praeclarumjj3's picture
:zap: Build App
9eae6e7
# Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
import numpy as np
from numpy.lib.type_check import imag
import torch
import torch.nn as nn
from torch_utils import misc
from torch_utils import persistence
from torch_utils.ops import conv2d_resample
from torch_utils.ops import upfirdn2d
from torch_utils.ops import bias_act
from torch_utils.ops import fma
from icecream import ic
import torch.nn.functional as F
from training.ffc import FFCResnetBlock, ConcatTupleLayer
import matplotlib.pyplot as plt
import PIL
#----------------------------------------------------------------------------
@misc.profiled_function
def normalize_2nd_moment(x, dim=1, eps=1e-8):
return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
def save_image_grid(feats, fname, gridsize):
gw, gh = gridsize
idx = gw * gh
max_num = torch.max(feats[:idx]).item()
min_num = torch.min(feats[:idx]).item()
feats = feats[:idx].cpu() * 255 / (max_num - min_num)
feats = np.asarray(feats, dtype=np.float32)
feats = np.rint(feats).clip(0, 255).astype(np.uint8)
C, H, W = feats.shape
feats = feats.reshape(gh, gw, 1, H, W)
feats = feats.transpose(0, 3, 1, 4, 2)
feats = feats.reshape(gh * H, gw * W, 1)
feats = np.stack([feats]*3, axis=2).squeeze() * 10
feats = np.rint(feats).clip(0, 255).astype(np.uint8)
from icecream import ic
ic(feats.shape)
feats = PIL.Image.fromarray(feats)
feats.save(fname + '.png')
#----------------------------------------------------------------------------
@misc.profiled_function
def modulated_conv2d(
x, # Input tensor of shape [batch_size, in_channels, in_height, in_width].
weight, # Weight tensor of shape [out_channels, in_channels, kernel_height, kernel_width].
styles, # Modulation coefficients of shape [batch_size, in_channels].
noise = None, # Optional noise tensor to add to the output activations.
up = 1, # Integer upsampling factor.
down = 1, # Integer downsampling factor.
padding = 0, # Padding with respect to the upsampled image.
resample_filter = None, # Low-pass filter to apply when resampling activations. Must be prepared beforehand by calling upfirdn2d.setup_filter().
demodulate = True, # Apply weight demodulation?
flip_weight = True, # False = convolution, True = correlation (matches torch.nn.functional.conv2d).
fused_modconv = True, # Perform modulation, convolution, and demodulation as a single fused operation?
):
batch_size = x.shape[0]
out_channels, in_channels, kh, kw = weight.shape
misc.assert_shape(weight, [out_channels, in_channels, kh, kw]) # [OIkk]
misc.assert_shape(x, [batch_size, in_channels, None, None]) # [NIHW]
misc.assert_shape(styles, [batch_size, in_channels]) # [NI]
# Pre-normalize inputs to avoid FP16 overflow.
if x.dtype == torch.float16 and demodulate:
weight = weight * (1 / np.sqrt(in_channels * kh * kw) / weight.norm(float('inf'), dim=[1,2,3], keepdim=True)) # max_Ikk
styles = styles / styles.norm(float('inf'), dim=1, keepdim=True) # max_I
# Calculate per-sample weights and demodulation coefficients.
w = None
dcoefs = None
if demodulate or fused_modconv:
w = weight.unsqueeze(0) # [NOIkk]
w = w * styles.reshape(batch_size, 1, -1, 1, 1) # [NOIkk]
if demodulate:
dcoefs = (w.square().sum(dim=[2,3,4]) + 1e-8).rsqrt() # [NO]
if demodulate and fused_modconv:
w = w * dcoefs.reshape(batch_size, -1, 1, 1, 1) # [NOIkk]
# Execute by scaling the activations before and after the convolution.
if not fused_modconv:
x = x * styles.to(x.dtype).reshape(batch_size, -1, 1, 1)
x = conv2d_resample.conv2d_resample(x=x, w=weight.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, flip_weight=flip_weight)
if demodulate and noise is not None:
x = fma.fma(x, dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1), noise.to(x.dtype))
elif demodulate:
x = x * dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1)
elif noise is not None:
x = x.add_(noise.to(x.dtype))
return x
# Execute as one fused op using grouped convolution.
with misc.suppress_tracer_warnings(): # this value will be treated as a constant
batch_size = int(batch_size)
misc.assert_shape(x, [batch_size, in_channels, None, None])
x = x.reshape(1, -1, *x.shape[2:])
w = w.reshape(-1, in_channels, kh, kw)
x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=resample_filter, up=up, down=down, padding=padding, groups=batch_size, flip_weight=flip_weight)
x = x.reshape(batch_size, -1, *x.shape[2:])
if noise is not None:
x = x.add_(noise)
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class FullyConnectedLayer(torch.nn.Module):
def __init__(self,
in_features, # Number of input features.
out_features, # Number of output features.
bias = True, # Apply additive bias before the activation function?
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
lr_multiplier = 1, # Learning rate multiplier.
bias_init = 0, # Initial value for the additive bias.
):
super().__init__()
self.activation = activation
self.weight = torch.nn.Parameter(torch.randn([out_features, in_features]) / lr_multiplier)
self.bias = torch.nn.Parameter(torch.full([out_features], np.float32(bias_init))) if bias else None
self.weight_gain = lr_multiplier / np.sqrt(in_features)
self.bias_gain = lr_multiplier
def forward(self, x):
w = self.weight.to(x.dtype) * self.weight_gain
b = self.bias
if b is not None:
b = b.to(x.dtype)
if self.bias_gain != 1:
b = b * self.bias_gain
if self.activation == 'linear' and b is not None:
x = torch.addmm(b.unsqueeze(0), x, w.t())
else:
x = x.matmul(w.t())
x = bias_act.bias_act(x, b, act=self.activation)
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class Conv2dLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
kernel_size, # Width and height of the convolution kernel.
bias = True, # Apply additive bias before the activation function?
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
up = 1, # Integer upsampling factor.
down = 1, # Integer downsampling factor.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output to +-X, None = disable clamping.
channels_last = False, # Expect the input to have memory_format=channels_last?
trainable = True, # Update the weights of this layer during training?
):
super().__init__()
self.activation = activation
self.up = up
self.down = down
self.conv_clamp = conv_clamp
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
self.act_gain = bias_act.activation_funcs[activation].def_gain
memory_format = torch.channels_last if channels_last else torch.contiguous_format
weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format)
bias = torch.zeros([out_channels]) if bias else None
if trainable:
self.weight = torch.nn.Parameter(weight)
self.bias = torch.nn.Parameter(bias) if bias is not None else None
else:
self.register_buffer('weight', weight)
if bias is not None:
self.register_buffer('bias', bias)
else:
self.bias = None
def forward(self, x, gain=1):
w = self.weight * self.weight_gain
b = self.bias.to(x.dtype) if self.bias is not None else None
flip_weight = (self.up == 1) # slightly faster
x = conv2d_resample.conv2d_resample(x=x, w=w.to(x.dtype), f=self.resample_filter, up=self.up, down=self.down, padding=self.padding, flip_weight=flip_weight)
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
x = bias_act.bias_act(x, b, act=self.activation, gain=act_gain, clamp=act_clamp)
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class FFCBlock(torch.nn.Module):
def __init__(self,
dim, # Number of output/input channels.
kernel_size, # Width and height of the convolution kernel.
padding,
ratio_gin=0.75,
ratio_gout=0.75,
activation = 'linear', # Activation function: 'relu', 'lrelu', etc.
):
super().__init__()
if activation == 'linear':
self.activation = nn.Identity
else:
self.activation = nn.ReLU
self.padding = padding
self.kernel_size = kernel_size
self.ffc_block = FFCResnetBlock(dim=dim,
padding_type='reflect',
norm_layer=nn.SyncBatchNorm,
activation_layer=self.activation,
dilation=1,
ratio_gin=ratio_gin,
ratio_gout=ratio_gout)
self.concat_layer = ConcatTupleLayer()
def forward(self, gen_ft, mask, fname=None):
x = gen_ft.float()
# x = mask*enc_ft + (1-mask)*gen_ft
x_l, x_g = x[:, :-self.ffc_block.conv1.ffc.global_in_num], x[:, -self.ffc_block.conv1.ffc.global_in_num:]
id_l, id_g = x_l, x_g
x_l, x_g = self.ffc_block((x_l, x_g), fname=fname)
x_l, x_g = id_l + x_l, id_g + x_g
x = self.concat_layer((x_l, x_g))
return x + gen_ft.float()
#----------------------------------------------------------------------------
@persistence.persistent_class
class EncoderEpilogue(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
cmap_dim, # Dimensionality of mapped conditioning label, 0 = no label.
z_dim, # Output Latent (Z) dimensionality.
resolution, # Resolution of this block.
img_channels, # Number of input color channels.
architecture = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
mbstd_group_size = 4, # Group size for the minibatch standard deviation layer, None = entire minibatch.
mbstd_num_channels = 1, # Number of features for the minibatch standard deviation layer, 0 = disable.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.cmap_dim = cmap_dim
self.resolution = resolution
self.img_channels = img_channels
self.architecture = architecture
if architecture == 'skip':
self.fromrgb = Conv2dLayer(self.img_channels, in_channels, kernel_size=1, activation=activation)
self.mbstd = MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels) if mbstd_num_channels > 0 else None
self.conv = Conv2dLayer(in_channels + mbstd_num_channels, in_channels, kernel_size=3, activation=activation, conv_clamp=conv_clamp)
self.fc = FullyConnectedLayer(in_channels * (resolution ** 2), z_dim, activation=activation)
# self.out = FullyConnectedLayer(in_channels, z_dim)
self.dropout = torch.nn.Dropout(p=0.5)
def forward(self, x, cmap, force_fp32=False):
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution]) # [NCHW]
_ = force_fp32 # unused
dtype = torch.float32
memory_format = torch.contiguous_format
# FromRGB.
x = x.to(dtype=dtype, memory_format=memory_format)
# Main layers.
if self.mbstd is not None:
x = self.mbstd(x)
const_e = self.conv(x)
x = self.fc(const_e.flatten(1))
# x = self.out(x)
x = self.dropout(x)
# Conditioning.
if self.cmap_dim > 0:
misc.assert_shape(cmap, [None, self.cmap_dim])
x = (x * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
assert x.dtype == dtype
return x, const_e
#----------------------------------------------------------------------------
@persistence.persistent_class
class EncoderBlock(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels, 0 = first block.
tmp_channels, # Number of intermediate channels.
out_channels, # Number of output channels.
resolution, # Resolution of this block.
img_channels, # Number of input color channels.
first_layer_idx, # Index of the first layer.
architecture = 'skip', # Architecture: 'orig', 'skip', 'resnet'.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
use_fp16 = False, # Use FP16 for this block?
fp16_channels_last = False, # Use channels-last memory format with FP16?
freeze_layers = 0, # Freeze-D: Number of layers to freeze.
):
assert in_channels in [0, tmp_channels]
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.resolution = resolution
self.img_channels = img_channels + 1
self.first_layer_idx = first_layer_idx
self.architecture = architecture
self.use_fp16 = use_fp16
self.channels_last = (use_fp16 and fp16_channels_last)
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.num_layers = 0
def trainable_gen():
while True:
layer_idx = self.first_layer_idx + self.num_layers
trainable = (layer_idx >= freeze_layers)
self.num_layers += 1
yield trainable
trainable_iter = trainable_gen()
if in_channels == 0:
self.fromrgb = Conv2dLayer(self.img_channels, tmp_channels, kernel_size=1, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv0 = Conv2dLayer(tmp_channels, tmp_channels, kernel_size=3, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv1 = Conv2dLayer(tmp_channels, out_channels, kernel_size=3, activation=activation, down=2,
trainable=next(trainable_iter), resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last)
if architecture == 'resnet':
self.skip = Conv2dLayer(tmp_channels, out_channels, kernel_size=1, bias=False, down=2,
trainable=next(trainable_iter), resample_filter=resample_filter, channels_last=self.channels_last)
def forward(self, x, img, force_fp32=False):
dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
# Input.
if x is not None:
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution])
x = x.to(dtype=dtype, memory_format=memory_format)
# FromRGB.
if self.in_channels == 0:
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
img = img.to(dtype=dtype, memory_format=memory_format)
y = self.fromrgb(img)
x = x + y if x is not None else y
img = upfirdn2d.downsample2d(img, self.resample_filter) if self.architecture == 'skip' else None
# Main layers.
if self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x)
feat = x.clone()
x = self.conv1(x, gain=np.sqrt(0.5))
x = y.add_(x)
else:
x = self.conv0(x)
feat = x.clone()
x = self.conv1(x)
assert x.dtype == dtype
return x, img, feat
#----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
w_dim, # Intermediate latent (W) dimensionality.
resolution, # Resolution of this layer.
kernel_size = 3, # Convolution kernel size.
up = 1, # Integer upsampling factor.
use_noise = True, # Enable noise input?
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
channels_last = False, # Use channels_last format for the weights?
):
super().__init__()
self.resolution = resolution
self.up = up
self.use_noise = use_noise
self.activation = activation
self.conv_clamp = conv_clamp
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.act_gain = bias_act.activation_funcs[activation].def_gain
self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
memory_format = torch.channels_last if channels_last else torch.contiguous_format
self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
if use_noise:
self.register_buffer('noise_const', torch.randn([resolution, resolution]))
self.noise_strength = torch.nn.Parameter(torch.zeros([]))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
def forward(self, x, w, noise_mode='random', fused_modconv=True, gain=1):
assert noise_mode in ['random', 'const', 'none']
in_resolution = self.resolution // self.up
misc.assert_shape(x, [None, self.weight.shape[1], in_resolution, in_resolution])
styles = self.affine(w)
noise = None
if self.use_noise and noise_mode == 'random':
noise = torch.randn([x.shape[0], 1, self.resolution, self.resolution], device=x.device) * self.noise_strength
if self.use_noise and noise_mode == 'const':
noise = self.noise_const * self.noise_strength
flip_weight = (self.up == 1) # slightly faster
x = modulated_conv2d(x=x, weight=self.weight, styles=styles, noise=noise, up=self.up,
padding=self.padding, resample_filter=self.resample_filter, flip_weight=flip_weight, fused_modconv=fused_modconv)
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
x = F.leaky_relu(x, negative_slope=0.2, inplace=False)
if act_gain != 1:
x = x * act_gain
if act_clamp is not None:
x = x.clamp(-act_clamp, act_clamp)
# x = bias_act.bias_act(x.clone(), self.bias.to(x.dtype), act=self.activation, gain=act_gain, clamp=act_clamp)
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class FFCSkipLayer(torch.nn.Module):
def __init__(self,
dim, # Number of input/output channels.
kernel_size = 3, # Convolution kernel size.
ratio_gin=0.75,
ratio_gout=0.75,
):
super().__init__()
self.padding = kernel_size // 2
self.ffc_act = FFCBlock(dim=dim, kernel_size=kernel_size, activation=nn.ReLU,
padding=self.padding, ratio_gin=ratio_gin, ratio_gout=ratio_gout)
def forward(self, gen_ft, mask, fname=None):
x = self.ffc_act(gen_ft, mask, fname=fname)
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class ToRGBLayer(torch.nn.Module):
def __init__(self, in_channels, out_channels, w_dim, kernel_size=1, conv_clamp=None, channels_last=False):
super().__init__()
self.conv_clamp = conv_clamp
self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
memory_format = torch.channels_last if channels_last else torch.contiguous_format
self.weight = torch.nn.Parameter(torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
def forward(self, x, w, fused_modconv=True):
styles = self.affine(w) * self.weight_gain
x = modulated_conv2d(x=x, weight=self.weight, styles=styles, demodulate=False, fused_modconv=fused_modconv)
x = bias_act.bias_act(x, self.bias.to(x.dtype), clamp=self.conv_clamp)
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisBlock(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels, 0 = first block.
out_channels, # Number of output channels.
w_dim, # Intermediate latent (W) dimensionality.
resolution, # Resolution of this block.
img_channels, # Number of output color channels.
is_last, # Is this the last block?
architecture = 'skip', # Architecture: 'orig', 'skip', 'resnet'.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
use_fp16 = False, # Use FP16 for this block?
fp16_channels_last = False, # Use channels-last memory format with FP16?
**layer_kwargs, # Arguments for SynthesisLayer.
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.w_dim = w_dim
self.resolution = resolution
self.img_channels = img_channels
self.is_last = is_last
self.architecture = architecture
self.use_fp16 = use_fp16
self.channels_last = (use_fp16 and fp16_channels_last)
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.num_conv = 0
self.num_torgb = 0
self.res_ffc = {4:0, 8: 0, 16: 0, 32: 1, 64: 1, 128: 1, 256: 1, 512: 1}
if in_channels != 0 and resolution >= 8:
self.ffc_skip = nn.ModuleList()
for _ in range(self.res_ffc[resolution]):
self.ffc_skip.append(FFCSkipLayer(dim=out_channels))
if in_channels == 0:
self.const = torch.nn.Parameter(torch.randn([out_channels, resolution, resolution]))
if in_channels != 0:
self.conv0 = SynthesisLayer(in_channels, out_channels, w_dim=w_dim*3, resolution=resolution, up=2,
resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last, **layer_kwargs)
self.num_conv += 1
self.conv1 = SynthesisLayer(out_channels, out_channels, w_dim=w_dim*3, resolution=resolution,
conv_clamp=conv_clamp, channels_last=self.channels_last, **layer_kwargs)
self.num_conv += 1
if is_last or architecture == 'skip':
self.torgb = ToRGBLayer(out_channels, img_channels, w_dim=w_dim*3,
conv_clamp=conv_clamp, channels_last=self.channels_last)
self.num_torgb += 1
if in_channels != 0 and architecture == 'resnet':
self.skip = Conv2dLayer(in_channels, out_channels, kernel_size=1, bias=False, up=2,
resample_filter=resample_filter, channels_last=self.channels_last)
def forward(self, x, mask, feats, img, ws, fname=None, force_fp32=False, fused_modconv=None, **layer_kwargs):
# misc.assert_shape(ws, [None, self.num_conv + self.num_torgb, self.w_dim])
# w_iter = iter(ws.unbind(dim=1))
dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
if fused_modconv is None:
with misc.suppress_tracer_warnings(): # this value will be treated as a constant
fused_modconv = (not self.training) and (dtype == torch.float32 or int(x.shape[0]) == 1)
# # Input.
# if self.in_channels == 0:
# ic(self.const.shape)
# x = self.const.to(dtype=dtype, memory_format=memory_format)
# x = x.unsqueeze(0).repeat([ws.shape[0], 1, 1, 1])
# ic(x.shape)
# else:
# misc.assert_shape(x, [None, self.in_channels, self.resolution // 2, self.resolution // 2])
# x = x.to(dtype=dtype, memory_format=memory_format)
# ic(x.shape, 'ELSE')
x = x.to(dtype=dtype, memory_format=memory_format)
x_skip = feats[self.resolution].clone().to(dtype=dtype, memory_format=memory_format)
# Main layers.
if self.in_channels == 0:
x = self.conv1(x, ws[1], fused_modconv=fused_modconv, **layer_kwargs)
elif self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x, ws[0].clone(), fused_modconv=fused_modconv, **layer_kwargs)
if len(self.ffc_skip) > 0:
mask = F.interpolate(mask, size=x_skip.shape[2:],)
z = x + x_skip
for fres in self.ffc_skip:
z = fres(z, mask)
x = x + z
else:
x = x + x_skip
x = self.conv1(x, ws[1].clone(), fused_modconv=fused_modconv, gain=np.sqrt(0.5), **layer_kwargs)
x = y.add_(x)
else:
x = self.conv0(x, ws[0].clone(), fused_modconv=fused_modconv, **layer_kwargs)
if len(self.ffc_skip) > 0:
# for i in range(x.shape[0]):
# c, h, w = x[i].shape
# gh = 3
# gw = 3
# save_image_grid(x[i].detach(), f'vis/{fname}_pre_{h}', (gh, gw))
mask = F.interpolate(mask, size=x_skip.shape[2:],)
z = x + x_skip
for fres in self.ffc_skip:
z = fres(z, mask)
# for i in range(z.shape[0]):
# c, h, w = z[i].shape
# gh = 3
# gw = 3
# save_image_grid(z[i].detach(), f'vis/{fname}_ffc_{h}', (gh, gw))
x = x + z
# for i in range(x.shape[0]):
# c, h, w = x[i].shape
# gh = 3
# gw = 3
# save_image_grid(x[i].detach(), f'vis/{fname}_post_{h}', (gh, gw))
else:
x = x + x_skip
x = self.conv1(x, ws[1].clone(), fused_modconv=fused_modconv, **layer_kwargs)
# ToRGB.
if img is not None:
misc.assert_shape(img, [None, self.img_channels, self.resolution // 2, self.resolution // 2])
img = upfirdn2d.upsample2d(img, self.resample_filter)
if self.is_last or self.architecture == 'skip':
y = self.torgb(x, ws[2].clone(), fused_modconv=fused_modconv)
y = y.to(dtype=torch.float32, memory_format=torch.contiguous_format)
img = img.add_(y) if img is not None else y
x = x.to(dtype=dtype)
assert x.dtype == dtype
assert img is None or img.dtype == torch.float32
return x, img
#----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisForeword(torch.nn.Module):
def __init__(self,
z_dim, # Output Latent (Z) dimensionality.
resolution, # Resolution of this block.
in_channels,
img_channels, # Number of input color channels.
architecture = 'skip', # Architecture: 'orig', 'skip', 'resnet'.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
):
super().__init__()
self.in_channels = in_channels
self.z_dim = z_dim
self.resolution = resolution
self.img_channels = img_channels
self.architecture = architecture
self.fc = FullyConnectedLayer(self.z_dim, (self.z_dim // 2) * 4 * 4, activation=activation)
self.conv = SynthesisLayer(self.in_channels, self.in_channels, w_dim=(z_dim // 2) * 3, resolution=4)
if architecture == 'skip':
self.torgb = ToRGBLayer(self.in_channels, self.img_channels, kernel_size=1, w_dim = (z_dim // 2) * 3)
def forward(self, x, ws, feats, img, force_fp32=False):
misc.assert_shape(x, [None, self.z_dim]) # [NC]
_ = force_fp32 # unused
dtype = torch.float32
memory_format = torch.contiguous_format
x_global = x.clone()
# ToRGB.
x = self.fc(x)
x = x.view(-1, self.z_dim // 2, 4, 4)
x = x.to(dtype=dtype, memory_format=memory_format)
# Main layers.
x_skip = feats[4].clone()
x = x + x_skip
mod_vector = []
mod_vector.append(ws[:, 0])
mod_vector.append(x_global.clone())
mod_vector = torch.cat(mod_vector, dim = 1)
x = self.conv(x, mod_vector)
mod_vector = []
mod_vector.append(ws[:, 2*2-3])
mod_vector.append(x_global.clone())
mod_vector = torch.cat(mod_vector, dim = 1)
if self.architecture == 'skip':
img = self.torgb(x, mod_vector)
img = img.to(dtype=torch.float32, memory_format=torch.contiguous_format)
assert x.dtype == dtype
return x, img
#----------------------------------------------------------------------------
@persistence.persistent_class
class DiscriminatorBlock(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels, 0 = first block.
tmp_channels, # Number of intermediate channels.
out_channels, # Number of output channels.
resolution, # Resolution of this block.
img_channels, # Number of input color channels.
first_layer_idx, # Index of the first layer.
architecture = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
resample_filter = [1,3,3,1], # Low-pass filter to apply when resampling activations.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
use_fp16 = False, # Use FP16 for this block?
fp16_channels_last = False, # Use channels-last memory format with FP16?
freeze_layers = 0, # Freeze-D: Number of layers to freeze.
):
assert in_channels in [0, tmp_channels]
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.resolution = resolution
self.img_channels = img_channels + 1
self.first_layer_idx = first_layer_idx
self.architecture = architecture
self.use_fp16 = use_fp16
self.channels_last = (use_fp16 and fp16_channels_last)
self.register_buffer('resample_filter', upfirdn2d.setup_filter(resample_filter))
self.num_layers = 0
def trainable_gen():
while True:
layer_idx = self.first_layer_idx + self.num_layers
trainable = (layer_idx >= freeze_layers)
self.num_layers += 1
yield trainable
trainable_iter = trainable_gen()
if in_channels == 0 or architecture == 'skip':
self.fromrgb = Conv2dLayer(self.img_channels, tmp_channels, kernel_size=1, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv0 = Conv2dLayer(tmp_channels, tmp_channels, kernel_size=3, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv1 = Conv2dLayer(tmp_channels, out_channels, kernel_size=3, activation=activation, down=2,
trainable=next(trainable_iter), resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last)
if architecture == 'resnet':
self.skip = Conv2dLayer(tmp_channels, out_channels, kernel_size=1, bias=False, down=2,
trainable=next(trainable_iter), resample_filter=resample_filter, channels_last=self.channels_last)
def forward(self, x, img, force_fp32=False):
dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
# Input.
if x is not None:
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution])
x = x.to(dtype=dtype, memory_format=memory_format)
# FromRGB.
if self.in_channels == 0 or self.architecture == 'skip':
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
img = img.to(dtype=dtype, memory_format=memory_format)
y = self.fromrgb(img)
x = x + y if x is not None else y
img = upfirdn2d.downsample2d(img, self.resample_filter) if self.architecture == 'skip' else None
# Main layers.
if self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x)
x = self.conv1(x, gain=np.sqrt(0.5))
x = y.add_(x)
else:
x = self.conv0(x)
x = self.conv1(x)
assert x.dtype == dtype
return x, img
#----------------------------------------------------------------------------
@persistence.persistent_class
class MinibatchStdLayer(torch.nn.Module):
def __init__(self, group_size, num_channels=1):
super().__init__()
self.group_size = group_size
self.num_channels = num_channels
def forward(self, x):
N, C, H, W = x.shape
with misc.suppress_tracer_warnings(): # as_tensor results are registered as constants
G = torch.min(torch.as_tensor(self.group_size), torch.as_tensor(N)) if self.group_size is not None else N
F = self.num_channels
c = C // F
y = x.reshape(G, -1, F, c, H, W) # [GnFcHW] Split minibatch N into n groups of size G, and channels C into F groups of size c.
y = y - y.mean(dim=0) # [GnFcHW] Subtract mean over group.
y = y.square().mean(dim=0) # [nFcHW] Calc variance over group.
y = (y + 1e-8).sqrt() # [nFcHW] Calc stddev over group.
y = y.mean(dim=[2,3,4]) # [nF] Take average over channels and pixels.
y = y.reshape(-1, F, 1, 1) # [nF11] Add missing dimensions.
y = y.repeat(G, 1, H, W) # [NFHW] Replicate over group and pixels.
x = torch.cat([x, y], dim=1) # [NCHW] Append to input as new channels.
return x
#----------------------------------------------------------------------------
@persistence.persistent_class
class DiscriminatorEpilogue(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
cmap_dim, # Dimensionality of mapped conditioning label, 0 = no label.
resolution, # Resolution of this block.
img_channels, # Number of input color channels.
architecture = 'resnet', # Architecture: 'orig', 'skip', 'resnet'.
mbstd_group_size = 4, # Group size for the minibatch standard deviation layer, None = entire minibatch.
mbstd_num_channels = 1, # Number of features for the minibatch standard deviation layer, 0 = disable.
activation = 'lrelu', # Activation function: 'relu', 'lrelu', etc.
conv_clamp = None, # Clamp the output of convolution layers to +-X, None = disable clamping.
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.cmap_dim = cmap_dim
self.resolution = resolution
self.img_channels = img_channels
self.architecture = architecture
if architecture == 'skip':
self.fromrgb = Conv2dLayer(img_channels, in_channels, kernel_size=1, activation=activation)
self.mbstd = MinibatchStdLayer(group_size=mbstd_group_size, num_channels=mbstd_num_channels) if mbstd_num_channels > 0 else None
self.conv = Conv2dLayer(in_channels + mbstd_num_channels, in_channels, kernel_size=3, activation=activation, conv_clamp=conv_clamp)
self.fc = FullyConnectedLayer(in_channels * (resolution ** 2), in_channels, activation=activation)
self.out = FullyConnectedLayer(in_channels, 1 if cmap_dim == 0 else cmap_dim)
def forward(self, x, img, cmap, force_fp32=False):
misc.assert_shape(x, [None, self.in_channels, self.resolution, self.resolution]) # [NCHW]
_ = force_fp32 # unused
dtype = torch.float32
memory_format = torch.contiguous_format
# FromRGB.
x = x.to(dtype=dtype, memory_format=memory_format)
if self.architecture == 'skip':
misc.assert_shape(img, [None, self.img_channels, self.resolution, self.resolution])
img = img.to(dtype=dtype, memory_format=memory_format)
x = x + self.fromrgb(img)
# Main layers.
if self.mbstd is not None:
x = self.mbstd(x)
x = self.conv(x)
x = self.fc(x.flatten(1))
x = self.out(x)
# Conditioning.
if self.cmap_dim > 0:
misc.assert_shape(cmap, [None, self.cmap_dim])
x = (x * cmap).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.cmap_dim))
assert x.dtype == dtype
return x
#----------------------------------------------------------------------------