244 lines
8.1 KiB
Python
244 lines
8.1 KiB
Python
from functools import partial
|
|
from math import log2
|
|
from random import random
|
|
import numpy as np
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
import models.archs.stylegan.stylegan2 as sg2
|
|
import models.steps.losses as L
|
|
|
|
|
|
def leaky_relu(p=0.2):
|
|
return nn.LeakyReLU(p)
|
|
|
|
|
|
def double_conv(chan_in, chan_out):
|
|
return nn.Sequential(
|
|
nn.Conv2d(chan_in, chan_out, 3, padding=1),
|
|
leaky_relu(),
|
|
nn.Conv2d(chan_out, chan_out, 3, padding=1),
|
|
leaky_relu()
|
|
)
|
|
|
|
|
|
class Flatten(nn.Module):
|
|
def __init__(self, index):
|
|
super().__init__()
|
|
self.index = index
|
|
def forward(self, x):
|
|
return x.flatten(self.index)
|
|
|
|
|
|
class DownBlock(nn.Module):
|
|
def __init__(self, input_channels, filters, downsample=True):
|
|
super().__init__()
|
|
self.conv_res = nn.Conv2d(input_channels, filters, 1, stride = (2 if downsample else 1))
|
|
|
|
self.net = double_conv(input_channels, filters)
|
|
self.down = nn.Conv2d(filters, filters, 3, padding = 1, stride = 2) if downsample else None
|
|
|
|
def forward(self, x):
|
|
res = self.conv_res(x)
|
|
x = self.net(x)
|
|
unet_res = x
|
|
|
|
if self.down is not None:
|
|
x = self.down(x)
|
|
|
|
x = x + res
|
|
return x, unet_res
|
|
|
|
|
|
class UpBlock(nn.Module):
|
|
def __init__(self, input_channels, filters):
|
|
super().__init__()
|
|
self.conv_res = nn.ConvTranspose2d(input_channels // 2, filters, 1, stride = 2)
|
|
self.net = double_conv(input_channels, filters)
|
|
self.up = nn.Upsample(scale_factor = 2, mode='bilinear', align_corners=False)
|
|
self.input_channels = input_channels
|
|
self.filters = filters
|
|
|
|
def forward(self, x, res):
|
|
*_, h, w = x.shape
|
|
conv_res = self.conv_res(x, output_size = (h * 2, w * 2))
|
|
x = self.up(x)
|
|
x = torch.cat((x, res), dim=1)
|
|
x = self.net(x)
|
|
x = x + conv_res
|
|
return x
|
|
|
|
|
|
class StyleGan2UnetDiscriminator(nn.Module):
|
|
def __init__(self, image_size, network_capacity = 16, fmap_max = 512, input_filters=3):
|
|
super().__init__()
|
|
num_layers = int(log2(image_size) - 3)
|
|
|
|
blocks = []
|
|
filters = [input_filters] + [(network_capacity) * (2 ** i) for i in range(num_layers + 1)]
|
|
|
|
set_fmap_max = partial(min, fmap_max)
|
|
filters = list(map(set_fmap_max, filters))
|
|
filters[-1] = filters[-2]
|
|
|
|
chan_in_out = list(zip(filters[:-1], filters[1:]))
|
|
chan_in_out = list(map(list, chan_in_out))
|
|
|
|
down_blocks = []
|
|
attn_blocks = []
|
|
|
|
for ind, (in_chan, out_chan) in enumerate(chan_in_out):
|
|
num_layer = ind + 1
|
|
is_not_last = ind != (len(chan_in_out) - 1)
|
|
|
|
block = DownBlock(in_chan, out_chan, downsample = is_not_last)
|
|
down_blocks.append(block)
|
|
|
|
attn_fn = sg2.attn_and_ff(out_chan)
|
|
attn_blocks.append(attn_fn)
|
|
|
|
self.down_blocks = nn.ModuleList(down_blocks)
|
|
self.attn_blocks = nn.ModuleList(attn_blocks)
|
|
|
|
last_chan = filters[-1]
|
|
|
|
self.to_logit = nn.Sequential(
|
|
leaky_relu(),
|
|
nn.AvgPool2d(image_size // (2 ** num_layers)),
|
|
Flatten(1),
|
|
nn.Linear(last_chan, 1)
|
|
)
|
|
|
|
self.conv = double_conv(last_chan, last_chan)
|
|
|
|
dec_chan_in_out = chan_in_out[:-1][::-1]
|
|
self.up_blocks = nn.ModuleList(list(map(lambda c: UpBlock(c[1] * 2, c[0]), dec_chan_in_out)))
|
|
self.conv_out = nn.Conv2d(input_filters, 1, 1)
|
|
|
|
def forward(self, x):
|
|
b, *_ = x.shape
|
|
|
|
residuals = []
|
|
|
|
for (down_block, attn_block) in zip(self.down_blocks, self.attn_blocks):
|
|
x, unet_res = down_block(x)
|
|
residuals.append(unet_res)
|
|
|
|
if attn_block is not None:
|
|
x = attn_block(x)
|
|
|
|
x = self.conv(x) + x
|
|
enc_out = self.to_logit(x)
|
|
|
|
for (up_block, res) in zip(self.up_blocks, residuals[:-1][::-1]):
|
|
x = up_block(x, res)
|
|
|
|
dec_out = self.conv_out(x)
|
|
return dec_out, enc_out
|
|
|
|
|
|
def warmup(start, end, max_steps, current_step):
|
|
if current_step > max_steps:
|
|
return end
|
|
return (end - start) * (current_step / max_steps) + start
|
|
|
|
|
|
def mask_src_tgt(source, target, mask):
|
|
return source * mask + (1 - mask) * target
|
|
|
|
|
|
def cutmix(source, target, coors, alpha = 1.):
|
|
source, target = map(torch.clone, (source, target))
|
|
((y0, y1), (x0, x1)), _ = coors
|
|
source[:, :, y0:y1, x0:x1] = target[:, :, y0:y1, x0:x1]
|
|
return source
|
|
|
|
|
|
def cutmix_coordinates(height, width, alpha = 1.):
|
|
lam = np.random.beta(alpha, alpha)
|
|
|
|
cx = np.random.uniform(0, width)
|
|
cy = np.random.uniform(0, height)
|
|
w = width * np.sqrt(1 - lam)
|
|
h = height * np.sqrt(1 - lam)
|
|
x0 = int(np.round(max(cx - w / 2, 0)))
|
|
x1 = int(np.round(min(cx + w / 2, width)))
|
|
y0 = int(np.round(max(cy - h / 2, 0)))
|
|
y1 = int(np.round(min(cy + h / 2, height)))
|
|
|
|
return ((y0, y1), (x0, x1)), lam
|
|
|
|
|
|
class StyleGan2UnetDivergenceLoss(L.ConfigurableLoss):
|
|
def __init__(self, opt, env):
|
|
super().__init__(opt, env)
|
|
self.real = opt['real']
|
|
self.fake = opt['fake']
|
|
self.discriminator = opt['discriminator']
|
|
self.for_gen = opt['gen_loss']
|
|
self.gp_frequency = opt['gradient_penalty_frequency']
|
|
self.noise = opt['noise'] if 'noise' in opt.keys() else 0
|
|
self.image_size = opt['image_size']
|
|
self.cr_weight = .2
|
|
|
|
def forward(self, net, state):
|
|
real_input = state[self.real]
|
|
fake_input = state[self.fake]
|
|
if self.noise != 0:
|
|
fake_input = fake_input + torch.rand_like(fake_input) * self.noise
|
|
real_input = real_input + torch.rand_like(real_input) * self.noise
|
|
|
|
D = self.env['discriminators'][self.discriminator]
|
|
fake_dec, fake_enc = D(fake_input)
|
|
fake_aug_images = D.module.aug_images
|
|
if self.for_gen:
|
|
return fake_enc.mean() + F.relu(1 + fake_dec).mean()
|
|
else:
|
|
dec_loss_coef = warmup(0, 1., 30000, self.env['step'])
|
|
cutmix_prob = warmup(0, 0.25, 30000, self.env['step'])
|
|
apply_cutmix = random() < cutmix_prob
|
|
|
|
real_input.requires_grad_() # <-- Needed to compute gradients on the input.
|
|
real_dec, real_enc = D(real_input)
|
|
real_aug_images = D.module.aug_images
|
|
enc_divergence = (F.relu(1 + real_enc) + F.relu(1 - fake_enc)).mean()
|
|
dec_divergence = (F.relu(1 + real_dec) + F.relu(1 - fake_dec)).mean()
|
|
disc_loss = enc_divergence + dec_divergence * dec_loss_coef
|
|
|
|
if apply_cutmix:
|
|
mask = cutmix(
|
|
torch.ones_like(real_dec),
|
|
torch.zeros_like(real_dec),
|
|
cutmix_coordinates(self.image_size, self.image_size)
|
|
)
|
|
|
|
if random() > 0.5:
|
|
mask = 1 - mask
|
|
|
|
cutmix_images = mask_src_tgt(real_aug_images, fake_aug_images, mask)
|
|
cutmix_dec_out, cutmix_enc_out = D.module.D(cutmix_images) # Bypass implied augmentor - hence D.module.D
|
|
|
|
cutmix_enc_divergence = F.relu(1 - cutmix_enc_out).mean()
|
|
cutmix_dec_divergence = F.relu(1 + (mask * 2 - 1) * cutmix_dec_out).mean()
|
|
disc_loss = disc_loss + cutmix_enc_divergence + cutmix_dec_divergence
|
|
|
|
cr_cutmix_dec_out = mask_src_tgt(real_dec, fake_dec, mask)
|
|
cr_loss = F.mse_loss(cutmix_dec_out, cr_cutmix_dec_out) * self.cr_weight
|
|
self.last_cr_loss = cr_loss.clone().detach().item()
|
|
|
|
disc_loss = disc_loss + cr_loss * dec_loss_coef
|
|
|
|
# Apply gradient penalty. TODO: migrate this elsewhere.
|
|
if self.env['step'] % self.gp_frequency == 0:
|
|
from models.archs.stylegan.stylegan2 import gradient_penalty
|
|
if random() < .5:
|
|
gp = gradient_penalty(real_input, real_enc)
|
|
else:
|
|
gp = gradient_penalty(real_input, real_dec) * dec_loss_coef
|
|
self.metrics.append(("gradient_penalty", gp.clone().detach()))
|
|
disc_loss = disc_loss + gp
|
|
|
|
real_input.requires_grad_(requires_grad=False)
|
|
return disc_loss
|