"vscode:/vscode.git/clone" did not exist on "b73b18b6db9cf2d15622435367ea3d07e0f2c800"
Unverified Commit 8cfd677c authored by Joey Ballentine's avatar Joey Ballentine Committed by GitHub
Browse files

Replace chainner_models with Spandrel package (#2146)

* Replace chainner_models with Spandrel

* Update to latest spandrel

* Use spandrel_foss instead

* update spandrel to new FOSS-compliant version
parent ffc4b7c3
# pylint: skip-file
# type: ignore
"""Modified from https://github.com/wzhouxiff/RestoreFormer
"""
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
class VectorQuantizer(nn.Module):
"""
see https://github.com/MishaLaskin/vqvae/blob/d761a999e2267766400dc646d82d3ac3657771d4/models/quantizer.py
____________________________________________
Discretization bottleneck part of the VQ-VAE.
Inputs:
- n_e : number of embeddings
- e_dim : dimension of embedding
- beta : commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2
_____________________________________________
"""
def __init__(self, n_e, e_dim, beta):
super(VectorQuantizer, self).__init__()
self.n_e = n_e
self.e_dim = e_dim
self.beta = beta
self.embedding = nn.Embedding(self.n_e, self.e_dim)
self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
def forward(self, z):
"""
Inputs the output of the encoder network z and maps it to a discrete
one-hot vector that is the index of the closest embedding vector e_j
z (continuous) -> z_q (discrete)
z.shape = (batch, channel, height, width)
quantization pipeline:
1. get encoder input (B,C,H,W)
2. flatten input to (B*H*W,C)
"""
# reshape z -> (batch, height, width, channel) and flatten
z = z.permute(0, 2, 3, 1).contiguous()
z_flattened = z.view(-1, self.e_dim)
# distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
d = (
torch.sum(z_flattened**2, dim=1, keepdim=True)
+ torch.sum(self.embedding.weight**2, dim=1)
- 2 * torch.matmul(z_flattened, self.embedding.weight.t())
)
# could possible replace this here
# #\start...
# find closest encodings
min_value, min_encoding_indices = torch.min(d, dim=1)
min_encoding_indices = min_encoding_indices.unsqueeze(1)
min_encodings = torch.zeros(min_encoding_indices.shape[0], self.n_e).to(z)
min_encodings.scatter_(1, min_encoding_indices, 1)
# dtype min encodings: torch.float32
# min_encodings shape: torch.Size([2048, 512])
# min_encoding_indices.shape: torch.Size([2048, 1])
# get quantized latent vectors
z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape)
# .........\end
# with:
# .........\start
# min_encoding_indices = torch.argmin(d, dim=1)
# z_q = self.embedding(min_encoding_indices)
# ......\end......... (TODO)
# compute loss for embedding
loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean(
(z_q - z.detach()) ** 2
)
# preserve gradients
z_q = z + (z_q - z).detach()
# perplexity
e_mean = torch.mean(min_encodings, dim=0)
perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10)))
# reshape back to match original input shape
z_q = z_q.permute(0, 3, 1, 2).contiguous()
return z_q, loss, (perplexity, min_encodings, min_encoding_indices, d)
def get_codebook_entry(self, indices, shape):
# shape specifying (batch, height, width, channel)
# TODO: check for more easy handling with nn.Embedding
min_encodings = torch.zeros(indices.shape[0], self.n_e).to(indices)
min_encodings.scatter_(1, indices[:, None], 1)
# get quantized latent vectors
z_q = torch.matmul(min_encodings.float(), self.embedding.weight)
if shape is not None:
z_q = z_q.view(shape)
# reshape back to match original input shape
z_q = z_q.permute(0, 3, 1, 2).contiguous()
return z_q
# pytorch_diffusion + derived encoder decoder
def nonlinearity(x):
# swish
return x * torch.sigmoid(x)
def Normalize(in_channels):
return torch.nn.GroupNorm(
num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
)
class Upsample(nn.Module):
def __init__(self, in_channels, with_conv):
super().__init__()
self.with_conv = with_conv
if self.with_conv:
self.conv = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=3, stride=1, padding=1
)
def forward(self, x):
x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
if self.with_conv:
x = self.conv(x)
return x
class Downsample(nn.Module):
def __init__(self, in_channels, with_conv):
super().__init__()
self.with_conv = with_conv
if self.with_conv:
# no asymmetric padding in torch conv, must do it ourselves
self.conv = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=3, stride=2, padding=0
)
def forward(self, x):
if self.with_conv:
pad = (0, 1, 0, 1)
x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
x = self.conv(x)
else:
x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
return x
class ResnetBlock(nn.Module):
def __init__(
self,
*,
in_channels,
out_channels=None,
conv_shortcut=False,
dropout,
temb_channels=512
):
super().__init__()
self.in_channels = in_channels
out_channels = in_channels if out_channels is None else out_channels
self.out_channels = out_channels
self.use_conv_shortcut = conv_shortcut
self.norm1 = Normalize(in_channels)
self.conv1 = torch.nn.Conv2d(
in_channels, out_channels, kernel_size=3, stride=1, padding=1
)
if temb_channels > 0:
self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
self.norm2 = Normalize(out_channels)
self.dropout = torch.nn.Dropout(dropout)
self.conv2 = torch.nn.Conv2d(
out_channels, out_channels, kernel_size=3, stride=1, padding=1
)
if self.in_channels != self.out_channels:
if self.use_conv_shortcut:
self.conv_shortcut = torch.nn.Conv2d(
in_channels, out_channels, kernel_size=3, stride=1, padding=1
)
else:
self.nin_shortcut = torch.nn.Conv2d(
in_channels, out_channels, kernel_size=1, stride=1, padding=0
)
def forward(self, x, temb):
h = x
h = self.norm1(h)
h = nonlinearity(h)
h = self.conv1(h)
if temb is not None:
h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
h = self.norm2(h)
h = nonlinearity(h)
h = self.dropout(h)
h = self.conv2(h)
if self.in_channels != self.out_channels:
if self.use_conv_shortcut:
x = self.conv_shortcut(x)
else:
x = self.nin_shortcut(x)
return x + h
class MultiHeadAttnBlock(nn.Module):
def __init__(self, in_channels, head_size=1):
super().__init__()
self.in_channels = in_channels
self.head_size = head_size
self.att_size = in_channels // head_size
assert (
in_channels % head_size == 0
), "The size of head should be divided by the number of channels."
self.norm1 = Normalize(in_channels)
self.norm2 = Normalize(in_channels)
self.q = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.k = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.v = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.proj_out = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.num = 0
def forward(self, x, y=None):
h_ = x
h_ = self.norm1(h_)
if y is None:
y = h_
else:
y = self.norm2(y)
q = self.q(y)
k = self.k(h_)
v = self.v(h_)
# compute attention
b, c, h, w = q.shape
q = q.reshape(b, self.head_size, self.att_size, h * w)
q = q.permute(0, 3, 1, 2) # b, hw, head, att
k = k.reshape(b, self.head_size, self.att_size, h * w)
k = k.permute(0, 3, 1, 2)
v = v.reshape(b, self.head_size, self.att_size, h * w)
v = v.permute(0, 3, 1, 2)
q = q.transpose(1, 2)
v = v.transpose(1, 2)
k = k.transpose(1, 2).transpose(2, 3)
scale = int(self.att_size) ** (-0.5)
q.mul_(scale)
w_ = torch.matmul(q, k)
w_ = F.softmax(w_, dim=3)
w_ = w_.matmul(v)
w_ = w_.transpose(1, 2).contiguous() # [b, h*w, head, att]
w_ = w_.view(b, h, w, -1)
w_ = w_.permute(0, 3, 1, 2)
w_ = self.proj_out(w_)
return x + w_
class MultiHeadEncoder(nn.Module):
def __init__(
self,
ch,
out_ch,
ch_mult=(1, 2, 4, 8),
num_res_blocks=2,
attn_resolutions=(16,),
dropout=0.0,
resamp_with_conv=True,
in_channels=3,
resolution=512,
z_channels=256,
double_z=True,
enable_mid=True,
head_size=1,
**ignore_kwargs
):
super().__init__()
self.ch = ch
self.temb_ch = 0
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
self.enable_mid = enable_mid
# downsampling
self.conv_in = torch.nn.Conv2d(
in_channels, self.ch, kernel_size=3, stride=1, padding=1
)
curr_res = resolution
in_ch_mult = (1,) + tuple(ch_mult)
self.down = nn.ModuleList()
for i_level in range(self.num_resolutions):
block = nn.ModuleList()
attn = nn.ModuleList()
block_in = ch * in_ch_mult[i_level]
block_out = ch * ch_mult[i_level]
for i_block in range(self.num_res_blocks):
block.append(
ResnetBlock(
in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout,
)
)
block_in = block_out
if curr_res in attn_resolutions:
attn.append(MultiHeadAttnBlock(block_in, head_size))
down = nn.Module()
down.block = block
down.attn = attn
if i_level != self.num_resolutions - 1:
down.downsample = Downsample(block_in, resamp_with_conv)
curr_res = curr_res // 2
self.down.append(down)
# middle
if self.enable_mid:
self.mid = nn.Module()
self.mid.block_1 = ResnetBlock(
in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout,
)
self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
self.mid.block_2 = ResnetBlock(
in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout,
)
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(
block_in,
2 * z_channels if double_z else z_channels,
kernel_size=3,
stride=1,
padding=1,
)
def forward(self, x):
hs = {}
# timestep embedding
temb = None
# downsampling
h = self.conv_in(x)
hs["in"] = h
for i_level in range(self.num_resolutions):
for i_block in range(self.num_res_blocks):
h = self.down[i_level].block[i_block](h, temb)
if len(self.down[i_level].attn) > 0:
h = self.down[i_level].attn[i_block](h)
if i_level != self.num_resolutions - 1:
# hs.append(h)
hs["block_" + str(i_level)] = h
h = self.down[i_level].downsample(h)
# middle
# h = hs[-1]
if self.enable_mid:
h = self.mid.block_1(h, temb)
hs["block_" + str(i_level) + "_atten"] = h
h = self.mid.attn_1(h)
h = self.mid.block_2(h, temb)
hs["mid_atten"] = h
# end
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
# hs.append(h)
hs["out"] = h
return hs
class MultiHeadDecoder(nn.Module):
def __init__(
self,
ch,
out_ch,
ch_mult=(1, 2, 4, 8),
num_res_blocks=2,
attn_resolutions=(16,),
dropout=0.0,
resamp_with_conv=True,
in_channels=3,
resolution=512,
z_channels=256,
give_pre_end=False,
enable_mid=True,
head_size=1,
**ignorekwargs
):
super().__init__()
self.ch = ch
self.temb_ch = 0
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
self.give_pre_end = give_pre_end
self.enable_mid = enable_mid
# compute in_ch_mult, block_in and curr_res at lowest res
block_in = ch * ch_mult[self.num_resolutions - 1]
curr_res = resolution // 2 ** (self.num_resolutions - 1)
self.z_shape = (1, z_channels, curr_res, curr_res)
print(
"Working with z of shape {} = {} dimensions.".format(
self.z_shape, np.prod(self.z_shape)
)
)
# z to block_in
self.conv_in = torch.nn.Conv2d(
z_channels, block_in, kernel_size=3, stride=1, padding=1
)
# middle
if self.enable_mid:
self.mid = nn.Module()
self.mid.block_1 = ResnetBlock(
in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout,
)
self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
self.mid.block_2 = ResnetBlock(
in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout,
)
# upsampling
self.up = nn.ModuleList()
for i_level in reversed(range(self.num_resolutions)):
block = nn.ModuleList()
attn = nn.ModuleList()
block_out = ch * ch_mult[i_level]
for i_block in range(self.num_res_blocks + 1):
block.append(
ResnetBlock(
in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout,
)
)
block_in = block_out
if curr_res in attn_resolutions:
attn.append(MultiHeadAttnBlock(block_in, head_size))
up = nn.Module()
up.block = block
up.attn = attn
if i_level != 0:
up.upsample = Upsample(block_in, resamp_with_conv)
curr_res = curr_res * 2
self.up.insert(0, up) # prepend to get consistent order
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(
block_in, out_ch, kernel_size=3, stride=1, padding=1
)
def forward(self, z):
# assert z.shape[1:] == self.z_shape[1:]
self.last_z_shape = z.shape
# timestep embedding
temb = None
# z to block_in
h = self.conv_in(z)
# middle
if self.enable_mid:
h = self.mid.block_1(h, temb)
h = self.mid.attn_1(h)
h = self.mid.block_2(h, temb)
# upsampling
for i_level in reversed(range(self.num_resolutions)):
for i_block in range(self.num_res_blocks + 1):
h = self.up[i_level].block[i_block](h, temb)
if len(self.up[i_level].attn) > 0:
h = self.up[i_level].attn[i_block](h)
if i_level != 0:
h = self.up[i_level].upsample(h)
# end
if self.give_pre_end:
return h
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
return h
class MultiHeadDecoderTransformer(nn.Module):
def __init__(
self,
ch,
out_ch,
ch_mult=(1, 2, 4, 8),
num_res_blocks=2,
attn_resolutions=(16,),
dropout=0.0,
resamp_with_conv=True,
in_channels=3,
resolution=512,
z_channels=256,
give_pre_end=False,
enable_mid=True,
head_size=1,
**ignorekwargs
):
super().__init__()
self.ch = ch
self.temb_ch = 0
self.num_resolutions = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
self.give_pre_end = give_pre_end
self.enable_mid = enable_mid
# compute in_ch_mult, block_in and curr_res at lowest res
block_in = ch * ch_mult[self.num_resolutions - 1]
curr_res = resolution // 2 ** (self.num_resolutions - 1)
self.z_shape = (1, z_channels, curr_res, curr_res)
print(
"Working with z of shape {} = {} dimensions.".format(
self.z_shape, np.prod(self.z_shape)
)
)
# z to block_in
self.conv_in = torch.nn.Conv2d(
z_channels, block_in, kernel_size=3, stride=1, padding=1
)
# middle
if self.enable_mid:
self.mid = nn.Module()
self.mid.block_1 = ResnetBlock(
in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout,
)
self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
self.mid.block_2 = ResnetBlock(
in_channels=block_in,
out_channels=block_in,
temb_channels=self.temb_ch,
dropout=dropout,
)
# upsampling
self.up = nn.ModuleList()
for i_level in reversed(range(self.num_resolutions)):
block = nn.ModuleList()
attn = nn.ModuleList()
block_out = ch * ch_mult[i_level]
for i_block in range(self.num_res_blocks + 1):
block.append(
ResnetBlock(
in_channels=block_in,
out_channels=block_out,
temb_channels=self.temb_ch,
dropout=dropout,
)
)
block_in = block_out
if curr_res in attn_resolutions:
attn.append(MultiHeadAttnBlock(block_in, head_size))
up = nn.Module()
up.block = block
up.attn = attn
if i_level != 0:
up.upsample = Upsample(block_in, resamp_with_conv)
curr_res = curr_res * 2
self.up.insert(0, up) # prepend to get consistent order
# end
self.norm_out = Normalize(block_in)
self.conv_out = torch.nn.Conv2d(
block_in, out_ch, kernel_size=3, stride=1, padding=1
)
def forward(self, z, hs):
# assert z.shape[1:] == self.z_shape[1:]
# self.last_z_shape = z.shape
# timestep embedding
temb = None
# z to block_in
h = self.conv_in(z)
# middle
if self.enable_mid:
h = self.mid.block_1(h, temb)
h = self.mid.attn_1(h, hs["mid_atten"])
h = self.mid.block_2(h, temb)
# upsampling
for i_level in reversed(range(self.num_resolutions)):
for i_block in range(self.num_res_blocks + 1):
h = self.up[i_level].block[i_block](h, temb)
if len(self.up[i_level].attn) > 0:
h = self.up[i_level].attn[i_block](
h, hs["block_" + str(i_level) + "_atten"]
)
# hfeature = h.clone()
if i_level != 0:
h = self.up[i_level].upsample(h)
# end
if self.give_pre_end:
return h
h = self.norm_out(h)
h = nonlinearity(h)
h = self.conv_out(h)
return h
class RestoreFormer(nn.Module):
def __init__(
self,
state_dict,
):
super(RestoreFormer, self).__init__()
n_embed = 1024
embed_dim = 256
ch = 64
out_ch = 3
ch_mult = (1, 2, 2, 4, 4, 8)
num_res_blocks = 2
attn_resolutions = (16,)
dropout = 0.0
in_channels = 3
resolution = 512
z_channels = 256
double_z = False
enable_mid = True
fix_decoder = False
fix_codebook = True
fix_encoder = False
head_size = 8
self.model_arch = "RestoreFormer"
self.sub_type = "Face SR"
self.scale = 8
self.in_nc = 3
self.out_nc = out_ch
self.state = state_dict
self.supports_fp16 = False
self.supports_bf16 = True
self.min_size_restriction = 16
self.encoder = MultiHeadEncoder(
ch=ch,
out_ch=out_ch,
ch_mult=ch_mult,
num_res_blocks=num_res_blocks,
attn_resolutions=attn_resolutions,
dropout=dropout,
in_channels=in_channels,
resolution=resolution,
z_channels=z_channels,
double_z=double_z,
enable_mid=enable_mid,
head_size=head_size,
)
self.decoder = MultiHeadDecoderTransformer(
ch=ch,
out_ch=out_ch,
ch_mult=ch_mult,
num_res_blocks=num_res_blocks,
attn_resolutions=attn_resolutions,
dropout=dropout,
in_channels=in_channels,
resolution=resolution,
z_channels=z_channels,
enable_mid=enable_mid,
head_size=head_size,
)
self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25)
self.quant_conv = torch.nn.Conv2d(z_channels, embed_dim, 1)
self.post_quant_conv = torch.nn.Conv2d(embed_dim, z_channels, 1)
if fix_decoder:
for _, param in self.decoder.named_parameters():
param.requires_grad = False
for _, param in self.post_quant_conv.named_parameters():
param.requires_grad = False
for _, param in self.quantize.named_parameters():
param.requires_grad = False
elif fix_codebook:
for _, param in self.quantize.named_parameters():
param.requires_grad = False
if fix_encoder:
for _, param in self.encoder.named_parameters():
param.requires_grad = False
self.load_state_dict(state_dict)
def encode(self, x):
hs = self.encoder(x)
h = self.quant_conv(hs["out"])
quant, emb_loss, info = self.quantize(h)
return quant, emb_loss, info, hs
def decode(self, quant, hs):
quant = self.post_quant_conv(quant)
dec = self.decoder(quant, hs)
return dec
def forward(self, input, **kwargs):
quant, diff, info, hs = self.encode(input)
dec = self.decode(quant, hs)
return dec, None
# pylint: skip-file
# type: ignore
import math
import random
import torch
from torch import nn
from torch.nn import functional as F
from .fused_act import FusedLeakyReLU, fused_leaky_relu
from .upfirdn2d import upfirdn2d
class NormStyleCode(nn.Module):
def forward(self, x):
"""Normalize the style codes.
Args:
x (Tensor): Style codes with shape (b, c).
Returns:
Tensor: Normalized tensor.
"""
return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
def make_resample_kernel(k):
"""Make resampling kernel for UpFirDn.
Args:
k (list[int]): A list indicating the 1D resample kernel magnitude.
Returns:
Tensor: 2D resampled kernel.
"""
k = torch.tensor(k, dtype=torch.float32)
if k.ndim == 1:
k = k[None, :] * k[:, None] # to 2D kernel, outer product
# normalize
k /= k.sum()
return k
class UpFirDnUpsample(nn.Module):
"""Upsample, FIR filter, and downsample (upsampole version).
References:
1. https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.upfirdn.html # noqa: E501
2. http://www.ece.northwestern.edu/local-apps/matlabhelp/toolbox/signal/upfirdn.html # noqa: E501
Args:
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude.
factor (int): Upsampling scale factor. Default: 2.
"""
def __init__(self, resample_kernel, factor=2):
super(UpFirDnUpsample, self).__init__()
self.kernel = make_resample_kernel(resample_kernel) * (factor**2)
self.factor = factor
pad = self.kernel.shape[0] - factor
self.pad = ((pad + 1) // 2 + factor - 1, pad // 2)
def forward(self, x):
out = upfirdn2d(x, self.kernel.type_as(x), up=self.factor, down=1, pad=self.pad)
return out
def __repr__(self):
return f"{self.__class__.__name__}(factor={self.factor})"
class UpFirDnDownsample(nn.Module):
"""Upsample, FIR filter, and downsample (downsampole version).
Args:
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude.
factor (int): Downsampling scale factor. Default: 2.
"""
def __init__(self, resample_kernel, factor=2):
super(UpFirDnDownsample, self).__init__()
self.kernel = make_resample_kernel(resample_kernel)
self.factor = factor
pad = self.kernel.shape[0] - factor
self.pad = ((pad + 1) // 2, pad // 2)
def forward(self, x):
out = upfirdn2d(x, self.kernel.type_as(x), up=1, down=self.factor, pad=self.pad)
return out
def __repr__(self):
return f"{self.__class__.__name__}(factor={self.factor})"
class UpFirDnSmooth(nn.Module):
"""Upsample, FIR filter, and downsample (smooth version).
Args:
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude.
upsample_factor (int): Upsampling scale factor. Default: 1.
downsample_factor (int): Downsampling scale factor. Default: 1.
kernel_size (int): Kernel size: Default: 1.
"""
def __init__(
self, resample_kernel, upsample_factor=1, downsample_factor=1, kernel_size=1
):
super(UpFirDnSmooth, self).__init__()
self.upsample_factor = upsample_factor
self.downsample_factor = downsample_factor
self.kernel = make_resample_kernel(resample_kernel)
if upsample_factor > 1:
self.kernel = self.kernel * (upsample_factor**2)
if upsample_factor > 1:
pad = (self.kernel.shape[0] - upsample_factor) - (kernel_size - 1)
self.pad = ((pad + 1) // 2 + upsample_factor - 1, pad // 2 + 1)
elif downsample_factor > 1:
pad = (self.kernel.shape[0] - downsample_factor) + (kernel_size - 1)
self.pad = ((pad + 1) // 2, pad // 2)
else:
raise NotImplementedError
def forward(self, x):
out = upfirdn2d(x, self.kernel.type_as(x), up=1, down=1, pad=self.pad)
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(upsample_factor={self.upsample_factor}"
f", downsample_factor={self.downsample_factor})"
)
class EqualLinear(nn.Module):
"""Equalized Linear as StyleGAN2.
Args:
in_channels (int): Size of each sample.
out_channels (int): Size of each output sample.
bias (bool): If set to ``False``, the layer will not learn an additive
bias. Default: ``True``.
bias_init_val (float): Bias initialized value. Default: 0.
lr_mul (float): Learning rate multiplier. Default: 1.
activation (None | str): The activation after ``linear`` operation.
Supported: 'fused_lrelu', None. Default: None.
"""
def __init__(
self,
in_channels,
out_channels,
bias=True,
bias_init_val=0,
lr_mul=1,
activation=None,
):
super(EqualLinear, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.lr_mul = lr_mul
self.activation = activation
if self.activation not in ["fused_lrelu", None]:
raise ValueError(
f"Wrong activation value in EqualLinear: {activation}"
"Supported ones are: ['fused_lrelu', None]."
)
self.scale = (1 / math.sqrt(in_channels)) * lr_mul
self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul))
if bias:
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
else:
self.register_parameter("bias", None)
def forward(self, x):
if self.bias is None:
bias = None
else:
bias = self.bias * self.lr_mul
if self.activation == "fused_lrelu":
out = F.linear(x, self.weight * self.scale)
out = fused_leaky_relu(out, bias)
else:
out = F.linear(x, self.weight * self.scale, bias=bias)
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, bias={self.bias is not None})"
)
class ModulatedConv2d(nn.Module):
"""Modulated Conv2d used in StyleGAN2.
There is no bias in ModulatedConv2d.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
num_style_feat (int): Channel number of style features.
demodulate (bool): Whether to demodulate in the conv layer.
Default: True.
sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
Default: None.
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude. Default: (1, 3, 3, 1).
eps (float): A value added to the denominator for numerical stability.
Default: 1e-8.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=True,
sample_mode=None,
resample_kernel=(1, 3, 3, 1),
eps=1e-8,
):
super(ModulatedConv2d, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.demodulate = demodulate
self.sample_mode = sample_mode
self.eps = eps
if self.sample_mode == "upsample":
self.smooth = UpFirDnSmooth(
resample_kernel,
upsample_factor=2,
downsample_factor=1,
kernel_size=kernel_size,
)
elif self.sample_mode == "downsample":
self.smooth = UpFirDnSmooth(
resample_kernel,
upsample_factor=1,
downsample_factor=2,
kernel_size=kernel_size,
)
elif self.sample_mode is None:
pass
else:
raise ValueError(
f"Wrong sample mode {self.sample_mode}, "
"supported ones are ['upsample', 'downsample', None]."
)
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
# modulation inside each modulated conv
self.modulation = EqualLinear(
num_style_feat,
in_channels,
bias=True,
bias_init_val=1,
lr_mul=1,
activation=None,
)
self.weight = nn.Parameter(
torch.randn(1, out_channels, in_channels, kernel_size, kernel_size)
)
self.padding = kernel_size // 2
def forward(self, x, style):
"""Forward function.
Args:
x (Tensor): Tensor with shape (b, c, h, w).
style (Tensor): Tensor with shape (b, num_style_feat).
Returns:
Tensor: Modulated tensor after convolution.
"""
b, c, h, w = x.shape # c = c_in
# weight modulation
style = self.modulation(style).view(b, 1, c, 1, 1)
# self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
weight = self.scale * self.weight * style # (b, c_out, c_in, k, k)
if self.demodulate:
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
weight = weight.view(
b * self.out_channels, c, self.kernel_size, self.kernel_size
)
if self.sample_mode == "upsample":
x = x.view(1, b * c, h, w)
weight = weight.view(
b, self.out_channels, c, self.kernel_size, self.kernel_size
)
weight = weight.transpose(1, 2).reshape(
b * c, self.out_channels, self.kernel_size, self.kernel_size
)
out = F.conv_transpose2d(x, weight, padding=0, stride=2, groups=b)
out = out.view(b, self.out_channels, *out.shape[2:4])
out = self.smooth(out)
elif self.sample_mode == "downsample":
x = self.smooth(x)
x = x.view(1, b * c, *x.shape[2:4])
out = F.conv2d(x, weight, padding=0, stride=2, groups=b)
out = out.view(b, self.out_channels, *out.shape[2:4])
else:
x = x.view(1, b * c, h, w)
# weight: (b*c_out, c_in, k, k), groups=b
out = F.conv2d(x, weight, padding=self.padding, groups=b)
out = out.view(b, self.out_channels, *out.shape[2:4])
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, "
f"kernel_size={self.kernel_size}, "
f"demodulate={self.demodulate}, sample_mode={self.sample_mode})"
)
class StyleConv(nn.Module):
"""Style conv.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
num_style_feat (int): Channel number of style features.
demodulate (bool): Whether demodulate in the conv layer. Default: True.
sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
Default: None.
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude. Default: (1, 3, 3, 1).
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=True,
sample_mode=None,
resample_kernel=(1, 3, 3, 1),
):
super(StyleConv, self).__init__()
self.modulated_conv = ModulatedConv2d(
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=demodulate,
sample_mode=sample_mode,
resample_kernel=resample_kernel,
)
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
self.activate = FusedLeakyReLU(out_channels)
def forward(self, x, style, noise=None):
# modulate
out = self.modulated_conv(x, style)
# noise injection
if noise is None:
b, _, h, w = out.shape
noise = out.new_empty(b, 1, h, w).normal_()
out = out + self.weight * noise
# activation (with bias)
out = self.activate(out)
return out
class ToRGB(nn.Module):
"""To RGB from features.
Args:
in_channels (int): Channel number of input.
num_style_feat (int): Channel number of style features.
upsample (bool): Whether to upsample. Default: True.
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude. Default: (1, 3, 3, 1).
"""
def __init__(
self, in_channels, num_style_feat, upsample=True, resample_kernel=(1, 3, 3, 1)
):
super(ToRGB, self).__init__()
if upsample:
self.upsample = UpFirDnUpsample(resample_kernel, factor=2)
else:
self.upsample = None
self.modulated_conv = ModulatedConv2d(
in_channels,
3,
kernel_size=1,
num_style_feat=num_style_feat,
demodulate=False,
sample_mode=None,
)
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
def forward(self, x, style, skip=None):
"""Forward function.
Args:
x (Tensor): Feature tensor with shape (b, c, h, w).
style (Tensor): Tensor with shape (b, num_style_feat).
skip (Tensor): Base/skip tensor. Default: None.
Returns:
Tensor: RGB images.
"""
out = self.modulated_conv(x, style)
out = out + self.bias
if skip is not None:
if self.upsample:
skip = self.upsample(skip)
out = out + skip
return out
class ConstantInput(nn.Module):
"""Constant input.
Args:
num_channel (int): Channel number of constant input.
size (int): Spatial size of constant input.
"""
def __init__(self, num_channel, size):
super(ConstantInput, self).__init__()
self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
def forward(self, batch):
out = self.weight.repeat(batch, 1, 1, 1)
return out
class StyleGAN2Generator(nn.Module):
"""StyleGAN2 Generator.
Args:
out_size (int): The spatial size of outputs.
num_style_feat (int): Channel number of style features. Default: 512.
num_mlp (int): Layer number of MLP style layers. Default: 8.
channel_multiplier (int): Channel multiplier for large networks of
StyleGAN2. Default: 2.
resample_kernel (list[int]): A list indicating the 1D resample kernel
magnitude. A cross production will be applied to extent 1D resample
kernel to 2D resample kernel. Default: (1, 3, 3, 1).
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
narrow (float): Narrow ratio for channels. Default: 1.0.
"""
def __init__(
self,
out_size,
num_style_feat=512,
num_mlp=8,
channel_multiplier=2,
resample_kernel=(1, 3, 3, 1),
lr_mlp=0.01,
narrow=1,
):
super(StyleGAN2Generator, self).__init__()
# Style MLP layers
self.num_style_feat = num_style_feat
style_mlp_layers = [NormStyleCode()]
for i in range(num_mlp):
style_mlp_layers.append(
EqualLinear(
num_style_feat,
num_style_feat,
bias=True,
bias_init_val=0,
lr_mul=lr_mlp,
activation="fused_lrelu",
)
)
self.style_mlp = nn.Sequential(*style_mlp_layers)
channels = {
"4": int(512 * narrow),
"8": int(512 * narrow),
"16": int(512 * narrow),
"32": int(512 * narrow),
"64": int(256 * channel_multiplier * narrow),
"128": int(128 * channel_multiplier * narrow),
"256": int(64 * channel_multiplier * narrow),
"512": int(32 * channel_multiplier * narrow),
"1024": int(16 * channel_multiplier * narrow),
}
self.channels = channels
self.constant_input = ConstantInput(channels["4"], size=4)
self.style_conv1 = StyleConv(
channels["4"],
channels["4"],
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode=None,
resample_kernel=resample_kernel,
)
self.to_rgb1 = ToRGB(
channels["4"],
num_style_feat,
upsample=False,
resample_kernel=resample_kernel,
)
self.log_size = int(math.log(out_size, 2))
self.num_layers = (self.log_size - 2) * 2 + 1
self.num_latent = self.log_size * 2 - 2
self.style_convs = nn.ModuleList()
self.to_rgbs = nn.ModuleList()
self.noises = nn.Module()
in_channels = channels["4"]
# noise
for layer_idx in range(self.num_layers):
resolution = 2 ** ((layer_idx + 5) // 2)
shape = [1, 1, resolution, resolution]
self.noises.register_buffer(f"noise{layer_idx}", torch.randn(*shape))
# style convs and to_rgbs
for i in range(3, self.log_size + 1):
out_channels = channels[f"{2**i}"]
self.style_convs.append(
StyleConv(
in_channels,
out_channels,
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode="upsample",
resample_kernel=resample_kernel,
)
)
self.style_convs.append(
StyleConv(
out_channels,
out_channels,
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode=None,
resample_kernel=resample_kernel,
)
)
self.to_rgbs.append(
ToRGB(
out_channels,
num_style_feat,
upsample=True,
resample_kernel=resample_kernel,
)
)
in_channels = out_channels
def make_noise(self):
"""Make noise for noise injection."""
device = self.constant_input.weight.device
noises = [torch.randn(1, 1, 4, 4, device=device)]
for i in range(3, self.log_size + 1):
for _ in range(2):
noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
return noises
def get_latent(self, x):
return self.style_mlp(x)
def mean_latent(self, num_latent):
latent_in = torch.randn(
num_latent, self.num_style_feat, device=self.constant_input.weight.device
)
latent = self.style_mlp(latent_in).mean(0, keepdim=True)
return latent
def forward(
self,
styles,
input_is_latent=False,
noise=None,
randomize_noise=True,
truncation=1,
truncation_latent=None,
inject_index=None,
return_latents=False,
):
"""Forward function for StyleGAN2Generator.
Args:
styles (list[Tensor]): Sample codes of styles.
input_is_latent (bool): Whether input is latent style.
Default: False.
noise (Tensor | None): Input noise or None. Default: None.
randomize_noise (bool): Randomize noise, used when 'noise' is
False. Default: True.
truncation (float): TODO. Default: 1.
truncation_latent (Tensor | None): TODO. Default: None.
inject_index (int | None): The injection index for mixing noise.
Default: None.
return_latents (bool): Whether to return style latents.
Default: False.
"""
# style codes -> latents with Style MLP layer
if not input_is_latent:
styles = [self.style_mlp(s) for s in styles]
# noises
if noise is None:
if randomize_noise:
noise = [None] * self.num_layers # for each style conv layer
else: # use the stored noise
noise = [
getattr(self.noises, f"noise{i}") for i in range(self.num_layers)
]
# style truncation
if truncation < 1:
style_truncation = []
for style in styles:
style_truncation.append(
truncation_latent + truncation * (style - truncation_latent)
)
styles = style_truncation
# get style latent with injection
if len(styles) == 1:
inject_index = self.num_latent
if styles[0].ndim < 3:
# repeat latent code for all the layers
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
else: # used for encoder with different latent code for each layer
latent = styles[0]
elif len(styles) == 2: # mixing noises
if inject_index is None:
inject_index = random.randint(1, self.num_latent - 1)
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
latent2 = (
styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
)
latent = torch.cat([latent1, latent2], 1)
# main generation
out = self.constant_input(latent.shape[0])
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
skip = self.to_rgb1(out, latent[:, 1])
i = 1
for conv1, conv2, noise1, noise2, to_rgb in zip(
self.style_convs[::2],
self.style_convs[1::2],
noise[1::2],
noise[2::2],
self.to_rgbs,
):
out = conv1(out, latent[:, i], noise=noise1)
out = conv2(out, latent[:, i + 1], noise=noise2)
skip = to_rgb(out, latent[:, i + 2], skip)
i += 2
image = skip
if return_latents:
return image, latent
else:
return image, None
class ScaledLeakyReLU(nn.Module):
"""Scaled LeakyReLU.
Args:
negative_slope (float): Negative slope. Default: 0.2.
"""
def __init__(self, negative_slope=0.2):
super(ScaledLeakyReLU, self).__init__()
self.negative_slope = negative_slope
def forward(self, x):
out = F.leaky_relu(x, negative_slope=self.negative_slope)
return out * math.sqrt(2)
class EqualConv2d(nn.Module):
"""Equalized Linear as StyleGAN2.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
stride (int): Stride of the convolution. Default: 1
padding (int): Zero-padding added to both sides of the input.
Default: 0.
bias (bool): If ``True``, adds a learnable bias to the output.
Default: ``True``.
bias_init_val (float): Bias initialized value. Default: 0.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
stride=1,
padding=0,
bias=True,
bias_init_val=0,
):
super(EqualConv2d, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
self.weight = nn.Parameter(
torch.randn(out_channels, in_channels, kernel_size, kernel_size)
)
if bias:
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
else:
self.register_parameter("bias", None)
def forward(self, x):
out = F.conv2d(
x,
self.weight * self.scale,
bias=self.bias,
stride=self.stride,
padding=self.padding,
)
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, "
f"kernel_size={self.kernel_size},"
f" stride={self.stride}, padding={self.padding}, "
f"bias={self.bias is not None})"
)
class ConvLayer(nn.Sequential):
"""Conv Layer used in StyleGAN2 Discriminator.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Kernel size.
downsample (bool): Whether downsample by a factor of 2.
Default: False.
resample_kernel (list[int]): A list indicating the 1D resample
kernel magnitude. A cross production will be applied to
extent 1D resample kernel to 2D resample kernel.
Default: (1, 3, 3, 1).
bias (bool): Whether with bias. Default: True.
activate (bool): Whether use activateion. Default: True.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
downsample=False,
resample_kernel=(1, 3, 3, 1),
bias=True,
activate=True,
):
layers = []
# downsample
if downsample:
layers.append(
UpFirDnSmooth(
resample_kernel,
upsample_factor=1,
downsample_factor=2,
kernel_size=kernel_size,
)
)
stride = 2
self.padding = 0
else:
stride = 1
self.padding = kernel_size // 2
# conv
layers.append(
EqualConv2d(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=self.padding,
bias=bias and not activate,
)
)
# activation
if activate:
if bias:
layers.append(FusedLeakyReLU(out_channels))
else:
layers.append(ScaledLeakyReLU(0.2))
super(ConvLayer, self).__init__(*layers)
class ResBlock(nn.Module):
"""Residual block used in StyleGAN2 Discriminator.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
resample_kernel (list[int]): A list indicating the 1D resample
kernel magnitude. A cross production will be applied to
extent 1D resample kernel to 2D resample kernel.
Default: (1, 3, 3, 1).
"""
def __init__(self, in_channels, out_channels, resample_kernel=(1, 3, 3, 1)):
super(ResBlock, self).__init__()
self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
self.conv2 = ConvLayer(
in_channels,
out_channels,
3,
downsample=True,
resample_kernel=resample_kernel,
bias=True,
activate=True,
)
self.skip = ConvLayer(
in_channels,
out_channels,
1,
downsample=True,
resample_kernel=resample_kernel,
bias=False,
activate=False,
)
def forward(self, x):
out = self.conv1(x)
out = self.conv2(out)
skip = self.skip(x)
out = (out + skip) / math.sqrt(2)
return out
# pylint: skip-file
# type: ignore
import math
import random
import torch
from torch import nn
from torch.nn import functional as F
from .fused_act import FusedLeakyReLU, fused_leaky_relu
class NormStyleCode(nn.Module):
def forward(self, x):
"""Normalize the style codes.
Args:
x (Tensor): Style codes with shape (b, c).
Returns:
Tensor: Normalized tensor.
"""
return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
class EqualLinear(nn.Module):
"""Equalized Linear as StyleGAN2.
Args:
in_channels (int): Size of each sample.
out_channels (int): Size of each output sample.
bias (bool): If set to ``False``, the layer will not learn an additive
bias. Default: ``True``.
bias_init_val (float): Bias initialized value. Default: 0.
lr_mul (float): Learning rate multiplier. Default: 1.
activation (None | str): The activation after ``linear`` operation.
Supported: 'fused_lrelu', None. Default: None.
"""
def __init__(
self,
in_channels,
out_channels,
bias=True,
bias_init_val=0,
lr_mul=1,
activation=None,
):
super(EqualLinear, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.lr_mul = lr_mul
self.activation = activation
if self.activation not in ["fused_lrelu", None]:
raise ValueError(
f"Wrong activation value in EqualLinear: {activation}"
"Supported ones are: ['fused_lrelu', None]."
)
self.scale = (1 / math.sqrt(in_channels)) * lr_mul
self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul))
if bias:
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
else:
self.register_parameter("bias", None)
def forward(self, x):
if self.bias is None:
bias = None
else:
bias = self.bias * self.lr_mul
if self.activation == "fused_lrelu":
out = F.linear(x, self.weight * self.scale)
out = fused_leaky_relu(out, bias)
else:
out = F.linear(x, self.weight * self.scale, bias=bias)
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, bias={self.bias is not None})"
)
class ModulatedConv2d(nn.Module):
"""Modulated Conv2d used in StyleGAN2.
There is no bias in ModulatedConv2d.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
num_style_feat (int): Channel number of style features.
demodulate (bool): Whether to demodulate in the conv layer.
Default: True.
sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
Default: None.
eps (float): A value added to the denominator for numerical stability.
Default: 1e-8.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=True,
sample_mode=None,
eps=1e-8,
interpolation_mode="bilinear",
):
super(ModulatedConv2d, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.demodulate = demodulate
self.sample_mode = sample_mode
self.eps = eps
self.interpolation_mode = interpolation_mode
if self.interpolation_mode == "nearest":
self.align_corners = None
else:
self.align_corners = False
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
# modulation inside each modulated conv
self.modulation = EqualLinear(
num_style_feat,
in_channels,
bias=True,
bias_init_val=1,
lr_mul=1,
activation=None,
)
self.weight = nn.Parameter(
torch.randn(1, out_channels, in_channels, kernel_size, kernel_size)
)
self.padding = kernel_size // 2
def forward(self, x, style):
"""Forward function.
Args:
x (Tensor): Tensor with shape (b, c, h, w).
style (Tensor): Tensor with shape (b, num_style_feat).
Returns:
Tensor: Modulated tensor after convolution.
"""
b, c, h, w = x.shape # c = c_in
# weight modulation
style = self.modulation(style).view(b, 1, c, 1, 1)
# self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
weight = self.scale * self.weight * style # (b, c_out, c_in, k, k)
if self.demodulate:
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
weight = weight.view(
b * self.out_channels, c, self.kernel_size, self.kernel_size
)
if self.sample_mode == "upsample":
x = F.interpolate(
x,
scale_factor=2,
mode=self.interpolation_mode,
align_corners=self.align_corners,
)
elif self.sample_mode == "downsample":
x = F.interpolate(
x,
scale_factor=0.5,
mode=self.interpolation_mode,
align_corners=self.align_corners,
)
b, c, h, w = x.shape
x = x.view(1, b * c, h, w)
# weight: (b*c_out, c_in, k, k), groups=b
out = F.conv2d(x, weight, padding=self.padding, groups=b)
out = out.view(b, self.out_channels, *out.shape[2:4])
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, "
f"kernel_size={self.kernel_size}, "
f"demodulate={self.demodulate}, sample_mode={self.sample_mode})"
)
class StyleConv(nn.Module):
"""Style conv.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
num_style_feat (int): Channel number of style features.
demodulate (bool): Whether demodulate in the conv layer. Default: True.
sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
Default: None.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=True,
sample_mode=None,
interpolation_mode="bilinear",
):
super(StyleConv, self).__init__()
self.modulated_conv = ModulatedConv2d(
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=demodulate,
sample_mode=sample_mode,
interpolation_mode=interpolation_mode,
)
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
self.activate = FusedLeakyReLU(out_channels)
def forward(self, x, style, noise=None):
# modulate
out = self.modulated_conv(x, style)
# noise injection
if noise is None:
b, _, h, w = out.shape
noise = out.new_empty(b, 1, h, w).normal_()
out = out + self.weight * noise
# activation (with bias)
out = self.activate(out)
return out
class ToRGB(nn.Module):
"""To RGB from features.
Args:
in_channels (int): Channel number of input.
num_style_feat (int): Channel number of style features.
upsample (bool): Whether to upsample. Default: True.
"""
def __init__(
self, in_channels, num_style_feat, upsample=True, interpolation_mode="bilinear"
):
super(ToRGB, self).__init__()
self.upsample = upsample
self.interpolation_mode = interpolation_mode
if self.interpolation_mode == "nearest":
self.align_corners = None
else:
self.align_corners = False
self.modulated_conv = ModulatedConv2d(
in_channels,
3,
kernel_size=1,
num_style_feat=num_style_feat,
demodulate=False,
sample_mode=None,
interpolation_mode=interpolation_mode,
)
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
def forward(self, x, style, skip=None):
"""Forward function.
Args:
x (Tensor): Feature tensor with shape (b, c, h, w).
style (Tensor): Tensor with shape (b, num_style_feat).
skip (Tensor): Base/skip tensor. Default: None.
Returns:
Tensor: RGB images.
"""
out = self.modulated_conv(x, style)
out = out + self.bias
if skip is not None:
if self.upsample:
skip = F.interpolate(
skip,
scale_factor=2,
mode=self.interpolation_mode,
align_corners=self.align_corners,
)
out = out + skip
return out
class ConstantInput(nn.Module):
"""Constant input.
Args:
num_channel (int): Channel number of constant input.
size (int): Spatial size of constant input.
"""
def __init__(self, num_channel, size):
super(ConstantInput, self).__init__()
self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
def forward(self, batch):
out = self.weight.repeat(batch, 1, 1, 1)
return out
class StyleGAN2GeneratorBilinear(nn.Module):
"""StyleGAN2 Generator.
Args:
out_size (int): The spatial size of outputs.
num_style_feat (int): Channel number of style features. Default: 512.
num_mlp (int): Layer number of MLP style layers. Default: 8.
channel_multiplier (int): Channel multiplier for large networks of
StyleGAN2. Default: 2.
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
narrow (float): Narrow ratio for channels. Default: 1.0.
"""
def __init__(
self,
out_size,
num_style_feat=512,
num_mlp=8,
channel_multiplier=2,
lr_mlp=0.01,
narrow=1,
interpolation_mode="bilinear",
):
super(StyleGAN2GeneratorBilinear, self).__init__()
# Style MLP layers
self.num_style_feat = num_style_feat
style_mlp_layers = [NormStyleCode()]
for i in range(num_mlp):
style_mlp_layers.append(
EqualLinear(
num_style_feat,
num_style_feat,
bias=True,
bias_init_val=0,
lr_mul=lr_mlp,
activation="fused_lrelu",
)
)
self.style_mlp = nn.Sequential(*style_mlp_layers)
channels = {
"4": int(512 * narrow),
"8": int(512 * narrow),
"16": int(512 * narrow),
"32": int(512 * narrow),
"64": int(256 * channel_multiplier * narrow),
"128": int(128 * channel_multiplier * narrow),
"256": int(64 * channel_multiplier * narrow),
"512": int(32 * channel_multiplier * narrow),
"1024": int(16 * channel_multiplier * narrow),
}
self.channels = channels
self.constant_input = ConstantInput(channels["4"], size=4)
self.style_conv1 = StyleConv(
channels["4"],
channels["4"],
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode=None,
interpolation_mode=interpolation_mode,
)
self.to_rgb1 = ToRGB(
channels["4"],
num_style_feat,
upsample=False,
interpolation_mode=interpolation_mode,
)
self.log_size = int(math.log(out_size, 2))
self.num_layers = (self.log_size - 2) * 2 + 1
self.num_latent = self.log_size * 2 - 2
self.style_convs = nn.ModuleList()
self.to_rgbs = nn.ModuleList()
self.noises = nn.Module()
in_channels = channels["4"]
# noise
for layer_idx in range(self.num_layers):
resolution = 2 ** ((layer_idx + 5) // 2)
shape = [1, 1, resolution, resolution]
self.noises.register_buffer(f"noise{layer_idx}", torch.randn(*shape))
# style convs and to_rgbs
for i in range(3, self.log_size + 1):
out_channels = channels[f"{2**i}"]
self.style_convs.append(
StyleConv(
in_channels,
out_channels,
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode="upsample",
interpolation_mode=interpolation_mode,
)
)
self.style_convs.append(
StyleConv(
out_channels,
out_channels,
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode=None,
interpolation_mode=interpolation_mode,
)
)
self.to_rgbs.append(
ToRGB(
out_channels,
num_style_feat,
upsample=True,
interpolation_mode=interpolation_mode,
)
)
in_channels = out_channels
def make_noise(self):
"""Make noise for noise injection."""
device = self.constant_input.weight.device
noises = [torch.randn(1, 1, 4, 4, device=device)]
for i in range(3, self.log_size + 1):
for _ in range(2):
noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
return noises
def get_latent(self, x):
return self.style_mlp(x)
def mean_latent(self, num_latent):
latent_in = torch.randn(
num_latent, self.num_style_feat, device=self.constant_input.weight.device
)
latent = self.style_mlp(latent_in).mean(0, keepdim=True)
return latent
def forward(
self,
styles,
input_is_latent=False,
noise=None,
randomize_noise=True,
truncation=1,
truncation_latent=None,
inject_index=None,
return_latents=False,
):
"""Forward function for StyleGAN2Generator.
Args:
styles (list[Tensor]): Sample codes of styles.
input_is_latent (bool): Whether input is latent style.
Default: False.
noise (Tensor | None): Input noise or None. Default: None.
randomize_noise (bool): Randomize noise, used when 'noise' is
False. Default: True.
truncation (float): TODO. Default: 1.
truncation_latent (Tensor | None): TODO. Default: None.
inject_index (int | None): The injection index for mixing noise.
Default: None.
return_latents (bool): Whether to return style latents.
Default: False.
"""
# style codes -> latents with Style MLP layer
if not input_is_latent:
styles = [self.style_mlp(s) for s in styles]
# noises
if noise is None:
if randomize_noise:
noise = [None] * self.num_layers # for each style conv layer
else: # use the stored noise
noise = [
getattr(self.noises, f"noise{i}") for i in range(self.num_layers)
]
# style truncation
if truncation < 1:
style_truncation = []
for style in styles:
style_truncation.append(
truncation_latent + truncation * (style - truncation_latent)
)
styles = style_truncation
# get style latent with injection
if len(styles) == 1:
inject_index = self.num_latent
if styles[0].ndim < 3:
# repeat latent code for all the layers
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
else: # used for encoder with different latent code for each layer
latent = styles[0]
elif len(styles) == 2: # mixing noises
if inject_index is None:
inject_index = random.randint(1, self.num_latent - 1)
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
latent2 = (
styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
)
latent = torch.cat([latent1, latent2], 1)
# main generation
out = self.constant_input(latent.shape[0])
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
skip = self.to_rgb1(out, latent[:, 1])
i = 1
for conv1, conv2, noise1, noise2, to_rgb in zip(
self.style_convs[::2],
self.style_convs[1::2],
noise[1::2],
noise[2::2],
self.to_rgbs,
):
out = conv1(out, latent[:, i], noise=noise1)
out = conv2(out, latent[:, i + 1], noise=noise2)
skip = to_rgb(out, latent[:, i + 2], skip)
i += 2
image = skip
if return_latents:
return image, latent
else:
return image, None
class ScaledLeakyReLU(nn.Module):
"""Scaled LeakyReLU.
Args:
negative_slope (float): Negative slope. Default: 0.2.
"""
def __init__(self, negative_slope=0.2):
super(ScaledLeakyReLU, self).__init__()
self.negative_slope = negative_slope
def forward(self, x):
out = F.leaky_relu(x, negative_slope=self.negative_slope)
return out * math.sqrt(2)
class EqualConv2d(nn.Module):
"""Equalized Linear as StyleGAN2.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
stride (int): Stride of the convolution. Default: 1
padding (int): Zero-padding added to both sides of the input.
Default: 0.
bias (bool): If ``True``, adds a learnable bias to the output.
Default: ``True``.
bias_init_val (float): Bias initialized value. Default: 0.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
stride=1,
padding=0,
bias=True,
bias_init_val=0,
):
super(EqualConv2d, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
self.weight = nn.Parameter(
torch.randn(out_channels, in_channels, kernel_size, kernel_size)
)
if bias:
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
else:
self.register_parameter("bias", None)
def forward(self, x):
out = F.conv2d(
x,
self.weight * self.scale,
bias=self.bias,
stride=self.stride,
padding=self.padding,
)
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, "
f"kernel_size={self.kernel_size},"
f" stride={self.stride}, padding={self.padding}, "
f"bias={self.bias is not None})"
)
class ConvLayer(nn.Sequential):
"""Conv Layer used in StyleGAN2 Discriminator.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Kernel size.
downsample (bool): Whether downsample by a factor of 2.
Default: False.
bias (bool): Whether with bias. Default: True.
activate (bool): Whether use activateion. Default: True.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
downsample=False,
bias=True,
activate=True,
interpolation_mode="bilinear",
):
layers = []
self.interpolation_mode = interpolation_mode
# downsample
if downsample:
if self.interpolation_mode == "nearest":
self.align_corners = None
else:
self.align_corners = False
layers.append(
torch.nn.Upsample(
scale_factor=0.5,
mode=interpolation_mode,
align_corners=self.align_corners,
)
)
stride = 1
self.padding = kernel_size // 2
# conv
layers.append(
EqualConv2d(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=self.padding,
bias=bias and not activate,
)
)
# activation
if activate:
if bias:
layers.append(FusedLeakyReLU(out_channels))
else:
layers.append(ScaledLeakyReLU(0.2))
super(ConvLayer, self).__init__(*layers)
class ResBlock(nn.Module):
"""Residual block used in StyleGAN2 Discriminator.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
"""
def __init__(self, in_channels, out_channels, interpolation_mode="bilinear"):
super(ResBlock, self).__init__()
self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
self.conv2 = ConvLayer(
in_channels,
out_channels,
3,
downsample=True,
interpolation_mode=interpolation_mode,
bias=True,
activate=True,
)
self.skip = ConvLayer(
in_channels,
out_channels,
1,
downsample=True,
interpolation_mode=interpolation_mode,
bias=False,
activate=False,
)
def forward(self, x):
out = self.conv1(x)
out = self.conv2(out)
skip = self.skip(x)
out = (out + skip) / math.sqrt(2)
return out
# pylint: skip-file
# type: ignore
import math
import torch
from torch import nn
from torch.nn import functional as F
from torch.nn import init
from torch.nn.modules.batchnorm import _BatchNorm
@torch.no_grad()
def default_init_weights(module_list, scale=1, bias_fill=0, **kwargs):
"""Initialize network weights.
Args:
module_list (list[nn.Module] | nn.Module): Modules to be initialized.
scale (float): Scale initialized weights, especially for residual
blocks. Default: 1.
bias_fill (float): The value to fill bias. Default: 0
kwargs (dict): Other arguments for initialization function.
"""
if not isinstance(module_list, list):
module_list = [module_list]
for module in module_list:
for m in module.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, **kwargs)
m.weight.data *= scale
if m.bias is not None:
m.bias.data.fill_(bias_fill)
elif isinstance(m, nn.Linear):
init.kaiming_normal_(m.weight, **kwargs)
m.weight.data *= scale
if m.bias is not None:
m.bias.data.fill_(bias_fill)
elif isinstance(m, _BatchNorm):
init.constant_(m.weight, 1)
if m.bias is not None:
m.bias.data.fill_(bias_fill)
class NormStyleCode(nn.Module):
def forward(self, x):
"""Normalize the style codes.
Args:
x (Tensor): Style codes with shape (b, c).
Returns:
Tensor: Normalized tensor.
"""
return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
class ModulatedConv2d(nn.Module):
"""Modulated Conv2d used in StyleGAN2.
There is no bias in ModulatedConv2d.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
num_style_feat (int): Channel number of style features.
demodulate (bool): Whether to demodulate in the conv layer. Default: True.
sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
eps (float): A value added to the denominator for numerical stability. Default: 1e-8.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=True,
sample_mode=None,
eps=1e-8,
):
super(ModulatedConv2d, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.demodulate = demodulate
self.sample_mode = sample_mode
self.eps = eps
# modulation inside each modulated conv
self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
# initialization
default_init_weights(
self.modulation,
scale=1,
bias_fill=1,
a=0,
mode="fan_in",
nonlinearity="linear",
)
self.weight = nn.Parameter(
torch.randn(1, out_channels, in_channels, kernel_size, kernel_size)
/ math.sqrt(in_channels * kernel_size**2)
)
self.padding = kernel_size // 2
def forward(self, x, style):
"""Forward function.
Args:
x (Tensor): Tensor with shape (b, c, h, w).
style (Tensor): Tensor with shape (b, num_style_feat).
Returns:
Tensor: Modulated tensor after convolution.
"""
b, c, h, w = x.shape # c = c_in
# weight modulation
style = self.modulation(style).view(b, 1, c, 1, 1)
# self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
weight = self.weight * style # (b, c_out, c_in, k, k)
if self.demodulate:
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
weight = weight.view(
b * self.out_channels, c, self.kernel_size, self.kernel_size
)
# upsample or downsample if necessary
if self.sample_mode == "upsample":
x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=False)
elif self.sample_mode == "downsample":
x = F.interpolate(x, scale_factor=0.5, mode="bilinear", align_corners=False)
b, c, h, w = x.shape
x = x.view(1, b * c, h, w)
# weight: (b*c_out, c_in, k, k), groups=b
out = F.conv2d(x, weight, padding=self.padding, groups=b)
out = out.view(b, self.out_channels, *out.shape[2:4])
return out
def __repr__(self):
return (
f"{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, "
f"kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})"
)
class StyleConv(nn.Module):
"""Style conv used in StyleGAN2.
Args:
in_channels (int): Channel number of the input.
out_channels (int): Channel number of the output.
kernel_size (int): Size of the convolving kernel.
num_style_feat (int): Channel number of style features.
demodulate (bool): Whether demodulate in the conv layer. Default: True.
sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
"""
def __init__(
self,
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=True,
sample_mode=None,
):
super(StyleConv, self).__init__()
self.modulated_conv = ModulatedConv2d(
in_channels,
out_channels,
kernel_size,
num_style_feat,
demodulate=demodulate,
sample_mode=sample_mode,
)
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
def forward(self, x, style, noise=None):
# modulate
out = self.modulated_conv(x, style) * 2**0.5 # for conversion
# noise injection
if noise is None:
b, _, h, w = out.shape
noise = out.new_empty(b, 1, h, w).normal_()
out = out + self.weight * noise
# add bias
out = out + self.bias
# activation
out = self.activate(out)
return out
class ToRGB(nn.Module):
"""To RGB (image space) from features.
Args:
in_channels (int): Channel number of input.
num_style_feat (int): Channel number of style features.
upsample (bool): Whether to upsample. Default: True.
"""
def __init__(self, in_channels, num_style_feat, upsample=True):
super(ToRGB, self).__init__()
self.upsample = upsample
self.modulated_conv = ModulatedConv2d(
in_channels,
3,
kernel_size=1,
num_style_feat=num_style_feat,
demodulate=False,
sample_mode=None,
)
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
def forward(self, x, style, skip=None):
"""Forward function.
Args:
x (Tensor): Feature tensor with shape (b, c, h, w).
style (Tensor): Tensor with shape (b, num_style_feat).
skip (Tensor): Base/skip tensor. Default: None.
Returns:
Tensor: RGB images.
"""
out = self.modulated_conv(x, style)
out = out + self.bias
if skip is not None:
if self.upsample:
skip = F.interpolate(
skip, scale_factor=2, mode="bilinear", align_corners=False
)
out = out + skip
return out
class ConstantInput(nn.Module):
"""Constant input.
Args:
num_channel (int): Channel number of constant input.
size (int): Spatial size of constant input.
"""
def __init__(self, num_channel, size):
super(ConstantInput, self).__init__()
self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
def forward(self, batch):
out = self.weight.repeat(batch, 1, 1, 1)
return out
class StyleGAN2GeneratorClean(nn.Module):
"""Clean version of StyleGAN2 Generator.
Args:
out_size (int): The spatial size of outputs.
num_style_feat (int): Channel number of style features. Default: 512.
num_mlp (int): Layer number of MLP style layers. Default: 8.
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
narrow (float): Narrow ratio for channels. Default: 1.0.
"""
def __init__(
self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1
):
super(StyleGAN2GeneratorClean, self).__init__()
# Style MLP layers
self.num_style_feat = num_style_feat
style_mlp_layers = [NormStyleCode()]
for i in range(num_mlp):
style_mlp_layers.extend(
[
nn.Linear(num_style_feat, num_style_feat, bias=True),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
]
)
self.style_mlp = nn.Sequential(*style_mlp_layers)
# initialization
default_init_weights(
self.style_mlp,
scale=1,
bias_fill=0,
a=0.2,
mode="fan_in",
nonlinearity="leaky_relu",
)
# channel list
channels = {
"4": int(512 * narrow),
"8": int(512 * narrow),
"16": int(512 * narrow),
"32": int(512 * narrow),
"64": int(256 * channel_multiplier * narrow),
"128": int(128 * channel_multiplier * narrow),
"256": int(64 * channel_multiplier * narrow),
"512": int(32 * channel_multiplier * narrow),
"1024": int(16 * channel_multiplier * narrow),
}
self.channels = channels
self.constant_input = ConstantInput(channels["4"], size=4)
self.style_conv1 = StyleConv(
channels["4"],
channels["4"],
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode=None,
)
self.to_rgb1 = ToRGB(channels["4"], num_style_feat, upsample=False)
self.log_size = int(math.log(out_size, 2))
self.num_layers = (self.log_size - 2) * 2 + 1
self.num_latent = self.log_size * 2 - 2
self.style_convs = nn.ModuleList()
self.to_rgbs = nn.ModuleList()
self.noises = nn.Module()
in_channels = channels["4"]
# noise
for layer_idx in range(self.num_layers):
resolution = 2 ** ((layer_idx + 5) // 2)
shape = [1, 1, resolution, resolution]
self.noises.register_buffer(f"noise{layer_idx}", torch.randn(*shape))
# style convs and to_rgbs
for i in range(3, self.log_size + 1):
out_channels = channels[f"{2**i}"]
self.style_convs.append(
StyleConv(
in_channels,
out_channels,
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode="upsample",
)
)
self.style_convs.append(
StyleConv(
out_channels,
out_channels,
kernel_size=3,
num_style_feat=num_style_feat,
demodulate=True,
sample_mode=None,
)
)
self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True))
in_channels = out_channels
def make_noise(self):
"""Make noise for noise injection."""
device = self.constant_input.weight.device
noises = [torch.randn(1, 1, 4, 4, device=device)]
for i in range(3, self.log_size + 1):
for _ in range(2):
noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
return noises
def get_latent(self, x):
return self.style_mlp(x)
def mean_latent(self, num_latent):
latent_in = torch.randn(
num_latent, self.num_style_feat, device=self.constant_input.weight.device
)
latent = self.style_mlp(latent_in).mean(0, keepdim=True)
return latent
def forward(
self,
styles,
input_is_latent=False,
noise=None,
randomize_noise=True,
truncation=1,
truncation_latent=None,
inject_index=None,
return_latents=False,
):
"""Forward function for StyleGAN2GeneratorClean.
Args:
styles (list[Tensor]): Sample codes of styles.
input_is_latent (bool): Whether input is latent style. Default: False.
noise (Tensor | None): Input noise or None. Default: None.
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
truncation (float): The truncation ratio. Default: 1.
truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
inject_index (int | None): The injection index for mixing noise. Default: None.
return_latents (bool): Whether to return style latents. Default: False.
"""
# style codes -> latents with Style MLP layer
if not input_is_latent:
styles = [self.style_mlp(s) for s in styles]
# noises
if noise is None:
if randomize_noise:
noise = [None] * self.num_layers # for each style conv layer
else: # use the stored noise
noise = [
getattr(self.noises, f"noise{i}") for i in range(self.num_layers)
]
# style truncation
if truncation < 1:
style_truncation = []
for style in styles:
style_truncation.append(
truncation_latent + truncation * (style - truncation_latent)
)
styles = style_truncation
# get style latents with injection
if len(styles) == 1:
inject_index = self.num_latent
if styles[0].ndim < 3:
# repeat latent code for all the layers
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
else: # used for encoder with different latent code for each layer
latent = styles[0]
elif len(styles) == 2: # mixing noises
if inject_index is None:
inject_index = random.randint(1, self.num_latent - 1)
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
latent2 = (
styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
)
latent = torch.cat([latent1, latent2], 1)
# main generation
out = self.constant_input(latent.shape[0])
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
skip = self.to_rgb1(out, latent[:, 1])
i = 1
for conv1, conv2, noise1, noise2, to_rgb in zip(
self.style_convs[::2],
self.style_convs[1::2],
noise[1::2],
noise[2::2],
self.to_rgbs,
):
out = conv1(out, latent[:, i], noise=noise1)
out = conv2(out, latent[:, i + 1], noise=noise2)
skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space
i += 2
image = skip
if return_latents:
return image, latent
else:
return image, None
# pylint: skip-file
# type: ignore
# modify from https://github.com/rosinality/stylegan2-pytorch/blob/master/op/upfirdn2d.py # noqa:E501
import os
import torch
from torch.autograd import Function
from torch.nn import functional as F
upfirdn2d_ext = None
class UpFirDn2dBackward(Function):
@staticmethod
def forward(
ctx, grad_output, kernel, grad_kernel, up, down, pad, g_pad, in_size, out_size
):
up_x, up_y = up
down_x, down_y = down
g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1 = g_pad
grad_output = grad_output.reshape(-1, out_size[0], out_size[1], 1)
grad_input = upfirdn2d_ext.upfirdn2d(
grad_output,
grad_kernel,
down_x,
down_y,
up_x,
up_y,
g_pad_x0,
g_pad_x1,
g_pad_y0,
g_pad_y1,
)
grad_input = grad_input.view(in_size[0], in_size[1], in_size[2], in_size[3])
ctx.save_for_backward(kernel)
pad_x0, pad_x1, pad_y0, pad_y1 = pad
ctx.up_x = up_x
ctx.up_y = up_y
ctx.down_x = down_x
ctx.down_y = down_y
ctx.pad_x0 = pad_x0
ctx.pad_x1 = pad_x1
ctx.pad_y0 = pad_y0
ctx.pad_y1 = pad_y1
ctx.in_size = in_size
ctx.out_size = out_size
return grad_input
@staticmethod
def backward(ctx, gradgrad_input):
(kernel,) = ctx.saved_tensors
gradgrad_input = gradgrad_input.reshape(-1, ctx.in_size[2], ctx.in_size[3], 1)
gradgrad_out = upfirdn2d_ext.upfirdn2d(
gradgrad_input,
kernel,
ctx.up_x,
ctx.up_y,
ctx.down_x,
ctx.down_y,
ctx.pad_x0,
ctx.pad_x1,
ctx.pad_y0,
ctx.pad_y1,
)
# gradgrad_out = gradgrad_out.view(ctx.in_size[0], ctx.out_size[0],
# ctx.out_size[1], ctx.in_size[3])
gradgrad_out = gradgrad_out.view(
ctx.in_size[0], ctx.in_size[1], ctx.out_size[0], ctx.out_size[1]
)
return gradgrad_out, None, None, None, None, None, None, None, None
class UpFirDn2d(Function):
@staticmethod
def forward(ctx, input, kernel, up, down, pad):
up_x, up_y = up
down_x, down_y = down
pad_x0, pad_x1, pad_y0, pad_y1 = pad
kernel_h, kernel_w = kernel.shape
_, channel, in_h, in_w = input.shape
ctx.in_size = input.shape
input = input.reshape(-1, in_h, in_w, 1)
ctx.save_for_backward(kernel, torch.flip(kernel, [0, 1]))
out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
ctx.out_size = (out_h, out_w)
ctx.up = (up_x, up_y)
ctx.down = (down_x, down_y)
ctx.pad = (pad_x0, pad_x1, pad_y0, pad_y1)
g_pad_x0 = kernel_w - pad_x0 - 1
g_pad_y0 = kernel_h - pad_y0 - 1
g_pad_x1 = in_w * up_x - out_w * down_x + pad_x0 - up_x + 1
g_pad_y1 = in_h * up_y - out_h * down_y + pad_y0 - up_y + 1
ctx.g_pad = (g_pad_x0, g_pad_x1, g_pad_y0, g_pad_y1)
out = upfirdn2d_ext.upfirdn2d(
input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1
)
# out = out.view(major, out_h, out_w, minor)
out = out.view(-1, channel, out_h, out_w)
return out
@staticmethod
def backward(ctx, grad_output):
kernel, grad_kernel = ctx.saved_tensors
grad_input = UpFirDn2dBackward.apply(
grad_output,
kernel,
grad_kernel,
ctx.up,
ctx.down,
ctx.pad,
ctx.g_pad,
ctx.in_size,
ctx.out_size,
)
return grad_input, None, None, None, None
def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
if input.device.type == "cpu":
out = upfirdn2d_native(
input, kernel, up, up, down, down, pad[0], pad[1], pad[0], pad[1]
)
else:
out = UpFirDn2d.apply(
input, kernel, (up, up), (down, down), (pad[0], pad[1], pad[0], pad[1])
)
return out
def upfirdn2d_native(
input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1
):
_, channel, in_h, in_w = input.shape
input = input.reshape(-1, in_h, in_w, 1)
_, in_h, in_w, minor = input.shape
kernel_h, kernel_w = kernel.shape
out = input.view(-1, in_h, 1, in_w, 1, minor)
out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
out = out.view(-1, in_h * up_y, in_w * up_x, minor)
out = F.pad(
out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)]
)
out = out[
:,
max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
:,
]
out = out.permute(0, 3, 1, 2)
out = out.reshape(
[-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1]
)
w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
out = F.conv2d(out, w)
out = out.reshape(
-1,
minor,
in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
)
out = out.permute(0, 2, 3, 1)
out = out[:, ::down_y, ::down_x, :]
out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
return out.view(-1, channel, out_h, out_w)
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 2019 Ross Wightman
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.
\ No newline at end of file
""" DropBlock, DropPath
PyTorch implementations of DropBlock and DropPath (Stochastic Depth) regularization layers.
Papers:
DropBlock: A regularization method for convolutional networks (https://arxiv.org/abs/1810.12890)
Deep Networks with Stochastic Depth (https://arxiv.org/abs/1603.09382)
Code:
DropBlock impl inspired by two Tensorflow impl that I liked:
- https://github.com/tensorflow/tpu/blob/master/models/official/resnet/resnet_model.py#L74
- https://github.com/clovaai/assembled-cnn/blob/master/nets/blocks.py
Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
def drop_block_2d(
x,
drop_prob: float = 0.1,
block_size: int = 7,
gamma_scale: float = 1.0,
with_noise: bool = False,
inplace: bool = False,
batchwise: bool = False,
):
"""DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
DropBlock with an experimental gaussian noise option. This layer has been tested on a few training
runs with success, but needs further validation and possibly optimization for lower runtime impact.
"""
_, C, H, W = x.shape
total_size = W * H
clipped_block_size = min(block_size, min(W, H))
# seed_drop_rate, the gamma parameter
gamma = (
gamma_scale
* drop_prob
* total_size
/ clipped_block_size**2
/ ((W - block_size + 1) * (H - block_size + 1))
)
# Forces the block to be inside the feature map.
w_i, h_i = torch.meshgrid(
torch.arange(W).to(x.device), torch.arange(H).to(x.device)
)
valid_block = (
(w_i >= clipped_block_size // 2) & (w_i < W - (clipped_block_size - 1) // 2)
) & ((h_i >= clipped_block_size // 2) & (h_i < H - (clipped_block_size - 1) // 2))
valid_block = torch.reshape(valid_block, (1, 1, H, W)).to(dtype=x.dtype)
if batchwise:
# one mask for whole batch, quite a bit faster
uniform_noise = torch.rand((1, C, H, W), dtype=x.dtype, device=x.device)
else:
uniform_noise = torch.rand_like(x)
block_mask = ((2 - gamma - valid_block + uniform_noise) >= 1).to(dtype=x.dtype)
block_mask = -F.max_pool2d(
-block_mask,
kernel_size=clipped_block_size, # block_size,
stride=1,
padding=clipped_block_size // 2,
)
if with_noise:
normal_noise = (
torch.randn((1, C, H, W), dtype=x.dtype, device=x.device)
if batchwise
else torch.randn_like(x)
)
if inplace:
x.mul_(block_mask).add_(normal_noise * (1 - block_mask))
else:
x = x * block_mask + normal_noise * (1 - block_mask)
else:
normalize_scale = (
block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)
).to(x.dtype)
if inplace:
x.mul_(block_mask * normalize_scale)
else:
x = x * block_mask * normalize_scale
return x
def drop_block_fast_2d(
x: torch.Tensor,
drop_prob: float = 0.1,
block_size: int = 7,
gamma_scale: float = 1.0,
with_noise: bool = False,
inplace: bool = False,
):
"""DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
DropBlock with an experimental gaussian noise option. Simplied from above without concern for valid
block mask at edges.
"""
_, _, H, W = x.shape
total_size = W * H
clipped_block_size = min(block_size, min(W, H))
gamma = (
gamma_scale
* drop_prob
* total_size
/ clipped_block_size**2
/ ((W - block_size + 1) * (H - block_size + 1))
)
block_mask = torch.empty_like(x).bernoulli_(gamma)
block_mask = F.max_pool2d(
block_mask.to(x.dtype),
kernel_size=clipped_block_size,
stride=1,
padding=clipped_block_size // 2,
)
if with_noise:
normal_noise = torch.empty_like(x).normal_()
if inplace:
x.mul_(1.0 - block_mask).add_(normal_noise * block_mask)
else:
x = x * (1.0 - block_mask) + normal_noise * block_mask
else:
block_mask = 1 - block_mask
normalize_scale = (
block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-6)
).to(dtype=x.dtype)
if inplace:
x.mul_(block_mask * normalize_scale)
else:
x = x * block_mask * normalize_scale
return x
class DropBlock2d(nn.Module):
"""DropBlock. See https://arxiv.org/pdf/1810.12890.pdf"""
def __init__(
self,
drop_prob: float = 0.1,
block_size: int = 7,
gamma_scale: float = 1.0,
with_noise: bool = False,
inplace: bool = False,
batchwise: bool = False,
fast: bool = True,
):
super(DropBlock2d, self).__init__()
self.drop_prob = drop_prob
self.gamma_scale = gamma_scale
self.block_size = block_size
self.with_noise = with_noise
self.inplace = inplace
self.batchwise = batchwise
self.fast = fast # FIXME finish comparisons of fast vs not
def forward(self, x):
if not self.training or not self.drop_prob:
return x
if self.fast:
return drop_block_fast_2d(
x,
self.drop_prob,
self.block_size,
self.gamma_scale,
self.with_noise,
self.inplace,
)
else:
return drop_block_2d(
x,
self.drop_prob,
self.block_size,
self.gamma_scale,
self.with_noise,
self.inplace,
self.batchwise,
)
def drop_path(
x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
'survival rate' as the argument.
"""
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (
x.ndim - 1
) # work with diff dim tensors, not just 2D ConvNets
random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
if keep_prob > 0.0 and scale_by_keep:
random_tensor.div_(keep_prob)
return x * random_tensor
class DropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
self.scale_by_keep = scale_by_keep
def forward(self, x):
return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
def extra_repr(self):
return f"drop_prob={round(self.drop_prob,3):0.3f}"
""" Layer/Module Helpers
Hacked together by / Copyright 2020 Ross Wightman
"""
import collections.abc
from itertools import repeat
# From PyTorch internals
def _ntuple(n):
def parse(x):
if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
return x
return tuple(repeat(x, n))
return parse
to_1tuple = _ntuple(1)
to_2tuple = _ntuple(2)
to_3tuple = _ntuple(3)
to_4tuple = _ntuple(4)
to_ntuple = _ntuple
def make_divisible(v, divisor=8, min_value=None, round_limit=0.9):
min_value = min_value or divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < round_limit * v:
new_v += divisor
return new_v
import math
import warnings
import torch
from torch.nn.init import _calculate_fan_in_and_fan_out
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
# Cut & paste from PyTorch official master until it's in a few official releases - RW
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
def norm_cdf(x):
# Computes standard normal cumulative distribution function
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
if (mean < a - 2 * std) or (mean > b + 2 * std):
warnings.warn(
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
"The distribution of values may be incorrect.",
stacklevel=2,
)
with torch.no_grad():
# Values are generated by using a truncated uniform distribution and
# then using the inverse CDF for the normal distribution.
# Get upper and lower cdf values
l = norm_cdf((a - mean) / std)
u = norm_cdf((b - mean) / std)
# Uniformly fill tensor with values from [l, u], then translate to
# [2l-1, 2u-1].
tensor.uniform_(2 * l - 1, 2 * u - 1)
# Use inverse cdf transform for normal distribution to get truncated
# standard normal
tensor.erfinv_()
# Transform to proper mean, std
tensor.mul_(std * math.sqrt(2.0))
tensor.add_(mean)
# Clamp to ensure it's in the proper range
tensor.clamp_(min=a, max=b)
return tensor
def trunc_normal_(
tensor: torch.Tensor, mean=0.0, std=1.0, a=-2.0, b=2.0
) -> torch.Tensor:
r"""Fills the input Tensor with values drawn from a truncated
normal distribution. The values are effectively drawn from the
normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
with values outside :math:`[a, b]` redrawn until they are within
the bounds. The method used for generating the random values works
best when :math:`a \leq \text{mean} \leq b`.
NOTE: this impl is similar to the PyTorch trunc_normal_, the bounds [a, b] are
applied while sampling the normal with mean/std applied, therefore a, b args
should be adjusted to match the range of mean, std args.
Args:
tensor: an n-dimensional `torch.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
a: the minimum cutoff value
b: the maximum cutoff value
Examples:
>>> w = torch.empty(3, 5)
>>> nn.init.trunc_normal_(w)
"""
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
def trunc_normal_tf_(
tensor: torch.Tensor, mean=0.0, std=1.0, a=-2.0, b=2.0
) -> torch.Tensor:
r"""Fills the input Tensor with values drawn from a truncated
normal distribution. The values are effectively drawn from the
normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
with values outside :math:`[a, b]` redrawn until they are within
the bounds. The method used for generating the random values works
best when :math:`a \leq \text{mean} \leq b`.
NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
and the result is subsquently scaled and shifted by the mean and std args.
Args:
tensor: an n-dimensional `torch.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
a: the minimum cutoff value
b: the maximum cutoff value
Examples:
>>> w = torch.empty(3, 5)
>>> nn.init.trunc_normal_(w)
"""
_no_grad_trunc_normal_(tensor, 0, 1.0, a, b)
with torch.no_grad():
tensor.mul_(std).add_(mean)
return tensor
def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
if mode == "fan_in":
denom = fan_in
elif mode == "fan_out":
denom = fan_out
elif mode == "fan_avg":
denom = (fan_in + fan_out) / 2
variance = scale / denom # type: ignore
if distribution == "truncated_normal":
# constant is stddev of standard normal truncated to (-2, 2)
trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
elif distribution == "normal":
tensor.normal_(std=math.sqrt(variance))
elif distribution == "uniform":
bound = math.sqrt(3 * variance)
# pylint: disable=invalid-unary-operand-type
tensor.uniform_(-bound, bound)
else:
raise ValueError(f"invalid distribution {distribution}")
def lecun_normal_(tensor):
variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
import logging as logger
from .architecture.DAT import DAT
from .architecture.face.codeformer import CodeFormer
from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean
from .architecture.face.restoreformer_arch import RestoreFormer
from .architecture.HAT import HAT
from .architecture.LaMa import LaMa
from .architecture.OmniSR.OmniSR import OmniSR
from .architecture.RRDB import RRDBNet as ESRGAN
from .architecture.SCUNet import SCUNet
from .architecture.SPSR import SPSRNet as SPSR
from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2
from .architecture.SwiftSRGAN import Generator as SwiftSRGAN
from .architecture.Swin2SR import Swin2SR
from .architecture.SwinIR import SwinIR
from .types import PyTorchModel
class UnsupportedModel(Exception):
pass
def load_state_dict(state_dict) -> PyTorchModel:
logger.debug(f"Loading state dict into pytorch model arch")
state_dict_keys = list(state_dict.keys())
if "params_ema" in state_dict_keys:
state_dict = state_dict["params_ema"]
elif "params-ema" in state_dict_keys:
state_dict = state_dict["params-ema"]
elif "params" in state_dict_keys:
state_dict = state_dict["params"]
state_dict_keys = list(state_dict.keys())
# SRVGGNet Real-ESRGAN (v2)
if "body.0.weight" in state_dict_keys and "body.1.weight" in state_dict_keys:
model = RealESRGANv2(state_dict)
# SPSR (ESRGAN with lots of extra layers)
elif "f_HR_conv1.0.weight" in state_dict:
model = SPSR(state_dict)
# Swift-SRGAN
elif (
"model" in state_dict_keys
and "initial.cnn.depthwise.weight" in state_dict["model"].keys()
):
model = SwiftSRGAN(state_dict)
# SwinIR, Swin2SR, HAT
elif "layers.0.residual_group.blocks.0.norm1.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)
else:
model = SwinIR(state_dict)
# GFPGAN
elif (
"toRGB.0.weight" in state_dict_keys
and "stylegan_decoder.style_mlp.1.weight" in state_dict_keys
):
model = GFPGANv1Clean(state_dict)
# RestoreFormer
elif (
"encoder.conv_in.weight" in state_dict_keys
and "encoder.down.0.block.0.norm1.weight" in state_dict_keys
):
model = RestoreFormer(state_dict)
elif (
"encoder.blocks.0.weight" in state_dict_keys
and "quantize.embedding.weight" in state_dict_keys
):
model = CodeFormer(state_dict)
# LaMa
elif (
"model.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)
# Omni-SR
elif "residual_layer.0.residual_layer.0.layer.0.fn.0.weight" in state_dict_keys:
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
else:
try:
model = ESRGAN(state_dict)
except:
# pylint: disable=raise-missing-from
raise UnsupportedModel
return model
from typing import Union
from .architecture.DAT import DAT
from .architecture.face.codeformer import CodeFormer
from .architecture.face.gfpganv1_clean_arch import GFPGANv1Clean
from .architecture.face.restoreformer_arch import RestoreFormer
from .architecture.HAT import HAT
from .architecture.LaMa import LaMa
from .architecture.OmniSR.OmniSR import OmniSR
from .architecture.RRDB import RRDBNet as ESRGAN
from .architecture.SCUNet import SCUNet
from .architecture.SPSR import SPSRNet as SPSR
from .architecture.SRVGG import SRVGGNetCompact as RealESRGANv2
from .architecture.SwiftSRGAN import Generator as SwiftSRGAN
from .architecture.Swin2SR import Swin2SR
from .architecture.SwinIR import SwinIR
PyTorchSRModels = (
RealESRGANv2,
SPSR,
SwiftSRGAN,
ESRGAN,
SwinIR,
Swin2SR,
HAT,
OmniSR,
SCUNet,
DAT,
)
PyTorchSRModel = Union[
RealESRGANv2,
SPSR,
SwiftSRGAN,
ESRGAN,
SwinIR,
Swin2SR,
HAT,
OmniSR,
SCUNet,
DAT,
]
def is_pytorch_sr_model(model: object):
return isinstance(model, PyTorchSRModels)
PyTorchFaceModels = (GFPGANv1Clean, RestoreFormer, CodeFormer)
PyTorchFaceModel = Union[GFPGANv1Clean, RestoreFormer, CodeFormer]
def is_pytorch_face_model(model: object):
return isinstance(model, PyTorchFaceModels)
PyTorchInpaintModels = (LaMa,)
PyTorchInpaintModel = Union[LaMa]
def is_pytorch_inpaint_model(model: object):
return isinstance(model, PyTorchInpaintModels)
PyTorchModels = (*PyTorchSRModels, *PyTorchFaceModels, *PyTorchInpaintModels)
PyTorchModel = Union[PyTorchSRModel, PyTorchFaceModel, PyTorchInpaintModel]
def is_pytorch_model(model: object):
return isinstance(model, PyTorchModels)
import os import os
from comfy_extras.chainner_models import model_loading from spandrel import ModelLoader, ImageModelDescriptor
from comfy import model_management from comfy import model_management
import torch import torch
import comfy.utils import comfy.utils
...@@ -20,7 +20,11 @@ class UpscaleModelLoader: ...@@ -20,7 +20,11 @@ class UpscaleModelLoader:
sd = comfy.utils.load_torch_file(model_path, safe_load=True) sd = comfy.utils.load_torch_file(model_path, safe_load=True)
if "module.layers.0.residual_group.blocks.0.norm1.weight" in sd: if "module.layers.0.residual_group.blocks.0.norm1.weight" in sd:
sd = comfy.utils.state_dict_prefix_replace(sd, {"module.":""}) sd = comfy.utils.state_dict_prefix_replace(sd, {"module.":""})
out = model_loading.load_state_dict(sd).eval() out = ModelLoader().load_from_state_dict(sd).eval()
if not isinstance(out, ImageModelDescriptor):
raise Exception("Upscale model must be a single-image model.")
return (out, ) return (out, )
...@@ -61,7 +65,7 @@ class ImageUpscaleWithModel: ...@@ -61,7 +65,7 @@ class ImageUpscaleWithModel:
if tile < 128: if tile < 128:
raise e raise e
upscale_model.cpu() upscale_model.to("cpu")
s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0) s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0)
return (s,) return (s,)
......
...@@ -11,3 +11,4 @@ scipy ...@@ -11,3 +11,4 @@ scipy
tqdm tqdm
psutil psutil
kornia>=0.7.1 kornia>=0.7.1
spandrel==0.3.1
\ No newline at end of file
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