_16) * 0.5
mul_map = F.dropout(mul_map, training=True).to(x.device)
add_n = self.to_square(ws[:, 0]).view(-1, 16, 16).unsqueeze(1)
add_n = F.interpolate(
add_n, size=fea_16.size()[-2:], mode="bilinear", align_corners=False
).to(x.device)
fea_16 = fea_16 * mul_map + add_n * (1 - mul_map)
E_features[4] = fea_16
# style
gs = self.to_style(fea_16).to(x.device)
# decoder
img = self.dec(fea_16, ws, gs, E_features, noise_mode=noise_mode).to(x.device)
# ensemble
img = img * (1 - masks_in) + images_in * masks_in
if not return_stg1:
return img
else:
return img, out_stg1
class Generator(nn.Module):
def __init__(
self,
z_dim, # Input latent (Z) dimensionality, 0 = no latent.
c_dim, # Conditioning label (C) dimensionality, 0 = no label.
w_dim, # Intermediate latent (W) dimensionality.
img_resolution, # resolution of generated image
img_channels, # Number of input color channels.
synthesis_kwargs={}, # Arguments for SynthesisNetwork.
mapping_kwargs={}, # Arguments for MappingNetwork.
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.img_resolution = img_resolution
self.img_channels = img_channels
self.synthesis = SynthesisNet(
w_dim=w_dim,
img_resolution=img_resolution,
img_channels=img_channels,
**synthesis_kwargs,
)
self.mapping = MappingNet(
z_dim=z_dim,
c_dim=c_dim,
w_dim=w_dim,
num_ws=self.synthesis.num_layers,
**mapping_kwargs,
)
def forward(
self,
images_in,
masks_in,
z,
c,
truncation_psi=1,
truncation_cutoff=None,
skip_w_avg_update=False,
noise_mode="none",
return_stg1=False,
):
ws = self.mapping(
z,
c,
truncation_psi=truncation_psi,
truncation_cutoff=truncation_cutoff,
skip_w_avg_update=skip_w_avg_update,
)
img = self.synthesis(images_in, masks_in, ws, noise_mode=noise_mode)
return img
class MAT(nn.Module):
def __init__(self, state_dict):
super(MAT, self).__init__()
self.model_arch = "MAT"
self.sub_type = "Inpaint"
self.in_nc = 3
self.out_nc = 3
self.scale = 1
self.supports_fp16 = False
self.supports_bf16 = True
self.min_size = 512
self.pad_mod = 512
self.pad_to_square = True
seed = 240 # pick up a random number
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
self.model = Generator(
z_dim=512, c_dim=0, w_dim=512, img_resolution=512, img_channels=3
)
self.z = torch.from_numpy(np.random.randn(1, self.model.z_dim)) # [1., 512]
self.label = torch.zeros([1, self.model.c_dim])
self.state = {
k.replace("synthesis", "model.synthesis").replace(
"mapping", "model.mapping"
): v
for k, v in state_dict.items()
}
self.load_state_dict(self.state, strict=False)
def forward(self, image, mask):
"""Input images and output images have same size
images: [H, W, C] RGB
masks: [H, W] mask area == 255
return: BGR IMAGE
"""
image = image * 2 - 1 # [0, 1] -> [-1, 1]
mask = 1 - mask
output = self.model(
image, mask, self.z, self.label, truncation_psi=1, noise_mode="none"
)
return output * 0.5 + 0.5
comfy_extras/chainner_models/architecture/RRDB.py 0 → 100644
View file @ 905857ed
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import functools
import math
import re
from collections import OrderedDict
import torch
import torch.nn as nn
import torch.nn.functional as F
from . import block as B
# Borrowed from https://github.com/rlaphoenix/VSGAN/blob/master/vsgan/archs/ESRGAN.py
# Which enhanced stuff that was already here
class RRDBNet(nn.Module):
def __init__(
self,
state_dict,
norm=None,
act: str = "leakyrelu",
upsampler: str = "upconv",
mode: B.ConvMode = "CNA",
) -> None:
"""
ESRGAN - Enhanced Super-Resolution Generative Adversarial Networks.
By Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao,
and Chen Change Loy.
This is old-arch Residual in Residual Dense Block Network and is not
the newest revision that's available at github.com/xinntao/ESRGAN.
This is on purpose, the newest Network has severely limited the
potential use of the Network with no benefits.
This network supports model files from both new and old-arch.
Args:
norm: Normalization layer
act: Activation layer
upsampler: Upsample layer. upconv, pixel_shuffle
mode: Convolution mode
"""
super(RRDBNet, self).__init__()
self.model_arch = "ESRGAN"
self.sub_type = "SR"
self.state = state_dict
self.norm = norm
self.act = act
self.upsampler = upsampler
self.mode = mode
self.state_map = {
# currently supports old, new, and newer RRDBNet arch models
# ESRGAN, BSRGAN/RealSR, Real-ESRGAN
"model.0.weight": ("conv_first.weight",),
"model.0.bias": ("conv_first.bias",),
"model.1.sub./NB/.weight": ("trunk_conv.weight", "conv_body.weight"),
"model.1.sub./NB/.bias": ("trunk_conv.bias", "conv_body.bias"),
r"model.1.sub.\1.RDB\2.conv\3.0.\4": (
r"RRDB_trunk\.(\d+)\.RDB(\d)\.conv(\d+)\.(weight|bias)",
r"body\.(\d+)\.rdb(\d)\.conv(\d+)\.(weight|bias)",
),
}
if "params_ema" in self.state:
self.state = self.state["params_ema"]
# self.model_arch = "RealESRGAN"
self.num_blocks = self.get_num_blocks()
self.plus = any("conv1x1" in k for k in self.state.keys())
if self.plus:
self.model_arch = "ESRGAN+"
self.state = self.new_to_old_arch(self.state)
self.key_arr = list(self.state.keys())
self.in_nc: int = self.state[self.key_arr[0]].shape[1]
self.out_nc: int = self.state[self.key_arr[-1]].shape[0]
self.scale: int = self.get_scale()
self.num_filters: int = self.state[self.key_arr[0]].shape[0]
self.supports_fp16 = True
self.supports_bfp16 = True
self.min_size_restriction = None
# Detect if pixelunshuffle was used (Real-ESRGAN)
if self.in_nc in (self.out_nc * 4, self.out_nc * 16) and self.out_nc in (
self.in_nc / 4,
self.in_nc / 16,
):
self.shuffle_factor = int(math.sqrt(self.in_nc / self.out_nc))
else:
self.shuffle_factor = None
upsample_block = {
"upconv": B.upconv_block,
"pixel_shuffle": B.pixelshuffle_block,
}.get(self.upsampler)
if upsample_block is None:
raise NotImplementedError(f"Upsample mode [{self.upsampler}] is not found")
if self.scale == 3:
upsample_blocks = upsample_block(
in_nc=self.num_filters,
out_nc=self.num_filters,
upscale_factor=3,
act_type=self.act,
)
else:
upsample_blocks = [
upsample_block(
in_nc=self.num_filters, out_nc=self.num_filters, act_type=self.act
)
for _ in range(int(math.log(self.scale, 2)))
]
self.model = B.sequential(
# fea conv
B.conv_block(
in_nc=self.in_nc,
out_nc=self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
),
B.ShortcutBlock(
B.sequential(
# rrdb blocks
*[
B.RRDB(
nf=self.num_filters,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=self.norm,
act_type=self.act,
mode="CNA",
plus=self.plus,
)
for _ in range(self.num_blocks)
],
# lr conv
B.conv_block(
in_nc=self.num_filters,
out_nc=self.num_filters,
kernel_size=3,
norm_type=self.norm,
act_type=None,
mode=self.mode,
),
)
),
*upsample_blocks,
# hr_conv0
B.conv_block(
in_nc=self.num_filters,
out_nc=self.num_filters,
kernel_size=3,
norm_type=None,
act_type=self.act,
),
# hr_conv1
B.conv_block(
in_nc=self.num_filters,
out_nc=self.out_nc,
kernel_size=3,
norm_type=None,
act_type=None,
),
)
# Adjust these properties for calculations outside of the model
if self.shuffle_factor:
self.in_nc //= self.shuffle_factor**2
self.scale //= self.shuffle_factor
self.load_state_dict(self.state, strict=False)
def new_to_old_arch(self, state):
"""Convert a new-arch model state dictionary to an old-arch dictionary."""
if "params_ema" in state:
state = state["params_ema"]
if "conv_first.weight" not in state:
# model is already old arch, this is a loose check, but should be sufficient
return state
# add nb to state keys
for kind in ("weight", "bias"):
self.state_map[f"model.1.sub.{self.num_blocks}.{kind}"] = self.state_map[
f"model.1.sub./NB/.{kind}"
]
del self.state_map[f"model.1.sub./NB/.{kind}"]
old_state = OrderedDict()
for old_key, new_keys in self.state_map.items():
for new_key in new_keys:
if r"\1" in old_key:
for k, v in state.items():
sub = re.sub(new_key, old_key, k)
if sub != k:
old_state[sub] = v
else:
if new_key in state:
old_state[old_key] = state[new_key]
# upconv layers
max_upconv = 0
for key in state.keys():
match = re.match(r"(upconv|conv_up)(\d)\.(weight|bias)", key)
if match is not None:
_, key_num, key_type = match.groups()
old_state[f"model.{int(key_num) * 3}.{key_type}"] = state[key]
max_upconv = max(max_upconv, int(key_num) * 3)
# final layers
for key in state.keys():
if key in ("HRconv.weight", "conv_hr.weight"):
old_state[f"model.{max_upconv + 2}.weight"] = state[key]
elif key in ("HRconv.bias", "conv_hr.bias"):
old_state[f"model.{max_upconv + 2}.bias"] = state[key]
elif key in ("conv_last.weight",):
old_state[f"model.{max_upconv + 4}.weight"] = state[key]
elif key in ("conv_last.bias",):
old_state[f"model.{max_upconv + 4}.bias"] = state[key]
# Sort by first numeric value of each layer
def compare(item1, item2):
parts1 = item1.split(".")
parts2 = item2.split(".")
int1 = int(parts1[1])
int2 = int(parts2[1])
return int1 - int2
sorted_keys = sorted(old_state.keys(), key=functools.cmp_to_key(compare))
# Rebuild the output dict in the right order
out_dict = OrderedDict((k, old_state[k]) for k in sorted_keys)
return out_dict
def get_scale(self, min_part: int = 6) -> int:
n = 0
for part in list(self.state):
parts = part.split(".")[1:]
if len(parts) == 2:
part_num = int(parts[0])
if part_num > min_part and parts[1] == "weight":
n += 1
return 2**n
def get_num_blocks(self) -> int:
nbs = []
state_keys = self.state_map[r"model.1.sub.\1.RDB\2.conv\3.0.\4"] + (
r"model\.\d+\.sub\.(\d+)\.RDB(\d+)\.conv(\d+)\.0\.(weight|bias)",
)
for state_key in state_keys:
for k in self.state:
m = re.search(state_key, k)
if m:
nbs.append(int(m.group(1)))
if nbs:
break
return max(*nbs) + 1
def forward(self, x):
if self.shuffle_factor:
_, _, h, w = x.size()
mod_pad_h = (
self.shuffle_factor - h % self.shuffle_factor
) % self.shuffle_factor
mod_pad_w = (
self.shuffle_factor - w % self.shuffle_factor
) % self.shuffle_factor
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
x = torch.pixel_unshuffle(x, downscale_factor=self.shuffle_factor)
x = self.model(x)
return x[:, :, : h * self.scale, : w * self.scale]
return self.model(x)
comfy_extras/chainner_models/architecture/SPSR.py 0 → 100644
View file @ 905857ed
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from . import block as B
class Get_gradient_nopadding(nn.Module):
def __init__(self):
super(Get_gradient_nopadding, self).__init__()
kernel_v = [[0, -1, 0], [0, 0, 0], [0, 1, 0]]
kernel_h = [[0, 0, 0], [-1, 0, 1], [0, 0, 0]]
kernel_h = torch.FloatTensor(kernel_h).unsqueeze(0).unsqueeze(0)
kernel_v = torch.FloatTensor(kernel_v).unsqueeze(0).unsqueeze(0)
self.weight_h = nn.Parameter(data=kernel_h, requires_grad=False) # type: ignore
self.weight_v = nn.Parameter(data=kernel_v, requires_grad=False) # type: ignore
def forward(self, x):
x_list = []
for i in range(x.shape[1]):
x_i = x[:, i]
x_i_v = F.conv2d(x_i.unsqueeze(1), self.weight_v, padding=1)
x_i_h = F.conv2d(x_i.unsqueeze(1), self.weight_h, padding=1)
x_i = torch.sqrt(torch.pow(x_i_v, 2) + torch.pow(x_i_h, 2) + 1e-6)
x_list.append(x_i)
x = torch.cat(x_list, dim=1)
return x
class SPSRNet(nn.Module):
def __init__(
self,
state_dict,
norm=None,
act: str = "leakyrelu",
upsampler: str = "upconv",
mode: B.ConvMode = "CNA",
):
super(SPSRNet, self).__init__()
self.model_arch = "SPSR"
self.sub_type = "SR"
self.state = state_dict
self.norm = norm
self.act = act
self.upsampler = upsampler
self.mode = mode
self.num_blocks = self.get_num_blocks()
self.in_nc: int = self.state["model.0.weight"].shape[1]
self.out_nc: int = self.state["f_HR_conv1.0.bias"].shape[0]
self.scale = self.get_scale(4)
print(self.scale)
self.num_filters: int = self.state["model.0.weight"].shape[0]
self.supports_fp16 = True
self.supports_bfp16 = True
self.min_size_restriction = None
n_upscale = int(math.log(self.scale, 2))
if self.scale == 3:
n_upscale = 1
fea_conv = B.conv_block(
self.in_nc, self.num_filters, kernel_size=3, norm_type=None, act_type=None
)
rb_blocks = [
B.RRDB(
self.num_filters,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=norm,
act_type=act,
mode="CNA",
)
for _ in range(self.num_blocks)
]
LR_conv = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=norm,
act_type=None,
mode=mode,
)
if upsampler == "upconv":
upsample_block = B.upconv_block
elif upsampler == "pixelshuffle":
upsample_block = B.pixelshuffle_block
else:
raise NotImplementedError(f"upsample mode [{upsampler}] is not found")
if self.scale == 3:
a_upsampler = upsample_block(
self.num_filters, self.num_filters, 3, act_type=act
)
else:
a_upsampler = [
upsample_block(self.num_filters, self.num_filters, act_type=act)
for _ in range(n_upscale)
]
self.HR_conv0_new = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=act,
)
self.HR_conv1_new = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.model = B.sequential(
fea_conv,
B.ShortcutBlockSPSR(B.sequential(*rb_blocks, LR_conv)),
*a_upsampler,
self.HR_conv0_new,
)
self.get_g_nopadding = Get_gradient_nopadding()
self.b_fea_conv = B.conv_block(
self.in_nc, self.num_filters, kernel_size=3, norm_type=None, act_type=None
)
self.b_concat_1 = B.conv_block(
2 * self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.b_block_1 = B.RRDB(
self.num_filters * 2,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=norm,
act_type=act,
mode="CNA",
)
self.b_concat_2 = B.conv_block(
2 * self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.b_block_2 = B.RRDB(
self.num_filters * 2,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=norm,
act_type=act,
mode="CNA",
)
self.b_concat_3 = B.conv_block(
2 * self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.b_block_3 = B.RRDB(
self.num_filters * 2,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=norm,
act_type=act,
mode="CNA",
)
self.b_concat_4 = B.conv_block(
2 * self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.b_block_4 = B.RRDB(
self.num_filters * 2,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=norm,
act_type=act,
mode="CNA",
)
self.b_LR_conv = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=norm,
act_type=None,
mode=mode,
)
if upsampler == "upconv":
upsample_block = B.upconv_block
elif upsampler == "pixelshuffle":
upsample_block = B.pixelshuffle_block
else:
raise NotImplementedError(f"upsample mode [{upsampler}] is not found")
if self.scale == 3:
b_upsampler = upsample_block(
self.num_filters, self.num_filters, 3, act_type=act
)
else:
b_upsampler = [
upsample_block(self.num_filters, self.num_filters, act_type=act)
for _ in range(n_upscale)
]
b_HR_conv0 = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=act,
)
b_HR_conv1 = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.b_module = B.sequential(*b_upsampler, b_HR_conv0, b_HR_conv1)
self.conv_w = B.conv_block(
self.num_filters, self.out_nc, kernel_size=1, norm_type=None, act_type=None
)
self.f_concat = B.conv_block(
self.num_filters * 2,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=None,
)
self.f_block = B.RRDB(
self.num_filters * 2,
kernel_size=3,
gc=32,
stride=1,
bias=True,
pad_type="zero",
norm_type=norm,
act_type=act,
mode="CNA",
)
self.f_HR_conv0 = B.conv_block(
self.num_filters,
self.num_filters,
kernel_size=3,
norm_type=None,
act_type=act,
)
self.f_HR_conv1 = B.conv_block(
self.num_filters, self.out_nc, kernel_size=3, norm_type=None, act_type=None
)
self.load_state_dict(self.state, strict=False)
def get_scale(self, min_part: int = 4) -> int:
n = 0
for part in list(self.state):
parts = part.split(".")
if len(parts) == 3:
part_num = int(parts[1])
if part_num > min_part and parts[0] == "model" and parts[2] == "weight":
n += 1
return 2**n
def get_num_blocks(self) -> int:
nb = 0
for part in list(self.state):
parts = part.split(".")
n_parts = len(parts)
if n_parts == 5 and parts[2] == "sub":
nb = int(parts[3])
return nb
def forward(self, x):
x_grad = self.get_g_nopadding(x)
x = self.model[0](x)
x, block_list = self.model[1](x)
x_ori = x
for i in range(5):
x = block_list[i](x)
x_fea1 = x
for i in range(5):
x = block_list[i + 5](x)
x_fea2 = x
for i in range(5):
x = block_list[i + 10](x)
x_fea3 = x
for i in range(5):
x = block_list[i + 15](x)
x_fea4 = x
x = block_list[20:](x)
# short cut
x = x_ori + x
x = self.model[2:](x)
x = self.HR_conv1_new(x)
x_b_fea = self.b_fea_conv(x_grad)
x_cat_1 = torch.cat([x_b_fea, x_fea1], dim=1)
x_cat_1 = self.b_block_1(x_cat_1)
x_cat_1 = self.b_concat_1(x_cat_1)
x_cat_2 = torch.cat([x_cat_1, x_fea2], dim=1)
x_cat_2 = self.b_block_2(x_cat_2)
x_cat_2 = self.b_concat_2(x_cat_2)
x_cat_3 = torch.cat([x_cat_2, x_fea3], dim=1)
x_cat_3 = self.b_block_3(x_cat_3)
x_cat_3 = self.b_concat_3(x_cat_3)
x_cat_4 = torch.cat([x_cat_3, x_fea4], dim=1)
x_cat_4 = self.b_block_4(x_cat_4)
x_cat_4 = self.b_concat_4(x_cat_4)
x_cat_4 = self.b_LR_conv(x_cat_4)
# short cut
x_cat_4 = x_cat_4 + x_b_fea
x_branch = self.b_module(x_cat_4)
# x_out_branch = self.conv_w(x_branch)
########
x_branch_d = x_branch
x_f_cat = torch.cat([x_branch_d, x], dim=1)
x_f_cat = self.f_block(x_f_cat)
x_out = self.f_concat(x_f_cat)
x_out = self.f_HR_conv0(x_out)
x_out = self.f_HR_conv1(x_out)
#########
# return x_out_branch, x_out, x_grad
return x_out
comfy_extras/chainner_models/architecture/SRVGG.py 0 → 100644
View file @ 905857ed
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import math
import torch.nn as nn
import torch.nn.functional as F
class SRVGGNetCompact(nn.Module):
"""A compact VGG-style network structure for super-resolution.
It is a compact network structure, which performs upsampling in the last layer and no convolution is
conducted on the HR feature space.
Args:
num_in_ch (int): Channel number of inputs. Default: 3.
num_out_ch (int): Channel number of outputs. Default: 3.
num_feat (int): Channel number of intermediate features. Default: 64.
num_conv (int): Number of convolution layers in the body network. Default: 16.
upscale (int): Upsampling factor. Default: 4.
act_type (str): Activation type, options: 'relu', 'prelu', 'leakyrelu'. Default: prelu.
"""
def __init__(
self,
state_dict,
act_type: str = "prelu",
):
super(SRVGGNetCompact, self).__init__()
self.model_arch = "SRVGG (RealESRGAN)"
self.sub_type = "SR"
self.act_type = act_type
self.state = state_dict
if "params" in self.state:
self.state = self.state["params"]
self.key_arr = list(self.state.keys())
self.in_nc = self.get_in_nc()
self.num_feat = self.get_num_feats()
self.num_conv = self.get_num_conv()
self.out_nc = self.in_nc # :(
self.pixelshuffle_shape = None # Defined in get_scale()
self.scale = self.get_scale()
self.supports_fp16 = True
self.supports_bfp16 = True
self.min_size_restriction = None
self.body = nn.ModuleList()
# the first conv
self.body.append(nn.Conv2d(self.in_nc, self.num_feat, 3, 1, 1))
# the first activation
if act_type == "relu":
activation = nn.ReLU(inplace=True)
elif act_type == "prelu":
activation = nn.PReLU(num_parameters=self.num_feat)
elif act_type == "leakyrelu":
activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)
self.body.append(activation) # type: ignore
# the body structure
for _ in range(self.num_conv):
self.body.append(nn.Conv2d(self.num_feat, self.num_feat, 3, 1, 1))
# activation
if act_type == "relu":
activation = nn.ReLU(inplace=True)
elif act_type == "prelu":
activation = nn.PReLU(num_parameters=self.num_feat)
elif act_type == "leakyrelu":
activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)
self.body.append(activation) # type: ignore
# the last conv
self.body.append(nn.Conv2d(self.num_feat, self.pixelshuffle_shape, 3, 1, 1)) # type: ignore
# upsample
self.upsampler = nn.PixelShuffle(self.scale)
self.load_state_dict(self.state, strict=False)
def get_num_conv(self) -> int:
return (int(self.key_arr[-1].split(".")[1]) - 2) // 2
def get_num_feats(self) -> int:
return self.state[self.key_arr[0]].shape[0]
def get_in_nc(self) -> int:
return self.state[self.key_arr[0]].shape[1]
def get_scale(self) -> int:
self.pixelshuffle_shape = self.state[self.key_arr[-1]].shape[0]
# Assume out_nc is the same as in_nc
# I cant think of a better way to do that
self.out_nc = self.in_nc
scale = math.sqrt(self.pixelshuffle_shape / self.out_nc)
if scale - int(scale) > 0:
print(
"out_nc is probably different than in_nc, scale calculation might be wrong"
)
scale = int(scale)
return scale
def forward(self, x):
out = x
for i in range(0, len(self.body)):
out = self.body[i](out)
out = self.upsampler(out)
# add the nearest upsampled image, so that the network learns the residual
base = F.interpolate(x, scale_factor=self.scale, mode="nearest")
out += base
return out
comfy_extras/chainner_models/architecture/SwiftSRGAN.py 0 → 100644
View file @ 905857ed
# From https://github.com/Koushik0901/Swift-SRGAN/blob/master/swift-srgan/models.py
import torch
from torch import nn
class SeperableConv2d(nn.Module):
def __init__(
self, in_channels, out_channels, kernel_size, stride=1, padding=1, bias=True
):
super(SeperableConv2d, self).__init__()
self.depthwise = nn.Conv2d(
in_channels,
in_channels,
kernel_size=kernel_size,
stride=stride,
groups=in_channels,
bias=bias,
padding=padding,
)
self.pointwise = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=bias)
def forward(self, x):
return self.pointwise(self.depthwise(x))
class ConvBlock(nn.Module):
def __init__(
self,
in_channels,
out_channels,
use_act=True,
use_bn=True,
discriminator=False,
**kwargs,
):
super(ConvBlock, self).__init__()
self.use_act = use_act
self.cnn = SeperableConv2d(in_channels, out_channels, **kwargs, bias=not use_bn)
self.bn = nn.BatchNorm2d(out_channels) if use_bn else nn.Identity()
self.act = (
nn.LeakyReLU(0.2, inplace=True)
if discriminator
else nn.PReLU(num_parameters=out_channels)
)
def forward(self, x):
return self.act(self.bn(self.cnn(x))) if self.use_act else self.bn(self.cnn(x))
class UpsampleBlock(nn.Module):
def __init__(self, in_channels, scale_factor):
super(UpsampleBlock, self).__init__()
self.conv = SeperableConv2d(
in_channels,
in_channels * scale_factor**2,
kernel_size=3,
stride=1,
padding=1,
)
self.ps = nn.PixelShuffle(
scale_factor
) # (in_channels * 4, H, W) -> (in_channels, H*2, W*2)
self.act = nn.PReLU(num_parameters=in_channels)
def forward(self, x):
return self.act(self.ps(self.conv(x)))
class ResidualBlock(nn.Module):
def __init__(self, in_channels):
super(ResidualBlock, self).__init__()
self.block1 = ConvBlock(
in_channels, in_channels, kernel_size=3, stride=1, padding=1
)
self.block2 = ConvBlock(
in_channels, in_channels, kernel_size=3, stride=1, padding=1, use_act=False
)
def forward(self, x):
out = self.block1(x)
out = self.block2(out)
return out + x
class Generator(nn.Module):
"""Swift-SRGAN Generator
Args:
in_channels (int): number of input image channels.
num_channels (int): number of hidden channels.
num_blocks (int): number of residual blocks.
upscale_factor (int): factor to upscale the image [2x, 4x, 8x].
Returns:
torch.Tensor: super resolution image
"""
def __init__(
self,
state_dict,
):
super(Generator, self).__init__()
self.model_arch = "Swift-SRGAN"
self.sub_type = "SR"
self.state = state_dict
if "model" in self.state:
self.state = self.state["model"]
self.in_nc: int = self.state["initial.cnn.depthwise.weight"].shape[0]
self.out_nc: int = self.state["final_conv.pointwise.weight"].shape[0]
self.num_filters: int = self.state["initial.cnn.pointwise.weight"].shape[0]
self.num_blocks = len(
set([x.split(".")[1] for x in self.state.keys() if "residual" in x])
)
self.scale: int = 2 ** len(
set([x.split(".")[1] for x in self.state.keys() if "upsampler" in x])
)
in_channels = self.in_nc
num_channels = self.num_filters
num_blocks = self.num_blocks
upscale_factor = self.scale
self.supports_fp16 = True
self.supports_bfp16 = True
self.min_size_restriction = None
self.initial = ConvBlock(
in_channels, num_channels, kernel_size=9, stride=1, padding=4, use_bn=False
)
self.residual = nn.Sequential(
*[ResidualBlock(num_channels) for _ in range(num_blocks)]
)
self.convblock = ConvBlock(
num_channels,
num_channels,
kernel_size=3,
stride=1,
padding=1,
use_act=False,
)
self.upsampler = nn.Sequential(
*[
UpsampleBlock(num_channels, scale_factor=2)
for _ in range(upscale_factor // 2)
]
)
self.final_conv = SeperableConv2d(
num_channels, in_channels, kernel_size=9, stride=1, padding=4
)
self.load_state_dict(self.state, strict=False)
def forward(self, x):
initial = self.initial(x)
x = self.residual(initial)
x = self.convblock(x) + initial
x = self.upsampler(x)
return (torch.tanh(self.final_conv(x)) + 1) / 2
comfy_extras/chainner_models/architecture/Swin2SR.py 0 → 100644
View file @ 905857ed
# pylint: skip-file
# -----------------------------------------------------------------------------------
# Swin2SR: Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration, https://arxiv.org/abs/2209.11345
# Written by Conde and Choi et al.
# From: https://raw.githubusercontent.com/mv-lab/swin2sr/main/models/network_swin2sr.py
# -----------------------------------------------------------------------------------
import math
import re
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
# Originally from the timm package
from .timm.drop import DropPath
from .timm.helpers import to_2tuple
from .timm.weight_init import trunc_normal_
class Mlp(nn.Module):
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.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 window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
windows = (
x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
)
return windows
def window_reverse(windows, window_size, H, W):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(
B, H // window_size, W // window_size, window_size, window_size, -1
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
class WindowAttention(nn.Module):
r"""Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
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_window_size (tuple[int]): The height and width of the window in pre-training.
"""
def __init__(
self,
dim,
window_size,
num_heads,
qkv_bias=True,
attn_drop=0.0,
proj_drop=0.0,
pretrained_window_size=[0, 0],
):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.pretrained_window_size = pretrained_window_size
self.num_heads = num_heads
self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True) # type: ignore
# mlp to generate continuous relative position bias
self.cpb_mlp = nn.Sequential(
nn.Linear(2, 512, bias=True),
nn.ReLU(inplace=True),
nn.Linear(512, num_heads, bias=False),
)
# get relative_coords_table
relative_coords_h = torch.arange(
-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32
)
relative_coords_w = torch.arange(
-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32
)
relative_coords_table = (
torch.stack(torch.meshgrid([relative_coords_h, relative_coords_w]))
.permute(1, 2, 0)
.contiguous()
.unsqueeze(0)
) # 1, 2*Wh-1, 2*Ww-1, 2
if pretrained_window_size[0] > 0:
relative_coords_table[:, :, :, 0] /= pretrained_window_size[0] - 1
relative_coords_table[:, :, :, 1] /= pretrained_window_size[1] - 1
else:
relative_coords_table[:, :, :, 0] /= self.window_size[0] - 1
relative_coords_table[:, :, :, 1] /= self.window_size[1] - 1
relative_coords_table *= 8 # normalize to -8, 8
relative_coords_table = (
torch.sign(relative_coords_table)
* torch.log2(torch.abs(relative_coords_table) + 1.0)
/ np.log2(8)
)
self.register_buffer("relative_coords_table", relative_coords_table)
# get pair-wise relative position index for each token inside the window
coords_h = torch.arange(self.window_size[0])
coords_w = torch.arange(self.window_size[1])
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
relative_coords = (
coords_flatten[:, :, None] - coords_flatten[:, None, :]
) # 2, Wh*Ww, Wh*Ww
relative_coords = relative_coords.permute(
1, 2, 0
).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size[1] - 1
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
self.qkv = nn.Linear(dim, dim * 3, bias=False)
if qkv_bias:
self.q_bias = nn.Parameter(torch.zeros(dim)) # type: ignore
self.v_bias = nn.Parameter(torch.zeros(dim)) # type: ignore
else:
self.q_bias = None
self.v_bias = None
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
"""
Args:
x: input features with shape of (num_windows*B, N, C)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
B_, N, C = x.shape
qkv_bias = None
if self.q_bias is not None:
qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias)) # type: ignore
qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
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],
) # make torchscript happy (cannot use tensor as tuple)
# cosine attention
attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
logit_scale = torch.clamp(
self.logit_scale,
max=torch.log(torch.tensor(1.0 / 0.01)).to(self.logit_scale.device),
).exp()
attn = attn * logit_scale
relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(
-1, self.num_heads
)
relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view( # type: ignore
self.window_size[0] * self.window_size[1],
self.window_size[0] * self.window_size[1],
-1,
) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous() # nH, Wh*Ww, Wh*Ww
relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
attn = attn + relative_position_bias.unsqueeze(0)
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
1
).unsqueeze(0)
attn = attn.view(-1, self.num_heads, N, N)
attn = self.softmax(attn)
else:
attn = self.softmax(attn)
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 extra_repr(self) -> str:
return (
f"dim={self.dim}, window_size={self.window_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 SwinTransformerBlock(nn.Module):
r"""Swin Transformer Block.
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.
shift_size (int): Shift size for SW-MSA.
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_window_size (int): Window size in pre-training.
"""
def __init__(
self,
dim,
input_resolution,
num_heads,
window_size=7,
shift_size=0,
mlp_ratio=4.0,
qkv_bias=True,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
pretrained_window_size=0,
):
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,
attn_drop=attn_drop,
proj_drop=drop,
pretrained_window_size=to_2tuple(pretrained_window_size),
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 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:
attn_mask = self.calculate_mask(self.input_resolution)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
H, W = x_size
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)
)
return attn_mask
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
# assert L == H * W, "input feature has wrong size"
shortcut = 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)
)
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 (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows = self.attn(
x_windows, mask=self.attn_mask
) # nW*B, window_size*window_size, C
else:
attn_windows = self.attn(
x_windows, mask=self.calculate_mask(x_size).to(x.device)
)
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(
shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
)
else:
x = shifted_x
x = x.view(B, H * W, C)
x = shortcut + self.drop_path(self.norm1(x))
# FFN
x = x + 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}, "
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 PatchMerging(nn.Module):
r"""Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(2 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
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)
x = self.norm(x)
return x
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
flops += H * W * self.dim // 2
return flops
class BasicLayer(nn.Module):
"""A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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 | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
pretrained_window_size (int): Local window size in pre-training.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
mlp_ratio=4.0,
qkv_bias=True,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
pretrained_window_size=0,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList(
[
SwinTransformerBlock(
dim=dim,
input_resolution=input_resolution,
num_heads=num_heads,
window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i]
if isinstance(drop_path, list)
else drop_path,
norm_layer=norm_layer,
pretrained_window_size=pretrained_window_size,
)
for i in range(depth)
]
)
# patch merging layer
if downsample is not None:
self.downsample = downsample(
input_resolution, dim=dim, norm_layer=norm_layer
)
else:
self.downsample = None
def forward(self, x, x_size):
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
if self.downsample is not None:
x = self.downsample(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
def flops(self):
flops = 0
for blk in self.blocks:
flops += blk.flops() # type: ignore
if self.downsample is not None:
flops += self.downsample.flops()
return flops
def _init_respostnorm(self):
for blk in self.blocks:
nn.init.constant_(blk.norm1.bias, 0) # type: ignore
nn.init.constant_(blk.norm1.weight, 0) # type: ignore
nn.init.constant_(blk.norm2.bias, 0) # type: ignore
nn.init.constant_(blk.norm2.weight, 0) # type: ignore
class PatchEmbed(nn.Module):
r"""Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] # type: ignore
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
self.proj = nn.Conv2d(
in_chans, embed_dim, kernel_size=patch_size, stride=patch_size # type: ignore
)
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
B, C, H, W = x.shape
# FIXME look at relaxing size constraints
# assert H == self.img_size[0] and W == self.img_size[1],
# f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.proj(x).flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
Ho, Wo = self.patches_resolution
flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1]) # type: ignore
if self.norm is not None:
flops += Ho * Wo * self.embed_dim
return flops
class RSTB(nn.Module):
"""Residual Swin Transformer Block (RSTB).
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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 | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
img_size: Input image size.
patch_size: Patch size.
resi_connection: The convolutional block before residual connection.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
mlp_ratio=4.0,
qkv_bias=True,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
img_size=224,
patch_size=4,
resi_connection="1conv",
):
super(RSTB, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.residual_group = BasicLayer(
dim=dim,
input_resolution=input_resolution,
depth=depth,
num_heads=num_heads,
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path,
norm_layer=norm_layer,
downsample=downsample,
use_checkpoint=use_checkpoint,
)
if resi_connection == "1conv":
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == "3conv":
# to save parameters and memory
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),
)
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=dim,
embed_dim=dim,
norm_layer=None,
)
self.patch_unembed = PatchUnEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=dim,
embed_dim=dim,
norm_layer=None,
)
def forward(self, x, x_size):
return (
self.patch_embed(
self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))
)
+ x
)
def flops(self):
flops = 0
flops += self.residual_group.flops()
H, W = self.input_resolution
flops += H * W * self.dim * self.dim * 9
flops += self.patch_embed.flops()
flops += self.patch_unembed.flops()
return flops
class PatchUnEmbed(nn.Module):
r"""Image to Patch Unembedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] # type: ignore
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
B, HW, C = x.shape
x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
return x
def flops(self):
flops = 0
return flops
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 Upsample_hf(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_hf, 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 # type: ignore
flops = H * W * self.num_feat * 3 * 9
return flops
class Swin2SR(nn.Module):
r"""Swin2SR
A PyTorch impl of : `Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration`.
Args:
img_size (int | tuple(int)): Input image size. Default 64
patch_size (int | tuple(int)): Patch size. Default: 1
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
window_size (int): Window size. Default: 7
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
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
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
patch_norm (bool): If True, add normalization after patch embedding. Default: True
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
img_range: Image range. 1. or 255.
upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(
self,
state_dict,
**kwargs,
):
super(Swin2SR, self).__init__()
# Defaults
img_size = 128
patch_size = 1
in_chans = 3
embed_dim = 96
depths = [6, 6, 6, 6]
num_heads = [6, 6, 6, 6]
window_size = 7
mlp_ratio = 4.0
qkv_bias = True
drop_rate = 0.0
attn_drop_rate = 0.0
drop_path_rate = 0.1
norm_layer = nn.LayerNorm
ape = False
patch_norm = True
use_checkpoint = False
upscale = 2
img_range = 1.0
upsampler = ""
resi_connection = "1conv"
num_in_ch = in_chans
num_out_ch = in_chans
num_feat = 64
self.model_arch = "Swin2SR"
self.sub_type = "SR"
self.state = state_dict
if "params_ema" in self.state:
self.state = self.state["params_ema"]
elif "params" in self.state:
self.state = self.state["params"]
state_keys = self.state.keys()
if "conv_before_upsample.0.weight" in state_keys:
if "conv_aux.weight" in state_keys:
upsampler = "pixelshuffle_aux"
elif "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 = (
self.state.get("conv_before_upsample.0.weight", None).shape[1]
if self.state.get("conv_before_upsample.weight", None)
else 64
)
num_in_ch = self.state["conv_first.weight"].shape[1]
in_chans = num_in_ch
if "conv_last.weight" in state_keys:
num_out_ch = self.state["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" or upsampler == "pixelshuffle_aux":
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 = self.state[upsample_key].shape[0]
upscale *= math.sqrt(shape // num_feat)
upscale = int(upscale)
elif upsampler == "pixelshuffledirect":
upscale = int(
math.sqrt(self.state["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*).residual_group.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))
depths = [max_block_num + 1 for _ in range(max_layer_num + 1)]
if (
"layers.0.residual_group.blocks.0.attn.relative_position_bias_table"
in state_keys
):
num_heads_num = self.state[
"layers.0.residual_group.blocks.0.attn.relative_position_bias_table"
].shape[-1]
num_heads = [num_heads_num for _ in range(max_layer_num + 1)]
else:
num_heads = depths
embed_dim = self.state["conv_first.weight"].shape[0]
mlp_ratio = float(
self.state["layers.0.residual_group.blocks.0.mlp.fc1.bias"].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"
window_size = int(
math.sqrt(
self.state[
"layers.0.residual_group.blocks.0.attn.relative_position_index"
].shape[0]
)
)
if "layers.0.residual_group.blocks.1.attn_mask" in state_keys:
img_size = int(
math.sqrt(
self.state["layers.0.residual_group.blocks.1.attn_mask"].shape[0]
)
* window_size
)
# The JPEG models are the only ones with window-size 7, and they also use this range
img_range = 255.0 if window_size == 7 else 1.0
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.depths = depths
self.window_size = window_size
self.mlp_ratio = mlp_ratio
self.scale = upscale
self.upsampler = upsampler
self.img_size = img_size
self.img_range = img_range
self.resi_connection = resi_connection
self.supports_fp16 = False # Too much weirdness to support this at the moment
self.supports_bfp16 = True
self.min_size_restriction = 16
## END AUTO DETECTION
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
self.window_size = window_size
#####################################################################################################
################################### 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(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = embed_dim
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=embed_dim,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None,
)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# merge non-overlapping patches into image
self.patch_unembed = PatchUnEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=embed_dim,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None,
)
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) # type: ignore
trunc_normal_(self.absolute_pos_embed, std=0.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
] # stochastic depth decay rule
# build Residual Swin Transformer blocks (RSTB)
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RSTB(
dim=embed_dim,
input_resolution=(patches_resolution[0], patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], # type: ignore # no impact on SR results
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection,
)
self.layers.append(layer)
if self.upsampler == "pixelshuffle_hf":
self.layers_hf = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RSTB(
dim=embed_dim,
input_resolution=(patches_resolution[0], patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])], # type: ignore # no impact on SR results # type: ignore
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection,
)
self.layers_hf.append(layer)
self.norm = norm_layer(self.num_features)
# 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, high quality image 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 == "pixelshuffle_aux":
self.conv_bicubic = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.conv_aux = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.conv_after_aux = nn.Sequential(
nn.Conv2d(3, 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 == "pixelshuffle_hf":
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.upsample_hf = Upsample_hf(upscale, num_feat)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.conv_first_hf = nn.Sequential(
nn.Conv2d(num_feat, embed_dim, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.conv_after_body_hf = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
self.conv_before_upsample_hf = nn.Sequential(
nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.conv_last_hf = 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,
(patches_resolution[0], patches_resolution[1]),
)
elif self.upsampler == "nearest+conv":
# for real-world SR (less artifacts)
assert self.upscale == 4, "only support x4 now."
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
else:
# for image denoising and JPEG compression artifact reduction
self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
self.apply(self._init_weights)
self.load_state_dict(state_dict)
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.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore # type: ignore
def no_weight_decay(self):
return {"absolute_pos_embed"}
@torch.jit.ignore # type: ignore
def no_weight_decay_keywords(self):
return {"relative_position_bias_table"}
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
return x
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x) # B L C
x = self.patch_unembed(x, x_size)
return x
def forward_features_hf(self, x):
x_size = (x.shape[2], x.shape[3])
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers_hf:
x = layer(x, x_size)
x = self.norm(x) # B L C
x = self.patch_unembed(x, x_size)
return x
def forward(self, x):
H, W = x.shape[2:]
x = self.check_image_size(x)
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == "pixelshuffle":
# for classical 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 == "pixelshuffle_aux":
bicubic = F.interpolate(
x,
size=(H * self.upscale, W * self.upscale),
mode="bicubic",
align_corners=False,
)
bicubic = self.conv_bicubic(bicubic)
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
aux = self.conv_aux(x) # b, 3, LR_H, LR_W
x = self.conv_after_aux(aux)
x = (
self.upsample(x)[:, :, : H * self.upscale, : W * self.upscale]
+ bicubic[:, :, : H * self.upscale, : W * self.upscale]
)
x = self.conv_last(x)
aux = aux / self.img_range + self.mean
elif self.upsampler == "pixelshuffle_hf":
# for classical SR with HF
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x_before = self.conv_before_upsample(x)
x_out = self.conv_last(self.upsample(x_before))
x_hf = self.conv_first_hf(x_before)
x_hf = self.conv_after_body_hf(self.forward_features_hf(x_hf)) + x_hf
x_hf = self.conv_before_upsample_hf(x_hf)
x_hf = self.conv_last_hf(self.upsample_hf(x_hf))
x = x_out + x_hf
x_hf = x_hf / self.img_range + self.mean
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)
elif self.upsampler == "nearest+conv":
# for real-world SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.lrelu(
self.conv_up1(
torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest")
)
)
x = self.lrelu(
self.conv_up2(
torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest")
)
)
x = self.conv_last(self.lrelu(self.conv_hr(x)))
else:
# for image denoising and JPEG compression artifact reduction
x_first = self.conv_first(x)
res = self.conv_after_body(self.forward_features(x_first)) + x_first
x = x + self.conv_last(res)
x = x / self.img_range + self.mean
if self.upsampler == "pixelshuffle_aux":
# NOTE: I removed an "aux" output here. not sure what that was for
return x[:, :, : H * self.upscale, : W * self.upscale] # type: ignore
elif self.upsampler == "pixelshuffle_hf":
x_out = x_out / self.img_range + self.mean # type: ignore
return x_out[:, :, : H * self.upscale, : W * self.upscale], x[:, :, : H * self.upscale, : W * self.upscale], x_hf[:, :, : H * self.upscale, : W * self.upscale] # type: ignore
else:
return x[:, :, : H * self.upscale, : W * self.upscale]
def flops(self):
flops = 0
H, W = self.patches_resolution
flops += H * W * 3 * self.embed_dim * 9
flops += self.patch_embed.flops()
for i, layer in enumerate(self.layers):
flops += layer.flops() # type: ignore
flops += H * W * 3 * self.embed_dim * self.embed_dim
flops += self.upsample.flops() # type: ignore
return flops
comfy_extras/chainner_models/architecture/SwinIR.py 0 → 100644
View file @ 905857ed
# pylint: skip-file
# -----------------------------------------------------------------------------------
# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
# Originally Written by Ze Liu, Modified by Jingyun Liang.
# -----------------------------------------------------------------------------------
import math
import re
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
# Originally from the timm package
from .timm.drop import DropPath
from .timm.helpers import to_2tuple
from .timm.weight_init import trunc_normal_
class Mlp(nn.Module):
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.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 window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
windows = (
x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
)
return windows
def window_reverse(windows, window_size, H, W):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(
B, H // window_size, W // window_size, window_size, window_size, -1
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
class WindowAttention(nn.Module):
r"""Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
"""
def __init__(
self,
dim,
window_size,
num_heads,
qkv_bias=True,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim**-0.5
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter( # type: ignore
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
) # 2*Wh-1 * 2*Ww-1, nH
# get pair-wise relative position index for each token inside the window
coords_h = torch.arange(self.window_size[0])
coords_w = torch.arange(self.window_size[1])
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
relative_coords = (
coords_flatten[:, :, None] - coords_flatten[:, None, :]
) # 2, Wh*Ww, Wh*Ww
relative_coords = relative_coords.permute(
1, 2, 0
).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size[1] - 1
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
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)
trunc_normal_(self.relative_position_bias_table, std=0.02)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
"""
Args:
x: input features with shape of (num_windows*B, N, C)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
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],
) # make torchscript happy (cannot use tensor as tuple)
q = q * self.scale
attn = q @ k.transpose(-2, -1)
relative_position_bias = self.relative_position_bias_table[
self.relative_position_index.view(-1) # type: ignore
].view(
self.window_size[0] * self.window_size[1],
self.window_size[0] * self.window_size[1],
-1,
) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous() # nH, Wh*Ww, Wh*Ww
attn = attn + relative_position_bias.unsqueeze(0)
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
1
).unsqueeze(0)
attn = attn.view(-1, self.num_heads, N, N)
attn = self.softmax(attn)
else:
attn = self.softmax(attn)
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 extra_repr(self) -> str:
return f"dim={self.dim}, window_size={self.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 SwinTransformerBlock(nn.Module):
r"""Swin Transformer Block.
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.
shift_size (int): Shift size for SW-MSA.
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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
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
"""
def __init__(
self,
dim,
input_resolution,
num_heads,
window_size=7,
shift_size=0,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
):
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 > 0.0 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:
attn_mask = self.calculate_mask(self.input_resolution)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
H, W = x_size
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)
)
return attn_mask
def forward(self, x, x_size):
H, W = x_size
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)
)
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 (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows = self.attn(
x_windows, mask=self.attn_mask
) # nW*B, window_size*window_size, C
else:
attn_windows = self.attn(
x_windows, mask=self.calculate_mask(x_size).to(x.device)
)
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(
shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
)
else:
x = shifted_x
x = x.view(B, H * W, C)
# FFN
x = shortcut + self.drop_path(x)
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 PatchMerging(nn.Module):
r"""Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
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.norm(x)
x = self.reduction(x)
return x
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = H * W * self.dim
flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
return flops
class BasicLayer(nn.Module):
"""A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList(
[
SwinTransformerBlock(
dim=dim,
input_resolution=input_resolution,
num_heads=num_heads,
window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i]
if isinstance(drop_path, list)
else drop_path,
norm_layer=norm_layer,
)
for i in range(depth)
]
)
# patch merging layer
if downsample is not None:
self.downsample = downsample(
input_resolution, dim=dim, norm_layer=norm_layer
)
else:
self.downsample = None
def forward(self, x, x_size):
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
if self.downsample is not None:
x = self.downsample(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
def flops(self):
flops = 0
for blk in self.blocks:
flops += blk.flops() # type: ignore
if self.downsample is not None:
flops += self.downsample.flops()
return flops
class RSTB(nn.Module):
"""Residual Swin Transformer Block (RSTB).
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local 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
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
img_size: Input image size.
patch_size: Patch size.
resi_connection: The convolutional block before residual connection.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads,
window_size,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False,
img_size=224,
patch_size=4,
resi_connection="1conv",
):
super(RSTB, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.residual_group = BasicLayer(
dim=dim,
input_resolution=input_resolution,
depth=depth,
num_heads=num_heads,
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path,
norm_layer=norm_layer,
downsample=downsample,
use_checkpoint=use_checkpoint,
)
if resi_connection == "1conv":
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == "3conv":
# to save parameters and memory
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),
)
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=0,
embed_dim=dim,
norm_layer=None,
)
self.patch_unembed = PatchUnEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=0,
embed_dim=dim,
norm_layer=None,
)
def forward(self, x, x_size):
return (
self.patch_embed(
self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))
)
+ x
)
def flops(self):
flops = 0
flops += self.residual_group.flops()
H, W = self.input_resolution
flops += H * W * self.dim * self.dim * 9
flops += self.patch_embed.flops()
flops += self.patch_unembed.flops()
return flops
class PatchEmbed(nn.Module):
r"""Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [
img_size[0] // patch_size[0], # type: ignore
img_size[1] // patch_size[1], # type: ignore
]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
x = x.flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
flops = 0
H, W = self.img_size
if self.norm is not None:
flops += H * W * self.embed_dim # type: ignore
return flops
class PatchUnEmbed(nn.Module):
r"""Image to Patch Unembedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [
img_size[0] // patch_size[0], # type: ignore
img_size[1] // patch_size[1], # type: ignore
]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
B, HW, C = x.shape
x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
return x
def flops(self):
flops = 0
return flops
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 # type: ignore
flops = H * W * self.num_feat * 3 * 9
return flops
class SwinIR(nn.Module):
r"""SwinIR
A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
Args:
img_size (int | tuple(int)): Input image size. Default 64
patch_size (int | tuple(int)): Patch size. Default: 1
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
window_size (int): Window size. Default: 7
mlp_ratio (float): Ratio of mlp 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): 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
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
patch_norm (bool): If True, add normalization after patch embedding. Default: True
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
img_range: Image range. 1. or 255.
upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(
self,
state_dict,
**kwargs,
):
super(SwinIR, self).__init__()
# Defaults
img_size = 64
patch_size = 1
in_chans = 3
embed_dim = 96
depths = [6, 6, 6, 6]
num_heads = [6, 6, 6, 6]
window_size = 7
mlp_ratio = 4.0
qkv_bias = True
qk_scale = None
drop_rate = 0.0
attn_drop_rate = 0.0
drop_path_rate = 0.1
norm_layer = nn.LayerNorm
ape = False
patch_norm = True
use_checkpoint = False
upscale = 2
img_range = 1.0
upsampler = ""
resi_connection = "1conv"
num_feat = 64
num_in_ch = in_chans
num_out_ch = in_chans
supports_fp16 = True
self.model_arch = "SwinIR"
self.sub_type = "SR"
self.state = state_dict
if "params_ema" in self.state:
self.state = self.state["params_ema"]
elif "params" in self.state:
self.state = self.state["params"]
state_keys = self.state.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 = (
self.state.get("conv_before_upsample.0.weight", None).shape[1]
if self.state.get("conv_before_upsample.weight", None)
else 64
)
num_in_ch = self.state["conv_first.weight"].shape[1]
in_chans = num_in_ch
if "conv_last.weight" in state_keys:
num_out_ch = self.state["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 = self.state[upsample_key].shape[0]
upscale *= math.sqrt(shape // num_feat)
upscale = int(upscale)
elif upsampler == "pixelshuffledirect":
upscale = int(
math.sqrt(self.state["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*).residual_group.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))
depths = [max_block_num + 1 for _ in range(max_layer_num + 1)]
if (
"layers.0.residual_group.blocks.0.attn.relative_position_bias_table"
in state_keys
):
num_heads_num = self.state[
"layers.0.residual_group.blocks.0.attn.relative_position_bias_table"
].shape[-1]
num_heads = [num_heads_num for _ in range(max_layer_num + 1)]
else:
num_heads = depths
embed_dim = self.state["conv_first.weight"].shape[0]
mlp_ratio = float(
self.state["layers.0.residual_group.blocks.0.mlp.fc1.bias"].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"
window_size = int(
math.sqrt(
self.state[
"layers.0.residual_group.blocks.0.attn.relative_position_index"
].shape[0]
)
)
if "layers.0.residual_group.blocks.1.attn_mask" in state_keys:
img_size = int(
math.sqrt(
self.state["layers.0.residual_group.blocks.1.attn_mask"].shape[0]
)
* window_size
)
# The JPEG models are the only ones with window-size 7, and they also use this range
img_range = 255.0 if window_size == 7 else 1.0
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.depths = depths
self.window_size = window_size
self.mlp_ratio = mlp_ratio
self.scale = upscale
self.upsampler = upsampler
self.img_size = img_size
self.img_range = img_range
self.supports_fp16 = False # Too much weirdness to support this at the moment
self.supports_bfp16 = True
self.min_size_restriction = 16
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
self.window_size = window_size
#####################################################################################################
################################### 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(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = embed_dim
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=embed_dim,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None,
)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# merge non-overlapping patches into image
self.patch_unembed = PatchUnEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=embed_dim,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None,
)
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter( # type: ignore
torch.zeros(1, num_patches, embed_dim)
)
trunc_normal_(self.absolute_pos_embed, std=0.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
] # stochastic depth decay rule
# build Residual Swin Transformer blocks (RSTB)
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RSTB(
dim=embed_dim,
input_resolution=(patches_resolution[0], patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[
sum(depths[:i_layer]) : sum(depths[: i_layer + 1]) # type: ignore
], # no impact on SR results
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection,
)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
# 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, high quality image 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,
(patches_resolution[0], patches_resolution[1]),
)
elif self.upsampler == "nearest+conv":
# for real-world SR (less artifacts)
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
if self.upscale == 4:
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
else:
# for image denoising and JPEG compression artifact reduction
self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
self.apply(self._init_weights)
self.load_state_dict(self.state, strict=False)
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.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore # type: ignore
def no_weight_decay(self):
return {"absolute_pos_embed"}
@torch.jit.ignore # type: ignore
def no_weight_decay_keywords(self):
return {"relative_position_bias_table"}
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
return x
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x) # B L C
x = self.patch_unembed(x, x_size)
return x
def forward(self, x):
H, W = x.shape[2:]
x = self.check_image_size(x)
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == "pixelshuffle":
# for classical 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)
elif self.upsampler == "nearest+conv":
# for real-world SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.lrelu(
self.conv_up1(
torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest") # type: ignore
)
)
if self.upscale == 4:
x = self.lrelu(
self.conv_up2(
torch.nn.functional.interpolate( # type: ignore
x, scale_factor=2, mode="nearest"
)
)
)
x = self.conv_last(self.lrelu(self.conv_hr(x)))
else:
# for image denoising and JPEG compression artifact reduction
x_first = self.conv_first(x)
res = self.conv_after_body(self.forward_features(x_first)) + x_first
x = x + self.conv_last(res)
x = x / self.img_range + self.mean
return x[:, :, : H * self.upscale, : W * self.upscale]
def flops(self):
flops = 0
H, W = self.patches_resolution
flops += H * W * 3 * self.embed_dim * 9
flops += self.patch_embed.flops()
for i, layer in enumerate(self.layers):
flops += layer.flops() # type: ignore
flops += H * W * 3 * self.embed_dim * self.embed_dim
flops += self.upsample.flops() # type: ignore
return flops
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment