Skip to content

GitLab

"src/routes/vscode:/vscode.git/clone" did not exist on "2d59847fb971f9079d0b305c2fa1b67d5b76ac2d"
Commit 766c7b38 authored by comfyanonymous's avatar comfyanonymous
Browse files

Update upscale model code to latest Chainner model code. 

Don't add SRFormer because the code license is incompatible with the GPL.

Remove MAT because it's unused and the license is incompatible with GPL.
parent 62efc78a
Show whitespace changes
Inline Side-by-side
Showing with 2084 additions and 2511 deletions
comfy_extras/chainner_models/architecture/DAT.py 0 → 100644
View file @ 766c7b38
# pylint: skip-file
import math
import re
import numpy as np
import torch
import torch.nn as nn
import torch.utils.checkpoint as checkpoint
from einops import rearrange
from einops.layers.torch import Rearrange
from torch import Tensor
from torch.nn import functional as F
from .timm.drop import DropPath
from .timm.weight_init import trunc_normal_
def img2windows(img, H_sp, W_sp):
"""
Input: Image (B, C, H, W)
Output: Window Partition (B', N, C)
"""
B, C, H, W = img.shape
img_reshape = img.view(B, C, H // H_sp, H_sp, W // W_sp, W_sp)
img_perm = (
img_reshape.permute(0, 2, 4, 3, 5, 1).contiguous().reshape(-1, H_sp * W_sp, C)
)
return img_perm
def windows2img(img_splits_hw, H_sp, W_sp, H, W):
"""
Input: Window Partition (B', N, C)
Output: Image (B, H, W, C)
"""
B = int(img_splits_hw.shape[0] / (H * W / H_sp / W_sp))
img = img_splits_hw.view(B, H // H_sp, W // W_sp, H_sp, W_sp, -1)
img = img.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return img
class SpatialGate(nn.Module):
"""Spatial-Gate.
Args:
dim (int): Half of input channels.
"""
def __init__(self, dim):
super().__init__()
self.norm = nn.LayerNorm(dim)
self.conv = nn.Conv2d(
dim, dim, kernel_size=3, stride=1, padding=1, groups=dim
) # DW Conv
def forward(self, x, H, W):
# Split
x1, x2 = x.chunk(2, dim=-1)
B, N, C = x.shape
x2 = (
self.conv(self.norm(x2).transpose(1, 2).contiguous().view(B, C // 2, H, W))
.flatten(2)
.transpose(-1, -2)
.contiguous()
)
return x1 * x2
class SGFN(nn.Module):
"""Spatial-Gate Feed-Forward Network.
Args:
in_features (int): Number of input channels.
hidden_features (int | None): Number of hidden channels. Default: None
out_features (int | None): Number of output channels. Default: None
act_layer (nn.Module): Activation layer. Default: nn.GELU
drop (float): Dropout rate. Default: 0.0
"""
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer=nn.GELU,
drop=0.0,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.sg = SpatialGate(hidden_features // 2)
self.fc2 = nn.Linear(hidden_features // 2, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x, H, W):
"""
Input: x: (B, H*W, C), H, W
Output: x: (B, H*W, C)
"""
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.sg(x, H, W)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class DynamicPosBias(nn.Module):
# The implementation builds on Crossformer code https://github.com/cheerss/CrossFormer/blob/main/models/crossformer.py
"""Dynamic Relative Position Bias.
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads.
residual (bool): If True, use residual strage to connect conv.
"""
def __init__(self, dim, num_heads, residual):
super().__init__()
self.residual = residual
self.num_heads = num_heads
self.pos_dim = dim // 4
self.pos_proj = nn.Linear(2, self.pos_dim)
self.pos1 = nn.Sequential(
nn.LayerNorm(self.pos_dim),
nn.ReLU(inplace=True),
nn.Linear(self.pos_dim, self.pos_dim),
)
self.pos2 = nn.Sequential(
nn.LayerNorm(self.pos_dim),
nn.ReLU(inplace=True),
nn.Linear(self.pos_dim, self.pos_dim),
)
self.pos3 = nn.Sequential(
nn.LayerNorm(self.pos_dim),
nn.ReLU(inplace=True),
nn.Linear(self.pos_dim, self.num_heads),
)
def forward(self, biases):
if self.residual:
pos = self.pos_proj(biases) # 2Gh-1 * 2Gw-1, heads
pos = pos + self.pos1(pos)
pos = pos + self.pos2(pos)
pos = self.pos3(pos)
else:
pos = self.pos3(self.pos2(self.pos1(self.pos_proj(biases))))
return pos
class Spatial_Attention(nn.Module):
"""Spatial Window Self-Attention.
It supports rectangle window (containing square window).
Args:
dim (int): Number of input channels.
idx (int): The indentix of window. (0/1)
split_size (tuple(int)): Height and Width of spatial window.
dim_out (int | None): The dimension of the attention output. Default: None
num_heads (int): Number of attention heads. Default: 6
attn_drop (float): Dropout ratio of attention weight. Default: 0.0
proj_drop (float): Dropout ratio of output. Default: 0.0
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set
position_bias (bool): The dynamic relative position bias. Default: True
"""
def __init__(
self,
dim,
idx,
split_size=[8, 8],
dim_out=None,
num_heads=6,
attn_drop=0.0,
proj_drop=0.0,
qk_scale=None,
position_bias=True,
):
super().__init__()
self.dim = dim
self.dim_out = dim_out or dim
self.split_size = split_size
self.num_heads = num_heads
self.idx = idx
self.position_bias = position_bias
head_dim = dim // num_heads
self.scale = qk_scale or head_dim**-0.5
if idx == 0:
H_sp, W_sp = self.split_size[0], self.split_size[1]
elif idx == 1:
W_sp, H_sp = self.split_size[0], self.split_size[1]
else:
print("ERROR MODE", idx)
exit(0)
self.H_sp = H_sp
self.W_sp = W_sp
if self.position_bias:
self.pos = DynamicPosBias(self.dim // 4, self.num_heads, residual=False)
# generate mother-set
position_bias_h = torch.arange(1 - self.H_sp, self.H_sp)
position_bias_w = torch.arange(1 - self.W_sp, self.W_sp)
biases = torch.stack(torch.meshgrid([position_bias_h, position_bias_w]))
biases = biases.flatten(1).transpose(0, 1).contiguous().float()
self.register_buffer("rpe_biases", biases)
# get pair-wise relative position index for each token inside the window
coords_h = torch.arange(self.H_sp)
coords_w = torch.arange(self.W_sp)
coords = torch.stack(torch.meshgrid([coords_h, coords_w]))
coords_flatten = torch.flatten(coords, 1)
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
relative_coords = relative_coords.permute(1, 2, 0).contiguous()
relative_coords[:, :, 0] += self.H_sp - 1
relative_coords[:, :, 1] += self.W_sp - 1
relative_coords[:, :, 0] *= 2 * self.W_sp - 1
relative_position_index = relative_coords.sum(-1)
self.register_buffer("relative_position_index", relative_position_index)
self.attn_drop = nn.Dropout(attn_drop)
def im2win(self, x, H, W):
B, N, C = x.shape
x = x.transpose(-2, -1).contiguous().view(B, C, H, W)
x = img2windows(x, self.H_sp, self.W_sp)
x = (
x.reshape(-1, self.H_sp * self.W_sp, self.num_heads, C // self.num_heads)
.permute(0, 2, 1, 3)
.contiguous()
)
return x
def forward(self, qkv, H, W, mask=None):
"""
Input: qkv: (B, 3*L, C), H, W, mask: (B, N, N), N is the window size
Output: x (B, H, W, C)
"""
q, k, v = qkv[0], qkv[1], qkv[2]
B, L, C = q.shape
assert L == H * W, "flatten img_tokens has wrong size"
# partition the q,k,v, image to window
q = self.im2win(q, H, W)
k = self.im2win(k, H, W)
v = self.im2win(v, H, W)
q = q * self.scale
attn = q @ k.transpose(-2, -1) # B head N C @ B head C N --> B head N N
# calculate drpe
if self.position_bias:
pos = self.pos(self.rpe_biases)
# select position bias
relative_position_bias = pos[self.relative_position_index.view(-1)].view(
self.H_sp * self.W_sp, self.H_sp * self.W_sp, -1
)
relative_position_bias = relative_position_bias.permute(
2, 0, 1
).contiguous()
attn = attn + relative_position_bias.unsqueeze(0)
N = attn.shape[3]
# use mask for shift window
if mask is not None:
nW = mask.shape[0]
attn = attn.view(B, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(
0
)
attn = attn.view(-1, self.num_heads, N, N)
attn = nn.functional.softmax(attn, dim=-1, dtype=attn.dtype)
attn = self.attn_drop(attn)
x = attn @ v
x = x.transpose(1, 2).reshape(
-1, self.H_sp * self.W_sp, C
) # B head N N @ B head N C
# merge the window, window to image
x = windows2img(x, self.H_sp, self.W_sp, H, W) # B H' W' C
return x
class Adaptive_Spatial_Attention(nn.Module):
# The implementation builds on CAT code https://github.com/Zhengchen1999/CAT
"""Adaptive Spatial Self-Attention
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads. Default: 6
split_size (tuple(int)): Height and Width of spatial window.
shift_size (tuple(int)): Shift size for spatial window.
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set.
drop (float): Dropout rate. Default: 0.0
attn_drop (float): Attention dropout rate. Default: 0.0
rg_idx (int): The indentix of Residual Group (RG)
b_idx (int): The indentix of Block in each RG
"""
def __init__(
self,
dim,
num_heads,
reso=64,
split_size=[8, 8],
shift_size=[1, 2],
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
rg_idx=0,
b_idx=0,
):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.split_size = split_size
self.shift_size = shift_size
self.b_idx = b_idx
self.rg_idx = rg_idx
self.patches_resolution = reso
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
assert (
0 <= self.shift_size[0] < self.split_size[0]
), "shift_size must in 0-split_size0"
assert (
0 <= self.shift_size[1] < self.split_size[1]
), "shift_size must in 0-split_size1"
self.branch_num = 2
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(drop)
self.attns = nn.ModuleList(
[
Spatial_Attention(
dim // 2,
idx=i,
split_size=split_size,
num_heads=num_heads // 2,
dim_out=dim // 2,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
position_bias=True,
)
for i in range(self.branch_num)
]
)
if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or (
self.rg_idx % 2 != 0 and self.b_idx % 4 == 0
):
attn_mask = self.calculate_mask(
self.patches_resolution, self.patches_resolution
)
self.register_buffer("attn_mask_0", attn_mask[0])
self.register_buffer("attn_mask_1", attn_mask[1])
else:
attn_mask = None
self.register_buffer("attn_mask_0", None)
self.register_buffer("attn_mask_1", None)
self.dwconv = nn.Sequential(
nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim),
nn.BatchNorm2d(dim),
nn.GELU(),
)
self.channel_interaction = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(dim, dim // 8, kernel_size=1),
nn.BatchNorm2d(dim // 8),
nn.GELU(),
nn.Conv2d(dim // 8, dim, kernel_size=1),
)
self.spatial_interaction = nn.Sequential(
nn.Conv2d(dim, dim // 16, kernel_size=1),
nn.BatchNorm2d(dim // 16),
nn.GELU(),
nn.Conv2d(dim // 16, 1, kernel_size=1),
)
def calculate_mask(self, H, W):
# The implementation builds on Swin Transformer code https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
# calculate attention mask for shift window
img_mask_0 = torch.zeros((1, H, W, 1)) # 1 H W 1 idx=0
img_mask_1 = torch.zeros((1, H, W, 1)) # 1 H W 1 idx=1
h_slices_0 = (
slice(0, -self.split_size[0]),
slice(-self.split_size[0], -self.shift_size[0]),
slice(-self.shift_size[0], None),
)
w_slices_0 = (
slice(0, -self.split_size[1]),
slice(-self.split_size[1], -self.shift_size[1]),
slice(-self.shift_size[1], None),
)
h_slices_1 = (
slice(0, -self.split_size[1]),
slice(-self.split_size[1], -self.shift_size[1]),
slice(-self.shift_size[1], None),
)
w_slices_1 = (
slice(0, -self.split_size[0]),
slice(-self.split_size[0], -self.shift_size[0]),
slice(-self.shift_size[0], None),
)
cnt = 0
for h in h_slices_0:
for w in w_slices_0:
img_mask_0[:, h, w, :] = cnt
cnt += 1
cnt = 0
for h in h_slices_1:
for w in w_slices_1:
img_mask_1[:, h, w, :] = cnt
cnt += 1
# calculate mask for window-0
img_mask_0 = img_mask_0.view(
1,
H // self.split_size[0],
self.split_size[0],
W // self.split_size[1],
self.split_size[1],
1,
)
img_mask_0 = (
img_mask_0.permute(0, 1, 3, 2, 4, 5)
.contiguous()
.view(-1, self.split_size[0], self.split_size[1], 1)
) # nW, sw[0], sw[1], 1
mask_windows_0 = img_mask_0.view(-1, self.split_size[0] * self.split_size[1])
attn_mask_0 = mask_windows_0.unsqueeze(1) - mask_windows_0.unsqueeze(2)
attn_mask_0 = attn_mask_0.masked_fill(
attn_mask_0 != 0, float(-100.0)
).masked_fill(attn_mask_0 == 0, float(0.0))
# calculate mask for window-1
img_mask_1 = img_mask_1.view(
1,
H // self.split_size[1],
self.split_size[1],
W // self.split_size[0],
self.split_size[0],
1,
)
img_mask_1 = (
img_mask_1.permute(0, 1, 3, 2, 4, 5)
.contiguous()
.view(-1, self.split_size[1], self.split_size[0], 1)
) # nW, sw[1], sw[0], 1
mask_windows_1 = img_mask_1.view(-1, self.split_size[1] * self.split_size[0])
attn_mask_1 = mask_windows_1.unsqueeze(1) - mask_windows_1.unsqueeze(2)
attn_mask_1 = attn_mask_1.masked_fill(
attn_mask_1 != 0, float(-100.0)
).masked_fill(attn_mask_1 == 0, float(0.0))
return attn_mask_0, attn_mask_1
def forward(self, x, H, W):
"""
Input: x: (B, H*W, C), H, W
Output: x: (B, H*W, C)
"""
B, L, C = x.shape
assert L == H * W, "flatten img_tokens has wrong size"
qkv = self.qkv(x).reshape(B, -1, 3, C).permute(2, 0, 1, 3) # 3, B, HW, C
# V without partition
v = qkv[2].transpose(-2, -1).contiguous().view(B, C, H, W)
# image padding
max_split_size = max(self.split_size[0], self.split_size[1])
pad_l = pad_t = 0
pad_r = (max_split_size - W % max_split_size) % max_split_size
pad_b = (max_split_size - H % max_split_size) % max_split_size
qkv = qkv.reshape(3 * B, H, W, C).permute(0, 3, 1, 2) # 3B C H W
qkv = (
F.pad(qkv, (pad_l, pad_r, pad_t, pad_b))
.reshape(3, B, C, -1)
.transpose(-2, -1)
) # l r t b
_H = pad_b + H
_W = pad_r + W
_L = _H * _W
# window-0 and window-1 on split channels [C/2, C/2]; for square windows (e.g., 8x8), window-0 and window-1 can be merged
# shift in block: (0, 4, 8, ...), (2, 6, 10, ...), (0, 4, 8, ...), (2, 6, 10, ...), ...
if (self.rg_idx % 2 == 0 and self.b_idx > 0 and (self.b_idx - 2) % 4 == 0) or (
self.rg_idx % 2 != 0 and self.b_idx % 4 == 0
):
qkv = qkv.view(3, B, _H, _W, C)
qkv_0 = torch.roll(
qkv[:, :, :, :, : C // 2],
shifts=(-self.shift_size[0], -self.shift_size[1]),
dims=(2, 3),
)
qkv_0 = qkv_0.view(3, B, _L, C // 2)
qkv_1 = torch.roll(
qkv[:, :, :, :, C // 2 :],
shifts=(-self.shift_size[1], -self.shift_size[0]),
dims=(2, 3),
)
qkv_1 = qkv_1.view(3, B, _L, C // 2)
if self.patches_resolution != _H or self.patches_resolution != _W:
mask_tmp = self.calculate_mask(_H, _W)
x1_shift = self.attns[0](qkv_0, _H, _W, mask=mask_tmp[0].to(x.device))
x2_shift = self.attns[1](qkv_1, _H, _W, mask=mask_tmp[1].to(x.device))
else:
x1_shift = self.attns[0](qkv_0, _H, _W, mask=self.attn_mask_0)
x2_shift = self.attns[1](qkv_1, _H, _W, mask=self.attn_mask_1)
x1 = torch.roll(
x1_shift, shifts=(self.shift_size[0], self.shift_size[1]), dims=(1, 2)
)
x2 = torch.roll(
x2_shift, shifts=(self.shift_size[1], self.shift_size[0]), dims=(1, 2)
)
x1 = x1[:, :H, :W, :].reshape(B, L, C // 2)
x2 = x2[:, :H, :W, :].reshape(B, L, C // 2)
# attention output
attened_x = torch.cat([x1, x2], dim=2)
else:
x1 = self.attns[0](qkv[:, :, :, : C // 2], _H, _W)[:, :H, :W, :].reshape(
B, L, C // 2
)
x2 = self.attns[1](qkv[:, :, :, C // 2 :], _H, _W)[:, :H, :W, :].reshape(
B, L, C // 2
)
# attention output
attened_x = torch.cat([x1, x2], dim=2)
# convolution output
conv_x = self.dwconv(v)
# Adaptive Interaction Module (AIM)
# C-Map (before sigmoid)
channel_map = (
self.channel_interaction(conv_x)
.permute(0, 2, 3, 1)
.contiguous()
.view(B, 1, C)
)
# S-Map (before sigmoid)
attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W)
spatial_map = self.spatial_interaction(attention_reshape)
# C-I
attened_x = attened_x * torch.sigmoid(channel_map)
# S-I
conv_x = torch.sigmoid(spatial_map) * conv_x
conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, L, C)
x = attened_x + conv_x
x = self.proj(x)
x = self.proj_drop(x)
return x
class Adaptive_Channel_Attention(nn.Module):
# The implementation builds on XCiT code https://github.com/facebookresearch/xcit
"""Adaptive Channel Self-Attention
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads. Default: 6
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set.
attn_drop (float): Attention dropout rate. Default: 0.0
drop_path (float): Stochastic depth rate. Default: 0.0
"""
def __init__(
self,
dim,
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
):
super().__init__()
self.num_heads = num_heads
self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1))
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.dwconv = nn.Sequential(
nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, groups=dim),
nn.BatchNorm2d(dim),
nn.GELU(),
)
self.channel_interaction = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Conv2d(dim, dim // 8, kernel_size=1),
nn.BatchNorm2d(dim // 8),
nn.GELU(),
nn.Conv2d(dim // 8, dim, kernel_size=1),
)
self.spatial_interaction = nn.Sequential(
nn.Conv2d(dim, dim // 16, kernel_size=1),
nn.BatchNorm2d(dim // 16),
nn.GELU(),
nn.Conv2d(dim // 16, 1, kernel_size=1),
)
def forward(self, x, H, W):
"""
Input: x: (B, H*W, C), H, W
Output: x: (B, H*W, C)
"""
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
qkv = qkv.permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
q = q.transpose(-2, -1)
k = k.transpose(-2, -1)
v = v.transpose(-2, -1)
v_ = v.reshape(B, C, N).contiguous().view(B, C, H, W)
q = torch.nn.functional.normalize(q, dim=-1)
k = torch.nn.functional.normalize(k, dim=-1)
attn = (q @ k.transpose(-2, -1)) * self.temperature
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
# attention output
attened_x = (attn @ v).permute(0, 3, 1, 2).reshape(B, N, C)
# convolution output
conv_x = self.dwconv(v_)
# Adaptive Interaction Module (AIM)
# C-Map (before sigmoid)
attention_reshape = attened_x.transpose(-2, -1).contiguous().view(B, C, H, W)
channel_map = self.channel_interaction(attention_reshape)
# S-Map (before sigmoid)
spatial_map = (
self.spatial_interaction(conv_x)
.permute(0, 2, 3, 1)
.contiguous()
.view(B, N, 1)
)
# S-I
attened_x = attened_x * torch.sigmoid(spatial_map)
# C-I
conv_x = conv_x * torch.sigmoid(channel_map)
conv_x = conv_x.permute(0, 2, 3, 1).contiguous().view(B, N, C)
x = attened_x + conv_x
x = self.proj(x)
x = self.proj_drop(x)
return x
class DATB(nn.Module):
def __init__(
self,
dim,
num_heads,
reso=64,
split_size=[2, 4],
shift_size=[1, 2],
expansion_factor=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
rg_idx=0,
b_idx=0,
):
super().__init__()
self.norm1 = norm_layer(dim)
if b_idx % 2 == 0:
# DSTB
self.attn = Adaptive_Spatial_Attention(
dim,
num_heads=num_heads,
reso=reso,
split_size=split_size,
shift_size=shift_size,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
rg_idx=rg_idx,
b_idx=b_idx,
)
else:
# DCTB
self.attn = Adaptive_Channel_Attention(
dim,
num_heads=num_heads,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
ffn_hidden_dim = int(dim * expansion_factor)
self.ffn = SGFN(
in_features=dim,
hidden_features=ffn_hidden_dim,
out_features=dim,
act_layer=act_layer,
)
self.norm2 = norm_layer(dim)
def forward(self, x, x_size):
"""
Input: x: (B, H*W, C), x_size: (H, W)
Output: x: (B, H*W, C)
"""
H, W = x_size
x = x + self.drop_path(self.attn(self.norm1(x), H, W))
x = x + self.drop_path(self.ffn(self.norm2(x), H, W))
return x
class ResidualGroup(nn.Module):
"""ResidualGroup
Args:
dim (int): Number of input channels.
reso (int): Input resolution.
num_heads (int): Number of attention heads.
split_size (tuple(int)): Height and Width of spatial window.
expansion_factor (float): Ratio of ffn hidden dim to embedding dim.
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None
drop (float): Dropout rate. Default: 0
attn_drop(float): Attention dropout rate. Default: 0
drop_paths (float | None): Stochastic depth rate.
act_layer (nn.Module): Activation layer. Default: nn.GELU
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm
depth (int): Number of dual aggregation Transformer blocks in residual group.
use_chk (bool): Whether to use checkpointing to save memory.
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(
self,
dim,
reso,
num_heads,
split_size=[2, 4],
expansion_factor=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_paths=None,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
depth=2,
use_chk=False,
resi_connection="1conv",
rg_idx=0,
):
super().__init__()
self.use_chk = use_chk
self.reso = reso
self.blocks = nn.ModuleList(
[
DATB(
dim=dim,
num_heads=num_heads,
reso=reso,
split_size=split_size,
shift_size=[split_size[0] // 2, split_size[1] // 2],
expansion_factor=expansion_factor,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_paths[i],
act_layer=act_layer,
norm_layer=norm_layer,
rg_idx=rg_idx,
b_idx=i,
)
for i in range(depth)
]
)
if resi_connection == "1conv":
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == "3conv":
self.conv = nn.Sequential(
nn.Conv2d(dim, dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim, 3, 1, 1),
)
def forward(self, x, x_size):
"""
Input: x: (B, H*W, C), x_size: (H, W)
Output: x: (B, H*W, C)
"""
H, W = x_size
res = x
for blk in self.blocks:
if self.use_chk:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W)
x = self.conv(x)
x = rearrange(x, "b c h w -> b (h w) c")
x = res + x
return x
class Upsample(nn.Sequential):
"""Upsample module.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat):
m = []
if (scale & (scale - 1)) == 0: # scale = 2^n
for _ in range(int(math.log(scale, 2))):
m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(2))
elif scale == 3:
m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(3))
else:
raise ValueError(
f"scale {scale} is not supported. " "Supported scales: 2^n and 3."
)
super(Upsample, self).__init__(*m)
class UpsampleOneStep(nn.Sequential):
"""UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
Used in lightweight SR to save parameters.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
self.num_feat = num_feat
self.input_resolution = input_resolution
m = []
m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1))
m.append(nn.PixelShuffle(scale))
super(UpsampleOneStep, self).__init__(*m)
def flops(self):
h, w = self.input_resolution
flops = h * w * self.num_feat * 3 * 9
return flops
class DAT(nn.Module):
"""Dual Aggregation Transformer
Args:
img_size (int): Input image size. Default: 64
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 180
depths (tuple(int)): Depth of each residual group (number of DATB in each RG).
split_size (tuple(int)): Height and Width of spatial window.
num_heads (tuple(int)): Number of attention heads in different residual groups.
expansion_factor (float): Ratio of ffn hidden dim to embedding dim. Default: 4
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None): Override default qk scale of head_dim ** -0.5 if set. Default: None
drop_rate (float): Dropout rate. Default: 0
attn_drop_rate (float): Attention dropout rate. Default: 0
drop_path_rate (float): Stochastic depth rate. Default: 0.1
act_layer (nn.Module): Activation layer. Default: nn.GELU
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm
use_chk (bool): Whether to use checkpointing to save memory.
upscale: Upscale factor. 2/3/4 for image SR
img_range: Image range. 1. or 255.
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(self, state_dict):
super().__init__()
# defaults
img_size = 64
in_chans = 3
embed_dim = 180
split_size = [2, 4]
depth = [2, 2, 2, 2]
num_heads = [2, 2, 2, 2]
expansion_factor = 4.0
qkv_bias = True
qk_scale = None
drop_rate = 0.0
attn_drop_rate = 0.0
drop_path_rate = 0.1
act_layer = nn.GELU
norm_layer = nn.LayerNorm
use_chk = False
upscale = 2
img_range = 1.0
resi_connection = "1conv"
upsampler = "pixelshuffle"
self.model_arch = "DAT"
self.sub_type = "SR"
self.state = state_dict
state_keys = state_dict.keys()
if "conv_before_upsample.0.weight" in state_keys:
if "conv_up1.weight" in state_keys:
upsampler = "nearest+conv"
else:
upsampler = "pixelshuffle"
supports_fp16 = False
elif "upsample.0.weight" in state_keys:
upsampler = "pixelshuffledirect"
else:
upsampler = ""
num_feat = (
state_dict.get("conv_before_upsample.0.weight", None).shape[1]
if state_dict.get("conv_before_upsample.weight", None)
else 64
)
num_in_ch = state_dict["conv_first.weight"].shape[1]
in_chans = num_in_ch
if "conv_last.weight" in state_keys:
num_out_ch = state_dict["conv_last.weight"].shape[0]
else:
num_out_ch = num_in_ch
upscale = 1
if upsampler == "nearest+conv":
upsample_keys = [
x for x in state_keys if "conv_up" in x and "bias" not in x
]
for upsample_key in upsample_keys:
upscale *= 2
elif upsampler == "pixelshuffle":
upsample_keys = [
x
for x in state_keys
if "upsample" in x and "conv" not in x and "bias" not in x
]
for upsample_key in upsample_keys:
shape = state_dict[upsample_key].shape[0]
upscale *= math.sqrt(shape // num_feat)
upscale = int(upscale)
elif upsampler == "pixelshuffledirect":
upscale = int(
math.sqrt(state_dict["upsample.0.bias"].shape[0] // num_out_ch)
)
max_layer_num = 0
max_block_num = 0
for key in state_keys:
result = re.match(r"layers.(\d*).blocks.(\d*).norm1.weight", key)
if result:
layer_num, block_num = result.groups()
max_layer_num = max(max_layer_num, int(layer_num))
max_block_num = max(max_block_num, int(block_num))
depth = [max_block_num + 1 for _ in range(max_layer_num + 1)]
if "layers.0.blocks.1.attn.temperature" in state_keys:
num_heads_num = state_dict["layers.0.blocks.1.attn.temperature"].shape[0]
num_heads = [num_heads_num for _ in range(max_layer_num + 1)]
else:
num_heads = depth
embed_dim = state_dict["conv_first.weight"].shape[0]
expansion_factor = float(
state_dict["layers.0.blocks.0.ffn.fc1.weight"].shape[0] / embed_dim
)
# TODO: could actually count the layers, but this should do
if "layers.0.conv.4.weight" in state_keys:
resi_connection = "3conv"
else:
resi_connection = "1conv"
if "layers.0.blocks.2.attn.attn_mask_0" in state_keys:
attn_mask_0_x, attn_mask_0_y, attn_mask_0_z = state_dict[
"layers.0.blocks.2.attn.attn_mask_0"
].shape
img_size = int(math.sqrt(attn_mask_0_x * attn_mask_0_y))
if "layers.0.blocks.0.attn.attns.0.rpe_biases" in state_keys:
split_sizes = (
state_dict["layers.0.blocks.0.attn.attns.0.rpe_biases"][-1] + 1
)
split_size = [int(x) for x in split_sizes]
self.in_nc = num_in_ch
self.out_nc = num_out_ch
self.num_feat = num_feat
self.embed_dim = embed_dim
self.num_heads = num_heads
self.depth = depth
self.scale = upscale
self.upsampler = upsampler
self.img_size = img_size
self.img_range = img_range
self.expansion_factor = expansion_factor
self.resi_connection = resi_connection
self.split_size = split_size
self.supports_fp16 = False # Too much weirdness to support this at the moment
self.supports_bfp16 = True
self.min_size_restriction = 16
num_in_ch = in_chans
num_out_ch = in_chans
num_feat = 64
self.img_range = img_range
if in_chans == 3:
rgb_mean = (0.4488, 0.4371, 0.4040)
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
else:
self.mean = torch.zeros(1, 1, 1, 1)
self.upscale = upscale
self.upsampler = upsampler
# ------------------------- 1, Shallow Feature Extraction ------------------------- #
self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
# ------------------------- 2, Deep Feature Extraction ------------------------- #
self.num_layers = len(depth)
self.use_chk = use_chk
self.num_features = (
self.embed_dim
) = embed_dim # num_features for consistency with other models
heads = num_heads
self.before_RG = nn.Sequential(
Rearrange("b c h w -> b (h w) c"), nn.LayerNorm(embed_dim)
)
curr_dim = embed_dim
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, np.sum(depth))
] # stochastic depth decay rule
self.layers = nn.ModuleList()
for i in range(self.num_layers):
layer = ResidualGroup(
dim=embed_dim,
num_heads=heads[i],
reso=img_size,
split_size=split_size,
expansion_factor=expansion_factor,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_paths=dpr[sum(depth[:i]) : sum(depth[: i + 1])],
act_layer=act_layer,
norm_layer=norm_layer,
depth=depth[i],
use_chk=use_chk,
resi_connection=resi_connection,
rg_idx=i,
)
self.layers.append(layer)
self.norm = norm_layer(curr_dim)
# build the last conv layer in deep feature extraction
if resi_connection == "1conv":
self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
elif resi_connection == "3conv":
# to save parameters and memory
self.conv_after_body = nn.Sequential(
nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1),
)
# ------------------------- 3, Reconstruction ------------------------- #
if self.upsampler == "pixelshuffle":
# for classical SR
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.upsample = Upsample(upscale, num_feat)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
elif self.upsampler == "pixelshuffledirect":
# for lightweight SR (to save parameters)
self.upsample = UpsampleOneStep(
upscale, embed_dim, num_out_ch, (img_size, img_size)
)
self.apply(self._init_weights)
self.load_state_dict(state_dict, strict=True)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(
m, (nn.LayerNorm, nn.BatchNorm2d, nn.GroupNorm, nn.InstanceNorm2d)
):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward_features(self, x):
_, _, H, W = x.shape
x_size = [H, W]
x = self.before_RG(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x)
x = rearrange(x, "b (h w) c -> b c h w", h=H, w=W)
return x
def forward(self, x):
"""
Input: x: (B, C, H, W)
"""
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == "pixelshuffle":
# for image SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.conv_last(self.upsample(x))
elif self.upsampler == "pixelshuffledirect":
# for lightweight SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.upsample(x)
x = x / self.img_range + self.mean
return x
comfy_extras/chainner_models/architecture/LICENSE-DAT 0 → 100644
View file @ 766c7b38
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-SCUNet 0 → 100644
View file @ 766c7b38
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 2022 Kai Zhang (cskaizhang@gmail.com, https://cszn.github.io/). All rights reserved.
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 deleted 100644 → 0
View file @ 62efc78a
## 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/MAT.py deleted 100644 → 0
View file @ 62efc78a
# 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/OmniSR/OmniSR.py
View file @ 766c7b38
...@@ -56,7 +56,17 @@ class OmniSR(nn.Module): ...@@ -56,7 +56,17 @@ class OmniSR(nn.Module):
residual_layer = [] residual_layer = []
self.res_num = res_num self.res_num = res_num
self.window_size = 8 # we can just assume this for now, but there's probably a way to calculate it (just need to get the sqrt of the right layer) if (
"residual_layer.0.residual_layer.0.layer.2.fn.rel_pos_bias.weight"
in state_dict.keys()
):
rel_pos_bias_weight = state_dict[
"residual_layer.0.residual_layer.0.layer.2.fn.rel_pos_bias.weight"
].shape[0]
self.window_size = int((math.sqrt(rel_pos_bias_weight) + 1) / 2)
else:
self.window_size = 8
self.up_scale = up_scale self.up_scale = up_scale
for _ in range(res_num): for _ in range(res_num):
......
comfy_extras/chainner_models/architecture/SCUNet.py 0 → 100644
View file @ 766c7b38
# pylint: skip-file
# -----------------------------------------------------------------------------------
# SCUNet: Practical Blind Denoising via Swin-Conv-UNet and Data Synthesis, https://arxiv.org/abs/2203.13278
# Zhang, Kai and Li, Yawei and Liang, Jingyun and Cao, Jiezhang and Zhang, Yulun and Tang, Hao and Timofte, Radu and Van Gool, Luc
# -----------------------------------------------------------------------------------
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from einops.layers.torch import Rearrange
from .timm.drop import DropPath
from .timm.weight_init import trunc_normal_
# Borrowed from https://github.com/cszn/SCUNet/blob/main/models/network_scunet.py
class WMSA(nn.Module):
"""Self-attention module in Swin Transformer"""
def __init__(self, input_dim, output_dim, head_dim, window_size, type):
super(WMSA, self).__init__()
self.input_dim = input_dim
self.output_dim = output_dim
self.head_dim = head_dim
self.scale = self.head_dim**-0.5
self.n_heads = input_dim // head_dim
self.window_size = window_size
self.type = type
self.embedding_layer = nn.Linear(self.input_dim, 3 * self.input_dim, bias=True)
self.relative_position_params = nn.Parameter(
torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads)
)
# TODO recover
# self.relative_position_params = nn.Parameter(torch.zeros(self.n_heads, 2 * window_size - 1, 2 * window_size -1))
self.relative_position_params = nn.Parameter(
torch.zeros((2 * window_size - 1) * (2 * window_size - 1), self.n_heads)
)
self.linear = nn.Linear(self.input_dim, self.output_dim)
trunc_normal_(self.relative_position_params, std=0.02)
self.relative_position_params = torch.nn.Parameter(
self.relative_position_params.view(
2 * window_size - 1, 2 * window_size - 1, self.n_heads
)
.transpose(1, 2)
.transpose(0, 1)
)
def generate_mask(self, h, w, p, shift):
"""generating the mask of SW-MSA
Args:
shift: shift parameters in CyclicShift.
Returns:
attn_mask: should be (1 1 w p p),
"""
# supporting square.
attn_mask = torch.zeros(
h,
w,
p,
p,
p,
p,
dtype=torch.bool,
device=self.relative_position_params.device,
)
if self.type == "W":
return attn_mask
s = p - shift
attn_mask[-1, :, :s, :, s:, :] = True
attn_mask[-1, :, s:, :, :s, :] = True
attn_mask[:, -1, :, :s, :, s:] = True
attn_mask[:, -1, :, s:, :, :s] = True
attn_mask = rearrange(
attn_mask, "w1 w2 p1 p2 p3 p4 -> 1 1 (w1 w2) (p1 p2) (p3 p4)"
)
return attn_mask
def forward(self, x):
"""Forward pass of Window Multi-head Self-attention module.
Args:
x: input tensor with shape of [b h w c];
attn_mask: attention mask, fill -inf where the value is True;
Returns:
output: tensor shape [b h w c]
"""
if self.type != "W":
x = torch.roll(
x,
shifts=(-(self.window_size // 2), -(self.window_size // 2)),
dims=(1, 2),
)
x = rearrange(
x,
"b (w1 p1) (w2 p2) c -> b w1 w2 p1 p2 c",
p1=self.window_size,
p2=self.window_size,
)
h_windows = x.size(1)
w_windows = x.size(2)
# square validation
# assert h_windows == w_windows
x = rearrange(
x,
"b w1 w2 p1 p2 c -> b (w1 w2) (p1 p2) c",
p1=self.window_size,
p2=self.window_size,
)
qkv = self.embedding_layer(x)
q, k, v = rearrange(
qkv, "b nw np (threeh c) -> threeh b nw np c", c=self.head_dim
).chunk(3, dim=0)
sim = torch.einsum("hbwpc,hbwqc->hbwpq", q, k) * self.scale
# Adding learnable relative embedding
sim = sim + rearrange(self.relative_embedding(), "h p q -> h 1 1 p q")
# Using Attn Mask to distinguish different subwindows.
if self.type != "W":
attn_mask = self.generate_mask(
h_windows, w_windows, self.window_size, shift=self.window_size // 2
)
sim = sim.masked_fill_(attn_mask, float("-inf"))
probs = nn.functional.softmax(sim, dim=-1)
output = torch.einsum("hbwij,hbwjc->hbwic", probs, v)
output = rearrange(output, "h b w p c -> b w p (h c)")
output = self.linear(output)
output = rearrange(
output,
"b (w1 w2) (p1 p2) c -> b (w1 p1) (w2 p2) c",
w1=h_windows,
p1=self.window_size,
)
if self.type != "W":
output = torch.roll(
output,
shifts=(self.window_size // 2, self.window_size // 2),
dims=(1, 2),
)
return output
def relative_embedding(self):
cord = torch.tensor(
np.array(
[
[i, j]
for i in range(self.window_size)
for j in range(self.window_size)
]
)
)
relation = cord[:, None, :] - cord[None, :, :] + self.window_size - 1
# negative is allowed
return self.relative_position_params[
:, relation[:, :, 0].long(), relation[:, :, 1].long()
]
class Block(nn.Module):
def __init__(
self,
input_dim,
output_dim,
head_dim,
window_size,
drop_path,
type="W",
input_resolution=None,
):
"""SwinTransformer Block"""
super(Block, self).__init__()
self.input_dim = input_dim
self.output_dim = output_dim
assert type in ["W", "SW"]
self.type = type
if input_resolution <= window_size:
self.type = "W"
self.ln1 = nn.LayerNorm(input_dim)
self.msa = WMSA(input_dim, input_dim, head_dim, window_size, self.type)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.ln2 = nn.LayerNorm(input_dim)
self.mlp = nn.Sequential(
nn.Linear(input_dim, 4 * input_dim),
nn.GELU(),
nn.Linear(4 * input_dim, output_dim),
)
def forward(self, x):
x = x + self.drop_path(self.msa(self.ln1(x)))
x = x + self.drop_path(self.mlp(self.ln2(x)))
return x
class ConvTransBlock(nn.Module):
def __init__(
self,
conv_dim,
trans_dim,
head_dim,
window_size,
drop_path,
type="W",
input_resolution=None,
):
"""SwinTransformer and Conv Block"""
super(ConvTransBlock, self).__init__()
self.conv_dim = conv_dim
self.trans_dim = trans_dim
self.head_dim = head_dim
self.window_size = window_size
self.drop_path = drop_path
self.type = type
self.input_resolution = input_resolution
assert self.type in ["W", "SW"]
if self.input_resolution <= self.window_size:
self.type = "W"
self.trans_block = Block(
self.trans_dim,
self.trans_dim,
self.head_dim,
self.window_size,
self.drop_path,
self.type,
self.input_resolution,
)
self.conv1_1 = nn.Conv2d(
self.conv_dim + self.trans_dim,
self.conv_dim + self.trans_dim,
1,
1,
0,
bias=True,
)
self.conv1_2 = nn.Conv2d(
self.conv_dim + self.trans_dim,
self.conv_dim + self.trans_dim,
1,
1,
0,
bias=True,
)
self.conv_block = nn.Sequential(
nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False),
nn.ReLU(True),
nn.Conv2d(self.conv_dim, self.conv_dim, 3, 1, 1, bias=False),
)
def forward(self, x):
conv_x, trans_x = torch.split(
self.conv1_1(x), (self.conv_dim, self.trans_dim), dim=1
)
conv_x = self.conv_block(conv_x) + conv_x
trans_x = Rearrange("b c h w -> b h w c")(trans_x)
trans_x = self.trans_block(trans_x)
trans_x = Rearrange("b h w c -> b c h w")(trans_x)
res = self.conv1_2(torch.cat((conv_x, trans_x), dim=1))
x = x + res
return x
class SCUNet(nn.Module):
def __init__(
self,
state_dict,
in_nc=3,
config=[4, 4, 4, 4, 4, 4, 4],
dim=64,
drop_path_rate=0.0,
input_resolution=256,
):
super(SCUNet, self).__init__()
self.model_arch = "SCUNet"
self.sub_type = "SR"
self.num_filters: int = 0
self.state = state_dict
self.config = config
self.dim = dim
self.head_dim = 32
self.window_size = 8
self.in_nc = in_nc
self.out_nc = self.in_nc
self.scale = 1
self.supports_fp16 = True
# drop path rate for each layer
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(config))]
self.m_head = [nn.Conv2d(in_nc, dim, 3, 1, 1, bias=False)]
begin = 0
self.m_down1 = [
ConvTransBlock(
dim // 2,
dim // 2,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution,
)
for i in range(config[0])
] + [nn.Conv2d(dim, 2 * dim, 2, 2, 0, bias=False)]
begin += config[0]
self.m_down2 = [
ConvTransBlock(
dim,
dim,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution // 2,
)
for i in range(config[1])
] + [nn.Conv2d(2 * dim, 4 * dim, 2, 2, 0, bias=False)]
begin += config[1]
self.m_down3 = [
ConvTransBlock(
2 * dim,
2 * dim,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution // 4,
)
for i in range(config[2])
] + [nn.Conv2d(4 * dim, 8 * dim, 2, 2, 0, bias=False)]
begin += config[2]
self.m_body = [
ConvTransBlock(
4 * dim,
4 * dim,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution // 8,
)
for i in range(config[3])
]
begin += config[3]
self.m_up3 = [
nn.ConvTranspose2d(8 * dim, 4 * dim, 2, 2, 0, bias=False),
] + [
ConvTransBlock(
2 * dim,
2 * dim,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution // 4,
)
for i in range(config[4])
]
begin += config[4]
self.m_up2 = [
nn.ConvTranspose2d(4 * dim, 2 * dim, 2, 2, 0, bias=False),
] + [
ConvTransBlock(
dim,
dim,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution // 2,
)
for i in range(config[5])
]
begin += config[5]
self.m_up1 = [
nn.ConvTranspose2d(2 * dim, dim, 2, 2, 0, bias=False),
] + [
ConvTransBlock(
dim // 2,
dim // 2,
self.head_dim,
self.window_size,
dpr[i + begin],
"W" if not i % 2 else "SW",
input_resolution,
)
for i in range(config[6])
]
self.m_tail = [nn.Conv2d(dim, in_nc, 3, 1, 1, bias=False)]
self.m_head = nn.Sequential(*self.m_head)
self.m_down1 = nn.Sequential(*self.m_down1)
self.m_down2 = nn.Sequential(*self.m_down2)
self.m_down3 = nn.Sequential(*self.m_down3)
self.m_body = nn.Sequential(*self.m_body)
self.m_up3 = nn.Sequential(*self.m_up3)
self.m_up2 = nn.Sequential(*self.m_up2)
self.m_up1 = nn.Sequential(*self.m_up1)
self.m_tail = nn.Sequential(*self.m_tail)
# self.apply(self._init_weights)
self.load_state_dict(state_dict, strict=True)
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (64 - h % 64) % 64
mod_pad_w = (64 - w % 64) % 64
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
return x
def forward(self, x0):
h, w = x0.size()[-2:]
x0 = self.check_image_size(x0)
x1 = self.m_head(x0)
x2 = self.m_down1(x1)
x3 = self.m_down2(x2)
x4 = self.m_down3(x3)
x = self.m_body(x4)
x = self.m_up3(x + x4)
x = self.m_up2(x + x3)
x = self.m_up1(x + x2)
x = self.m_tail(x + x1)
x = x[:, :, :h, :w]
return x
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=0.02)
if 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)
comfy_extras/chainner_models/architecture/SPSR.py
View file @ 766c7b38
...@@ -60,7 +60,6 @@ class SPSRNet(nn.Module): ...@@ -60,7 +60,6 @@ class SPSRNet(nn.Module):
self.out_nc: int = self.state["f_HR_conv1.0.bias"].shape[0] self.out_nc: int = self.state["f_HR_conv1.0.bias"].shape[0]
self.scale = self.get_scale(4) self.scale = self.get_scale(4)
print(self.scale)
self.num_filters: int = self.state["model.0.weight"].shape[0] self.num_filters: int = self.state["model.0.weight"].shape[0]
self.supports_fp16 = True self.supports_fp16 = True
......
comfy_extras/chainner_models/architecture/SwinIR.py
View file @ 766c7b38
...@@ -972,6 +972,7 @@ class SwinIR(nn.Module): ...@@ -972,6 +972,7 @@ class SwinIR(nn.Module):
self.upsampler = upsampler self.upsampler = upsampler
self.img_size = img_size self.img_size = img_size
self.img_range = img_range 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_fp16 = False # Too much weirdness to support this at the moment
self.supports_bfp16 = True self.supports_bfp16 = True
......
comfy_extras/chainner_models/architecture/mat/utils.py deleted 100644 → 0
View file @ 62efc78a
"""Code used for this implementation of the MAT helper utils is modified from
lama-cleaner, copyright of Sanster: https://github.com/fenglinglwb/MAT"""
import collections
from itertools import repeat
from typing import Any
import numpy as np
import torch
from torch import conv2d, conv_transpose2d
def normalize_2nd_moment(x, dim=1, eps=1e-8):
return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
class EasyDict(dict):
"""Convenience class that behaves like a dict but allows access with the attribute syntax."""
def __getattr__(self, name: str) -> Any:
try:
return self[name]
except KeyError:
raise AttributeError(name)
def __setattr__(self, name: str, value: Any) -> None:
self[name] = value
def __delattr__(self, name: str) -> None:
del self[name]
activation_funcs = {
"linear": EasyDict(
func=lambda x, **_: x,
def_alpha=0,
def_gain=1,
cuda_idx=1,
ref="",
has_2nd_grad=False,
),
"relu": EasyDict(
func=lambda x, **_: torch.nn.functional.relu(x),
def_alpha=0,
def_gain=np.sqrt(2),
cuda_idx=2,
ref="y",
has_2nd_grad=False,
),
"lrelu": EasyDict(
func=lambda x, alpha, **_: torch.nn.functional.leaky_relu(x, alpha),
def_alpha=0.2,
def_gain=np.sqrt(2),
cuda_idx=3,
ref="y",
has_2nd_grad=False,
),
"tanh": EasyDict(
func=lambda x, **_: torch.tanh(x),
def_alpha=0,
def_gain=1,
cuda_idx=4,
ref="y",
has_2nd_grad=True,
),
"sigmoid": EasyDict(
func=lambda x, **_: torch.sigmoid(x),
def_alpha=0,
def_gain=1,
cuda_idx=5,
ref="y",
has_2nd_grad=True,
),
"elu": EasyDict(
func=lambda x, **_: torch.nn.functional.elu(x),
def_alpha=0,
def_gain=1,
cuda_idx=6,
ref="y",
has_2nd_grad=True,
),
"selu": EasyDict(
func=lambda x, **_: torch.nn.functional.selu(x),
def_alpha=0,
def_gain=1,
cuda_idx=7,
ref="y",
has_2nd_grad=True,
),
"softplus": EasyDict(
func=lambda x, **_: torch.nn.functional.softplus(x),
def_alpha=0,
def_gain=1,
cuda_idx=8,
ref="y",
has_2nd_grad=True,
),
"swish": EasyDict(
func=lambda x, **_: torch.sigmoid(x) * x,
def_alpha=0,
def_gain=np.sqrt(2),
cuda_idx=9,
ref="x",
has_2nd_grad=True,
),
}
def _bias_act_ref(x, b=None, dim=1, act="linear", alpha=None, gain=None, clamp=None):
"""Slow reference implementation of `bias_act()` using standard TensorFlow ops."""
assert isinstance(x, torch.Tensor)
assert clamp is None or clamp >= 0
spec = activation_funcs[act]
alpha = float(alpha if alpha is not None else spec.def_alpha)
gain = float(gain if gain is not None else spec.def_gain)
clamp = float(clamp if clamp is not None else -1)
# Add bias.
if b is not None:
assert isinstance(b, torch.Tensor) and b.ndim == 1
assert 0 <= dim < x.ndim
assert b.shape[0] == x.shape[dim]
x = x + b.reshape([-1 if i == dim else 1 for i in range(x.ndim)]).to(x.device)
# Evaluate activation function.
alpha = float(alpha)
x = spec.func(x, alpha=alpha)
# Scale by gain.
gain = float(gain)
if gain != 1:
x = x * gain
# Clamp.
if clamp >= 0:
x = x.clamp(-clamp, clamp) # pylint: disable=invalid-unary-operand-type
return x
def bias_act(
x, b=None, dim=1, act="linear", alpha=None, gain=None, clamp=None, impl="ref"
):
r"""Fused bias and activation function.
Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
and scales the result by `gain`. Each of the steps is optional. In most cases,
the fused op is considerably more efficient than performing the same calculation
using standard PyTorch ops. It supports first and second order gradients,
but not third order gradients.
Args:
x: Input activation tensor. Can be of any shape.
b: Bias vector, or `None` to disable. Must be a 1D tensor of the same type
as `x`. The shape must be known, and it must match the dimension of `x`
corresponding to `dim`.
dim: The dimension in `x` corresponding to the elements of `b`.
The value of `dim` is ignored if `b` is not specified.
act: Name of the activation function to evaluate, or `"linear"` to disable.
Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
See `activation_funcs` for a full list. `None` is not allowed.
alpha: Shape parameter for the activation function, or `None` to use the default.
gain: Scaling factor for the output tensor, or `None` to use default.
See `activation_funcs` for the default scaling of each activation function.
If unsure, consider specifying 1.
clamp: Clamp the output values to `[-clamp, +clamp]`, or `None` to disable
the clamping (default).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the same shape and datatype as `x`.
"""
assert isinstance(x, torch.Tensor)
assert impl in ["ref", "cuda"]
return _bias_act_ref(
x=x, b=b, dim=dim, act=act, alpha=alpha, gain=gain, clamp=clamp
)
def setup_filter(
f,
device=torch.device("cpu"),
normalize=True,
flip_filter=False,
gain=1,
separable=None,
):
r"""Convenience function to setup 2D FIR filter for `upfirdn2d()`.
Args:
f: Torch tensor, numpy array, or python list of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable),
`[]` (impulse), or
`None` (identity).
device: Result device (default: cpu).
normalize: Normalize the filter so that it retains the magnitude
for constant input signal (DC)? (default: True).
flip_filter: Flip the filter? (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
separable: Return a separable filter? (default: select automatically).
Returns:
Float32 tensor of the shape
`[filter_height, filter_width]` (non-separable) or
`[filter_taps]` (separable).
"""
# Validate.
if f is None:
f = 1
f = torch.as_tensor(f, dtype=torch.float32)
assert f.ndim in [0, 1, 2]
assert f.numel() > 0
if f.ndim == 0:
f = f[np.newaxis]
# Separable?
if separable is None:
separable = f.ndim == 1 and f.numel() >= 8
if f.ndim == 1 and not separable:
f = f.ger(f)
assert f.ndim == (1 if separable else 2)
# Apply normalize, flip, gain, and device.
if normalize:
f /= f.sum()
if flip_filter:
f = f.flip(list(range(f.ndim)))
f = f * (gain ** (f.ndim / 2))
f = f.to(device=device)
return f
def _get_filter_size(f):
if f is None:
return 1, 1
assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
fw = f.shape[-1]
fh = f.shape[0]
fw = int(fw)
fh = int(fh)
assert fw >= 1 and fh >= 1
return fw, fh
def _get_weight_shape(w):
shape = [int(sz) for sz in w.shape]
return shape
def _parse_scaling(scaling):
if isinstance(scaling, int):
scaling = [scaling, scaling]
assert isinstance(scaling, (list, tuple))
assert all(isinstance(x, int) for x in scaling)
sx, sy = scaling
assert sx >= 1 and sy >= 1
return sx, sy
def _parse_padding(padding):
if isinstance(padding, int):
padding = [padding, padding]
assert isinstance(padding, (list, tuple))
assert all(isinstance(x, int) for x in padding)
if len(padding) == 2:
padx, pady = padding
padding = [padx, padx, pady, pady]
padx0, padx1, pady0, pady1 = padding
return padx0, padx1, pady0, pady1
def _ntuple(n):
def parse(x):
if isinstance(x, collections.abc.Iterable):
return x
return tuple(repeat(x, n))
return parse
to_2tuple = _ntuple(2)
def _upfirdn2d_ref(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1):
"""Slow reference implementation of `upfirdn2d()` using standard PyTorch ops."""
# Validate arguments.
assert isinstance(x, torch.Tensor) and x.ndim == 4
if f is None:
f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
assert f.dtype == torch.float32 and not f.requires_grad
batch_size, num_channels, in_height, in_width = x.shape
# upx, upy = _parse_scaling(up)
# downx, downy = _parse_scaling(down)
upx, upy = up, up
downx, downy = down, down
# padx0, padx1, pady0, pady1 = _parse_padding(padding)
padx0, padx1, pady0, pady1 = padding[0], padding[1], padding[2], padding[3]
# Upsample by inserting zeros.
x = x.reshape([batch_size, num_channels, in_height, 1, in_width, 1])
x = torch.nn.functional.pad(x, [0, upx - 1, 0, 0, 0, upy - 1])
x = x.reshape([batch_size, num_channels, in_height * upy, in_width * upx])
# Pad or crop.
x = torch.nn.functional.pad(
x, [max(padx0, 0), max(padx1, 0), max(pady0, 0), max(pady1, 0)]
)
x = x[
:,
:,
max(-pady0, 0) : x.shape[2] - max(-pady1, 0),
max(-padx0, 0) : x.shape[3] - max(-padx1, 0),
]
# Setup filter.
f = f * (gain ** (f.ndim / 2))
f = f.to(x.dtype)
if not flip_filter:
f = f.flip(list(range(f.ndim)))
# Convolve with the filter.
f = f[np.newaxis, np.newaxis].repeat([num_channels, 1] + [1] * f.ndim)
if f.ndim == 4:
x = conv2d(input=x, weight=f, groups=num_channels)
else:
x = conv2d(input=x, weight=f.unsqueeze(2), groups=num_channels)
x = conv2d(input=x, weight=f.unsqueeze(3), groups=num_channels)
# Downsample by throwing away pixels.
x = x[:, :, ::downy, ::downx]
return x
def upfirdn2d(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1, impl="cuda"):
r"""Pad, upsample, filter, and downsample a batch of 2D images.
Performs the following sequence of operations for each channel:
1. Upsample the image by inserting N-1 zeros after each pixel (`up`).
2. Pad the image with the specified number of zeros on each side (`padding`).
Negative padding corresponds to cropping the image.
3. Convolve the image with the specified 2D FIR filter (`f`), shrinking it
so that the footprint of all output pixels lies within the input image.
4. Downsample the image by keeping every Nth pixel (`down`).
This sequence of operations bears close resemblance to scipy.signal.upfirdn().
The fused op is considerably more efficient than performing the same calculation
using standard PyTorch ops. It supports gradients of arbitrary order.
Args:
x: Float32/float64/float16 input tensor of the shape
`[batch_size, num_channels, in_height, in_width]`.
f: Float32 FIR filter of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable), or
`None` (identity).
up: Integer upsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
down: Integer downsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
padding: Padding with respect to the upsampled image. Can be a single number
or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
flip_filter: False = convolution, True = correlation (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
impl: Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
# assert isinstance(x, torch.Tensor)
# assert impl in ['ref', 'cuda']
return _upfirdn2d_ref(
x, f, up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain
)
def upsample2d(x, f, up=2, padding=0, flip_filter=False, gain=1, impl="cuda"):
r"""Upsample a batch of 2D images using the given 2D FIR filter.
By default, the result is padded so that its shape is a multiple of the input.
User-specified padding is applied on top of that, with negative values
indicating cropping. Pixels outside the image are assumed to be zero.
Args:
x: Float32/float64/float16 input tensor of the shape
`[batch_size, num_channels, in_height, in_width]`.
f: Float32 FIR filter of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable), or
`None` (identity).
up: Integer upsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
padding: Padding with respect to the output. Can be a single number or a
list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
flip_filter: False = convolution, True = correlation (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
impl: Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
upx, upy = _parse_scaling(up)
# upx, upy = up, up
padx0, padx1, pady0, pady1 = _parse_padding(padding)
# padx0, padx1, pady0, pady1 = padding, padding, padding, padding
fw, fh = _get_filter_size(f)
p = [
padx0 + (fw + upx - 1) // 2,
padx1 + (fw - upx) // 2,
pady0 + (fh + upy - 1) // 2,
pady1 + (fh - upy) // 2,
]
return upfirdn2d(
x,
f,
up=up,
padding=p,
flip_filter=flip_filter,
gain=gain * upx * upy,
impl=impl,
)
class FullyConnectedLayer(torch.nn.Module):
def __init__(
self,
in_features, # Number of input features.
out_features, # Number of output features.
bias=True, # Apply additive bias before the activation function?
activation="linear", # Activation function: 'relu', 'lrelu', etc.
lr_multiplier=1, # Learning rate multiplier.
bias_init=0, # Initial value for the additive bias.
):
super().__init__()
self.weight = torch.nn.Parameter(
torch.randn([out_features, in_features]) / lr_multiplier
)
self.bias = (
torch.nn.Parameter(torch.full([out_features], np.float32(bias_init)))
if bias
else None
)
self.activation = activation
self.weight_gain = lr_multiplier / np.sqrt(in_features)
self.bias_gain = lr_multiplier
def forward(self, x):
w = self.weight * self.weight_gain
b = self.bias
if b is not None and self.bias_gain != 1:
b = b * self.bias_gain
if self.activation == "linear" and b is not None:
# out = torch.addmm(b.unsqueeze(0), x, w.t())
x = x.matmul(w.t().to(x.device))
out = x + b.reshape(
[-1 if i == x.ndim - 1 else 1 for i in range(x.ndim)]
).to(x.device)
else:
x = x.matmul(w.t().to(x.device))
out = bias_act(x, b, act=self.activation, dim=x.ndim - 1).to(x.device)
return out
def _conv2d_wrapper(
x, w, stride=1, padding=0, groups=1, transpose=False, flip_weight=True
):
"""Wrapper for the underlying `conv2d()` and `conv_transpose2d()` implementations."""
out_channels, in_channels_per_group, kh, kw = _get_weight_shape(w)
# Flip weight if requested.
if (
not flip_weight
): # conv2d() actually performs correlation (flip_weight=True) not convolution (flip_weight=False).
w = w.flip([2, 3])
# Workaround performance pitfall in cuDNN 8.0.5, triggered when using
# 1x1 kernel + memory_format=channels_last + less than 64 channels.
if (
kw == 1
and kh == 1
and stride == 1
and padding in [0, [0, 0], (0, 0)]
and not transpose
):
if x.stride()[1] == 1 and min(out_channels, in_channels_per_group) < 64:
if out_channels <= 4 and groups == 1:
in_shape = x.shape
x = w.squeeze(3).squeeze(2) @ x.reshape(
[in_shape[0], in_channels_per_group, -1]
)
x = x.reshape([in_shape[0], out_channels, in_shape[2], in_shape[3]])
else:
x = x.to(memory_format=torch.contiguous_format)
w = w.to(memory_format=torch.contiguous_format)
x = conv2d(x, w, groups=groups)
return x.to(memory_format=torch.channels_last)
# Otherwise => execute using conv2d_gradfix.
op = conv_transpose2d if transpose else conv2d
return op(x, w, stride=stride, padding=padding, groups=groups)
def conv2d_resample(
x, w, f=None, up=1, down=1, padding=0, groups=1, flip_weight=True, flip_filter=False
):
r"""2D convolution with optional up/downsampling.
Padding is performed only once at the beginning, not between the operations.
Args:
x: Input tensor of shape
`[batch_size, in_channels, in_height, in_width]`.
w: Weight tensor of shape
`[out_channels, in_channels//groups, kernel_height, kernel_width]`.
f: Low-pass filter for up/downsampling. Must be prepared beforehand by
calling setup_filter(). None = identity (default).
up: Integer upsampling factor (default: 1).
down: Integer downsampling factor (default: 1).
padding: Padding with respect to the upsampled image. Can be a single number
or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
groups: Split input channels into N groups (default: 1).
flip_weight: False = convolution, True = correlation (default: True).
flip_filter: False = convolution, True = correlation (default: False).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
# Validate arguments.
assert isinstance(x, torch.Tensor) and (x.ndim == 4)
assert isinstance(w, torch.Tensor) and (w.ndim == 4) and (w.dtype == x.dtype)
assert f is None or (
isinstance(f, torch.Tensor) and f.ndim in [1, 2] and f.dtype == torch.float32
)
assert isinstance(up, int) and (up >= 1)
assert isinstance(down, int) and (down >= 1)
# assert isinstance(groups, int) and (groups >= 1), f"!!!!!! groups: {groups} isinstance(groups, int) {isinstance(groups, int)} {type(groups)}"
out_channels, in_channels_per_group, kh, kw = _get_weight_shape(w)
fw, fh = _get_filter_size(f)
# px0, px1, py0, py1 = _parse_padding(padding)
px0, px1, py0, py1 = padding, padding, padding, padding
# Adjust padding to account for up/downsampling.
if up > 1:
px0 += (fw + up - 1) // 2
px1 += (fw - up) // 2
py0 += (fh + up - 1) // 2
py1 += (fh - up) // 2
if down > 1:
px0 += (fw - down + 1) // 2
px1 += (fw - down) // 2
py0 += (fh - down + 1) // 2
py1 += (fh - down) // 2
# Fast path: 1x1 convolution with downsampling only => downsample first, then convolve.
if kw == 1 and kh == 1 and (down > 1 and up == 1):
x = upfirdn2d(
x=x, f=f, down=down, padding=[px0, px1, py0, py1], flip_filter=flip_filter
)
x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
return x
# Fast path: 1x1 convolution with upsampling only => convolve first, then upsample.
if kw == 1 and kh == 1 and (up > 1 and down == 1):
x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
x = upfirdn2d(
x=x,
f=f,
up=up,
padding=[px0, px1, py0, py1],
gain=up**2,
flip_filter=flip_filter,
)
return x
# Fast path: downsampling only => use strided convolution.
if down > 1 and up == 1:
x = upfirdn2d(x=x, f=f, padding=[px0, px1, py0, py1], flip_filter=flip_filter)
x = _conv2d_wrapper(
x=x, w=w, stride=down, groups=groups, flip_weight=flip_weight
)
return x
# Fast path: upsampling with optional downsampling => use transpose strided convolution.
if up > 1:
if groups == 1:
w = w.transpose(0, 1)
else:
w = w.reshape(groups, out_channels // groups, in_channels_per_group, kh, kw)
w = w.transpose(1, 2)
w = w.reshape(
groups * in_channels_per_group, out_channels // groups, kh, kw
)
px0 -= kw - 1
px1 -= kw - up
py0 -= kh - 1
py1 -= kh - up
pxt = max(min(-px0, -px1), 0)
pyt = max(min(-py0, -py1), 0)
x = _conv2d_wrapper(
x=x,
w=w,
stride=up,
padding=[pyt, pxt],
groups=groups,
transpose=True,
flip_weight=(not flip_weight),
)
x = upfirdn2d(
x=x,
f=f,
padding=[px0 + pxt, px1 + pxt, py0 + pyt, py1 + pyt],
gain=up**2,
flip_filter=flip_filter,
)
if down > 1:
x = upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
return x
# Fast path: no up/downsampling, padding supported by the underlying implementation => use plain conv2d.
if up == 1 and down == 1:
if px0 == px1 and py0 == py1 and px0 >= 0 and py0 >= 0:
return _conv2d_wrapper(
x=x, w=w, padding=[py0, px0], groups=groups, flip_weight=flip_weight
)
# Fallback: Generic reference implementation.
x = upfirdn2d(
x=x,
f=(f if up > 1 else None),
up=up,
padding=[px0, px1, py0, py1],
gain=up**2,
flip_filter=flip_filter,
)
x = _conv2d_wrapper(x=x, w=w, groups=groups, flip_weight=flip_weight)
if down > 1:
x = upfirdn2d(x=x, f=f, down=down, flip_filter=flip_filter)
return x
class Conv2dLayer(torch.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.
channels_last=False, # Expect the input to have memory_format=channels_last?
trainable=True, # Update the weights of this layer during training?
):
super().__init__()
self.activation = activation
self.up = up
self.down = down
self.register_buffer("resample_filter", setup_filter(resample_filter))
self.conv_clamp = conv_clamp
self.padding = kernel_size // 2
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size**2))
self.act_gain = activation_funcs[activation].def_gain
memory_format = (
torch.channels_last if channels_last else torch.contiguous_format
)
weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(
memory_format=memory_format
)
bias = torch.zeros([out_channels]) if bias else None
if trainable:
self.weight = torch.nn.Parameter(weight)
self.bias = torch.nn.Parameter(bias) if bias is not None else None
else:
self.register_buffer("weight", weight)
if bias is not None:
self.register_buffer("bias", bias)
else:
self.bias = None
def forward(self, x, gain=1):
w = self.weight * self.weight_gain
x = conv2d_resample(
x=x,
w=w,
f=self.resample_filter,
up=self.up,
down=self.down,
padding=self.padding,
)
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
comfy_extras/chainner_models/model_loading.py
View file @ 766c7b38
import logging as logger import logging as logger
from .architecture.DAT import DAT
from .architecture.face.codeformer import CodeFormer from .architecture.face.codeformer import CodeFormer
from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean
from .architecture.face.restoreformer_arch import RestoreFormer from .architecture.face.restoreformer_arch import RestoreFormer
from .architecture.HAT import HAT from .architecture.HAT import HAT
from .architecture.LaMa import LaMa from .architecture.LaMa import LaMa
from .architecture.MAT import MAT
from .architecture.OmniSR.OmniSR import OmniSR from .architecture.OmniSR.OmniSR import OmniSR
from .architecture.RRDB import RRDBNet as ESRGAN from .architecture.RRDB import RRDBNet as ESRGAN
from .architecture.SCUNet import SCUNet
from .architecture.SPSR import SPSRNet as SPSR from .architecture.SPSR import SPSRNet as SPSR
from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2 from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2
from .architecture.SwiftSRGAN import Generator as SwiftSRGAN from .architecture.SwiftSRGAN import Generator as SwiftSRGAN
...@@ -33,7 +34,6 @@ def load_state_dict(state_dict) -> PyTorchModel: ...@@ -33,7 +34,6 @@ def load_state_dict(state_dict) -> PyTorchModel:
state_dict = state_dict["params"] state_dict = state_dict["params"]
state_dict_keys = list(state_dict.keys()) state_dict_keys = list(state_dict.keys())
# SRVGGNet Real-ESRGAN (v2) # SRVGGNet Real-ESRGAN (v2)
if "body.0.weight" in state_dict_keys and "body.1.weight" in state_dict_keys: if "body.0.weight" in state_dict_keys and "body.1.weight" in state_dict_keys:
model = RealESRGANv2(state_dict) model = RealESRGANv2(state_dict)
...@@ -46,12 +46,14 @@ def load_state_dict(state_dict) -> PyTorchModel: ...@@ -46,12 +46,14 @@ def load_state_dict(state_dict) -> PyTorchModel:
and "initial.cnn.depthwise.weight" in state_dict["model"].keys() and "initial.cnn.depthwise.weight" in state_dict["model"].keys()
): ):
model = SwiftSRGAN(state_dict) model = SwiftSRGAN(state_dict)
# HAT -- be sure it is above swinir # SwinIR, Swin2SR, HAT
elif "layers.0.residual_group.blocks.0.conv_block.cab.0.weight" in state_dict_keys:
model = HAT(state_dict)
# SwinIR
elif "layers.0.residual_group.blocks.0.norm1.weight" in state_dict_keys: elif "layers.0.residual_group.blocks.0.norm1.weight" in state_dict_keys:
if "patch_embed.proj.weight" in state_dict_keys: if (
"layers.0.residual_group.blocks.0.conv_block.cab.0.weight"
in state_dict_keys
):
model = HAT(state_dict)
elif "patch_embed.proj.weight" in state_dict_keys:
model = Swin2SR(state_dict) model = Swin2SR(state_dict)
else: else:
model = SwinIR(state_dict) model = SwinIR(state_dict)
...@@ -78,12 +80,15 @@ def load_state_dict(state_dict) -> PyTorchModel: ...@@ -78,12 +80,15 @@ def load_state_dict(state_dict) -> PyTorchModel:
or "generator.model.1.bn_l.running_mean" in state_dict_keys or "generator.model.1.bn_l.running_mean" in state_dict_keys
): ):
model = LaMa(state_dict) model = LaMa(state_dict)
# MAT
elif "synthesis.first_stage.conv_first.conv.resample_filter" in state_dict_keys:
model = MAT(state_dict)
# Omni-SR # Omni-SR
elif "residual_layer.0.residual_layer.0.layer.0.fn.0.weight" in state_dict_keys: elif "residual_layer.0.residual_layer.0.layer.0.fn.0.weight" in state_dict_keys:
model = OmniSR(state_dict) model = OmniSR(state_dict)
# SCUNet
elif "m_head.0.weight" in state_dict_keys and "m_tail.0.weight" in state_dict_keys:
model = SCUNet(state_dict)
# DAT
elif "layers.0.blocks.2.attn.attn_mask_0" in state_dict_keys:
model = DAT(state_dict)
# Regular ESRGAN, "new-arch" ESRGAN, Real-ESRGAN v1 # Regular ESRGAN, "new-arch" ESRGAN, Real-ESRGAN v1
else: else:
try: try:
......
comfy_extras/chainner_models/types.py
View file @ 766c7b38
from typing import Union from typing import Union
from .architecture.DAT import DAT
from .architecture.face.codeformer import CodeFormer from .architecture.face.codeformer import CodeFormer
from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean
from .architecture.face.restoreformer_arch import RestoreFormer from .architecture.face.restoreformer_arch import RestoreFormer
from .architecture.HAT import HAT from .architecture.HAT import HAT
from .architecture.LaMa import LaMa from .architecture.LaMa import LaMa
from .architecture.MAT import MAT
from .architecture.OmniSR.OmniSR import OmniSR from .architecture.OmniSR.OmniSR import OmniSR
from .architecture.RRDB import RRDBNet as ESRGAN from .architecture.RRDB import RRDBNet as ESRGAN
from .architecture.SCUNet import SCUNet
from .architecture.SPSR import SPSRNet as SPSR from .architecture.SPSR import SPSRNet as SPSR
from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2 from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2
from .architecture.SwiftSRGAN import Generator as SwiftSRGAN from .architecture.SwiftSRGAN import Generator as SwiftSRGAN
from .architecture.Swin2SR import Swin2SR from .architecture.Swin2SR import Swin2SR
from .architecture.SwinIR import SwinIR from .architecture.SwinIR import SwinIR
PyTorchSRModels = (RealESRGANv2, SPSR, SwiftSRGAN, ESRGAN, SwinIR, Swin2SR, HAT, OmniSR) PyTorchSRModels = (
RealESRGANv2,
SPSR,
SwiftSRGAN,
ESRGAN,
SwinIR,
Swin2SR,
HAT,
OmniSR,
SCUNet,
DAT,
)
PyTorchSRModel = Union[ PyTorchSRModel = Union[
RealESRGANv2, RealESRGANv2,
SPSR, SPSR,
...@@ -24,6 +36,8 @@ PyTorchSRModel = Union[ ...@@ -24,6 +36,8 @@ PyTorchSRModel = Union[
Swin2SR, Swin2SR,
HAT, HAT,
OmniSR, OmniSR,
SCUNet,
DAT,
] ]
...@@ -39,8 +53,8 @@ def is_pytorch_face_model(model: object): ...@@ -39,8 +53,8 @@ def is_pytorch_face_model(model: object):
return isinstance(model, PyTorchFaceModels) return isinstance(model, PyTorchFaceModels)
PyTorchInpaintModels = (LaMa, MAT) PyTorchInpaintModels = (LaMa,)
PyTorchInpaintModel = Union[LaMa, MAT] PyTorchInpaintModel = Union[LaMa]
def is_pytorch_inpaint_model(model: object): def is_pytorch_inpaint_model(model: object):
......
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