DL-Art-School/codes/models/networks.py

203 lines
11 KiB
Python
Raw Normal View History

2019-08-23 13:42:47 +00:00
import torch
2020-08-26 00:14:45 +00:00
import logging
2020-09-07 23:01:48 +00:00
from munch import munchify
2019-08-23 13:42:47 +00:00
import models.archs.SRResNet_arch as SRResNet_arch
import models.archs.discriminator_vgg_arch as SRGAN_arch
import models.archs.DiscriminatorResnet_arch as DiscriminatorResnet_arch
import models.archs.DiscriminatorResnet_arch_passthrough as DiscriminatorResnet_arch_passthrough
2019-08-23 13:42:47 +00:00
import models.archs.RRDBNet_arch as RRDBNet_arch
import models.archs.feature_arch as feature_arch
import models.archs.SwitchedResidualGenerator_arch as SwitchedGen_arch
import models.archs.SPSR_arch as spsr
import models.archs.StructuredSwitchedGenerator as ssg
from collections import OrderedDict
2019-08-23 13:42:47 +00:00
2020-08-26 00:14:45 +00:00
logger = logging.getLogger('base')
2019-08-23 13:42:47 +00:00
# Generator
2020-08-22 14:24:34 +00:00
def define_G(opt, net_key='network_G', scale=None):
if net_key is not None:
opt_net = opt[net_key]
else:
opt_net = opt
if scale is None:
scale = opt['scale']
2019-08-23 13:42:47 +00:00
which_model = opt_net['which_model_G']
# image restoration
if which_model == 'MSRResNet':
netG = SRResNet_arch.MSRResNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'],
nf=opt_net['nf'], nb=opt_net['nb'], upscale=opt_net['scale'])
elif which_model == 'RRDBNet':
2020-04-22 06:37:41 +00:00
# RRDB does scaling in two steps, so take the sqrt of the scale we actually want to achieve and feed it to RRDB.
initial_stride = 1 if 'initial_stride' not in opt_net else opt_net['initial_stride']
assert initial_stride == 1 or initial_stride == 2
# Need to adjust the scale the generator sees by the stride since the stride causes a down-sample.
gen_scale = scale * initial_stride
2019-08-23 13:42:47 +00:00
netG = RRDBNet_arch.RRDBNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'],
2020-09-20 03:47:10 +00:00
nf=opt_net['nf'], nb=opt_net['nb'], scale=opt_net['scale'] if 'scale' in opt_net.keys() else gen_scale,
initial_stride=initial_stride)
2020-06-25 01:49:37 +00:00
elif which_model == "ConfigurableSwitchedResidualGenerator2":
netG = SwitchedGen_arch.ConfigurableSwitchedResidualGenerator2(switch_depth=opt_net['switch_depth'], switch_filters=opt_net['switch_filters'],
2020-06-25 01:49:37 +00:00
switch_reductions=opt_net['switch_reductions'],
switch_processing_layers=opt_net['switch_processing_layers'], trans_counts=opt_net['trans_counts'],
trans_kernel_sizes=opt_net['trans_kernel_sizes'], trans_layers=opt_net['trans_layers'],
2020-07-18 13:24:02 +00:00
transformation_filters=opt_net['transformation_filters'], attention_norm=opt_net['attention_norm'],
2020-06-25 01:49:37 +00:00
initial_temp=opt_net['temperature'], final_temperature_step=opt_net['temperature_final_step'],
heightened_temp_min=opt_net['heightened_temp_min'], heightened_final_step=opt_net['heightened_final_step'],
upsample_factor=scale, add_scalable_noise_to_transforms=opt_net['add_noise'])
elif which_model == 'spsr_net_improved':
netG = spsr.SPSRNetSimplified(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], nf=opt_net['nf'],
nb=opt_net['nb'], upscale=opt_net['scale'])
elif which_model == "spsr_switched":
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
netG = spsr.SwitchedSpsr(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
elif which_model == "spsr_switched_with_ref2" or which_model == "spsr3":
2020-08-29 15:27:18 +00:00
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
2020-09-07 23:01:48 +00:00
netG = spsr.SwitchedSpsrWithRef2(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
2020-08-29 15:27:18 +00:00
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
elif which_model == "spsr4":
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
netG = spsr.Spsr4(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
elif which_model == "spsr5":
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
netG = spsr.Spsr5(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
elif which_model == "ssgr1":
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
netG = ssg.SSGr1(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
2020-09-20 03:47:10 +00:00
elif which_model == 'ssg_no_embedding':
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
netG = ssg.SSGNoEmbedding(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
elif which_model == 'ssg_lite':
xforms = opt_net['num_transforms'] if 'num_transforms' in opt_net.keys() else 8
netG = ssg.SSGLite(in_nc=3, out_nc=3, nf=opt_net['nf'], xforms=xforms, upscale=opt_net['scale'],
init_temperature=opt_net['temperature'] if 'temperature' in opt_net.keys() else 10)
elif which_model == "backbone_encoder":
netG = SwitchedGen_arch.BackboneEncoder(pretrained_backbone=opt_net['pretrained_spinenet'])
2020-09-15 22:55:38 +00:00
elif which_model == "backbone_encoder_no_ref":
netG = SwitchedGen_arch.BackboneEncoderNoRef(pretrained_backbone=opt_net['pretrained_spinenet'])
2020-09-21 18:36:49 +00:00
elif which_model == "backbone_resnet":
netG = SwitchedGen_arch.BackboneResnet()
2019-08-23 13:42:47 +00:00
else:
raise NotImplementedError('Generator model [{:s}] not recognized'.format(which_model))
return netG
2020-08-25 23:58:20 +00:00
class GradDiscWrapper(torch.nn.Module):
def __init__(self, m):
super(GradDiscWrapper, self).__init__()
2020-08-26 00:14:45 +00:00
logger.info("Wrapping a discriminator..")
2020-08-25 23:58:20 +00:00
self.m = m
2020-08-26 00:14:45 +00:00
def forward(self, x):
return self.m(x)
2020-08-25 23:58:20 +00:00
def define_D_net(opt_net, img_sz=None, wrap=False):
2019-08-23 13:42:47 +00:00
which_model = opt_net['which_model_D']
2020-09-20 03:47:10 +00:00
if 'image_size' in opt_net.keys():
img_sz = opt_net['image_size']
2019-08-23 13:42:47 +00:00
if which_model == 'discriminator_vgg_128':
2020-08-31 15:50:30 +00:00
netD = SRGAN_arch.Discriminator_VGG_128(in_nc=opt_net['in_nc'], nf=opt_net['nf'], input_img_factor=img_sz / 128, extra_conv=opt_net['extra_conv'])
elif which_model == 'discriminator_vgg_128_gn':
2020-08-31 15:50:30 +00:00
netD = SRGAN_arch.Discriminator_VGG_128_GN(in_nc=opt_net['in_nc'], nf=opt_net['nf'], input_img_factor=img_sz / 128)
2020-08-26 00:14:45 +00:00
if wrap:
netD = GradDiscWrapper(netD)
elif which_model == 'discriminator_resnet':
2020-05-02 01:56:14 +00:00
netD = DiscriminatorResnet_arch.fixup_resnet34(num_filters=opt_net['nf'], num_classes=1, input_img_size=img_sz)
elif which_model == 'discriminator_resnet_passthrough':
netD = DiscriminatorResnet_arch_passthrough.fixup_resnet34(num_filters=opt_net['nf'], num_classes=1, input_img_size=img_sz,
number_skips=opt_net['number_skips'], use_bn=True,
disable_passthrough=opt_net['disable_passthrough'])
2020-07-06 03:49:09 +00:00
elif which_model == 'discriminator_pix':
netD = SRGAN_arch.Discriminator_VGG_PixLoss(in_nc=opt_net['in_nc'], nf=opt_net['nf'])
elif which_model == "discriminator_unet":
netD = SRGAN_arch.Discriminator_UNet(in_nc=opt_net['in_nc'], nf=opt_net['nf'])
elif which_model == "discriminator_unet_fea":
netD = SRGAN_arch.Discriminator_UNet_FeaOut(in_nc=opt_net['in_nc'], nf=opt_net['nf'], feature_mode=opt_net['feature_mode'])
elif which_model == "discriminator_switched":
netD = SRGAN_arch.Discriminator_switched(in_nc=opt_net['in_nc'], nf=opt_net['nf'], initial_temp=opt_net['initial_temp'],
final_temperature_step=opt_net['final_temperature_step'])
2020-08-06 14:56:21 +00:00
elif which_model == "cross_compare_vgg128":
netD = SRGAN_arch.CrossCompareDiscriminator(in_nc=opt_net['in_nc'], ref_channels=opt_net['ref_channels'] if 'ref_channels' in opt_net.keys() else 3, nf=opt_net['nf'], scale=opt_net['scale'])
elif which_model == "discriminator_refvgg":
netD = SRGAN_arch.RefDiscriminatorVgg128(in_nc=opt_net['in_nc'], nf=opt_net['nf'], input_img_factor=img_sz / 128)
2019-08-23 13:42:47 +00:00
else:
raise NotImplementedError('Discriminator model [{:s}] not recognized'.format(which_model))
return netD
# Discriminator
2020-08-25 23:58:20 +00:00
def define_D(opt, wrap=False):
img_sz = opt['datasets']['train']['target_size']
opt_net = opt['network_D']
2020-08-25 23:58:20 +00:00
return define_D_net(opt_net, img_sz, wrap=wrap)
def define_fixed_D(opt):
# Note that this will not work with "old" VGG-style discriminators with dense blocks until the img_size parameter is added.
net = define_D_net(opt)
# Load the model parameters:
load_net = torch.load(opt['pretrained_path'])
load_net_clean = OrderedDict() # remove unnecessary 'module.'
for k, v in load_net.items():
if k.startswith('module.'):
load_net_clean[k[7:]] = v
else:
load_net_clean[k] = v
net.load_state_dict(load_net_clean)
# Put into eval mode, freeze the parameters and set the 'weight' field.
net.eval()
for k, v in net.named_parameters():
v.requires_grad = False
net.fdisc_weight = opt['weight']
2020-07-31 21:07:10 +00:00
return net
2019-08-23 13:42:47 +00:00
# Define network used for perceptual loss
def define_F(which_model='vgg', use_bn=False, for_training=False, load_path=None, feature_layers=None):
2020-08-22 19:08:33 +00:00
if which_model == 'vgg':
# PyTorch pretrained VGG19-54, before ReLU.
if feature_layers is None:
if use_bn:
feature_layers = [49]
else:
feature_layers = [34]
if for_training:
netF = feature_arch.TrainableVGGFeatureExtractor(feature_layers=feature_layers, use_bn=use_bn,
2020-08-22 19:08:33 +00:00
use_input_norm=True)
else:
netF = feature_arch.VGGFeatureExtractor(feature_layers=feature_layers, use_bn=use_bn,
2020-08-22 19:08:33 +00:00
use_input_norm=True)
elif which_model == 'wide_resnet':
netF = feature_arch.WideResnetFeatureExtractor(use_input_norm=True)
else:
raise NotImplementedError
if load_path:
# Load the model parameters:
load_net = torch.load(load_path)
load_net_clean = OrderedDict() # remove unnecessary 'module.'
for k, v in load_net.items():
if k.startswith('module.'):
load_net_clean[k[7:]] = v
else:
load_net_clean[k] = v
netF.load_state_dict(load_net_clean)
if not for_training:
# Put into eval mode, freeze the parameters and set the 'weight' field.
netF.eval()
for k, v in netF.named_parameters():
v.requires_grad = False
2019-08-23 13:42:47 +00:00
return netF