From 9b44f6f5c0a029f335967713a85953425ca54449 Mon Sep 17 00:00:00 2001 From: James Betker Date: Sat, 23 May 2020 21:09:21 -0600 Subject: [PATCH] Add AssistedRRDB and remove RRDBNetXL --- codes/models/archs/RRDBNetXL_arch.py | 98 ---------------------------- codes/models/archs/RRDBNet_arch.py | 80 +++++++++++++++++++++++ codes/models/networks.py | 15 ++--- 3 files changed, 86 insertions(+), 107 deletions(-) delete mode 100644 codes/models/archs/RRDBNetXL_arch.py diff --git a/codes/models/archs/RRDBNetXL_arch.py b/codes/models/archs/RRDBNetXL_arch.py deleted file mode 100644 index 9831aa9f..00000000 --- a/codes/models/archs/RRDBNetXL_arch.py +++ /dev/null @@ -1,98 +0,0 @@ -import functools -import torch -import torch.nn as nn -import torch.nn.functional as F -import models.archs.arch_util as arch_util - - -class ResidualDenseBlock_5C(nn.Module): - def __init__(self, nf=64, gc=32, bias=True): - super(ResidualDenseBlock_5C, self).__init__() - # gc: growth channel, i.e. intermediate channels - self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias) - self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias) - self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias) - self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias) - self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias) - self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) - - # initialization - arch_util.initialize_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], - 0.1) - - def forward(self, x): - x1 = self.lrelu(self.conv1(x)) - x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1))) - x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1))) - x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1))) - x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1)) - return x5 * 0.2 + x - - -class RRDB(nn.Module): - '''Residual in Residual Dense Block''' - - def __init__(self, nf, gc=32): - super(RRDB, self).__init__() - self.RDB1 = ResidualDenseBlock_5C(nf, gc) - self.RDB2 = ResidualDenseBlock_5C(nf, gc) - self.RDB3 = ResidualDenseBlock_5C(nf, gc) - - def forward(self, x): - out = self.RDB1(x) - out = self.RDB2(out) - out = self.RDB3(out) - return out * 0.2 + x - - -class RRDBNet(nn.Module): - def __init__(self, in_nc, out_nc, nf, nb_lo, nb_med, nb_hi, gc=32, interpolation_scale_factor=2): - super(RRDBNet, self).__init__() - nfmed = int(nf/2) - nfhi = int(nf/8) - gcmed = int(gc/2) - gchi = int(gc/8) - RRDB_block_f_lo = functools.partial(RRDB, nf=nf, gc=gc) - RRDB_block_f_lo_med = functools.partial(RRDB, nf=nfmed, gc=gcmed) - RRDB_block_f_lo_hi = functools.partial(RRDB, nf=nfhi, gc=gchi) - - self.conv_first = nn.Conv2d(in_nc, nf, 7, 1, padding=3, bias=True) - self.RRDB_trunk_lo = arch_util.make_layer(RRDB_block_f_lo, nb_lo) - self.trunk_conv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True) - self.lo_skip_conv1 = nn.Conv2d(nf, nf, 3, 1, padding=1, bias=True) - self.lo_skip_conv2 = nn.Conv2d(nf, out_nc, 3, 1, 1, bias=True) - - #### upsampling - self.upconv1 = nn.Conv2d(nf, nfmed, 3, 1, padding=1, bias=True) - self.RRDB_trunk_med = arch_util.make_layer(RRDB_block_f_lo_med, nb_med) - self.trunk_conv_med = nn.Conv2d(nfmed, nfmed, 3, 1, 1, bias=True) - self.med_skip_conv1 = nn.Conv2d(nfmed, nfmed, 3, 1, padding=1, bias=True) - self.med_skip_conv2 = nn.Conv2d(nfmed, out_nc, 3, 1, 1, bias=True) - - self.upconv2 = nn.Conv2d(nfmed, nfhi, 3, 1, padding=1, bias=True) - self.RRDB_trunk_hi = arch_util.make_layer(RRDB_block_f_lo_hi, nb_hi) - self.trunk_conv_hi = nn.Conv2d(nfhi, nfhi, 3, 1, 1, bias=True) - self.HRconv = nn.Conv2d(nfhi, nfhi, 5, 1, padding=2, bias=True) - self.conv_last = nn.Conv2d(nfhi, out_nc, 3, 1, 1, bias=True) - - self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) - - self.interpolation_scale_factor = interpolation_scale_factor - - def forward(self, x): - fea = self.conv_first(x) - branch = self.trunk_conv(self.RRDB_trunk_lo(fea)) - fea = (fea + branch) / 2 - lo_skip = self.lo_skip_conv2(self.lrelu(self.lo_skip_conv1(fea))) - - fea = self.lrelu(self.upconv1(F.interpolate(fea, scale_factor=self.interpolation_scale_factor, mode='nearest'))) - branch = self.trunk_conv_med(self.RRDB_trunk_med(fea)) - fea = (fea + branch) / 2 - med_skip = self.med_skip_conv2(self.lrelu(self.med_skip_conv1(fea))) - - fea = self.lrelu(self.upconv2(F.interpolate(fea, scale_factor=self.interpolation_scale_factor, mode='nearest'))) - branch = self.trunk_conv_hi(self.RRDB_trunk_hi(fea)) - fea = (fea + branch) / 2 - out = self.conv_last(self.lrelu(self.HRconv(fea))) - - return out, med_skip, lo_skip \ No newline at end of file diff --git a/codes/models/archs/RRDBNet_arch.py b/codes/models/archs/RRDBNet_arch.py index d11170d9..5a73c490 100644 --- a/codes/models/archs/RRDBNet_arch.py +++ b/codes/models/archs/RRDBNet_arch.py @@ -3,6 +3,7 @@ import torch import torch.nn as nn import torch.nn.functional as F import models.archs.arch_util as arch_util +import torchvision class ResidualDenseBlock_5C(nn.Module): @@ -76,3 +77,82 @@ class RRDBNet(nn.Module): out = self.conv_last(self.lrelu(self.HRconv(fea))) return (out,) + +# Variant of RRDBNet that is "assisted" by an external pretrained image classifier whose +# intermediate layers have been splayed out, pixel-shuffled, and fed back in. +class AssistedRRDBNet(nn.Module): + # in_nc=number of input channels. + # out_nc=number of output channels. + # nf=internal filter count + # nb=number of additional blocks after the assistance layers. + # gc=growth channel inside of residual blocks + # scale=the number of times the output is doubled in size. + def __init__(self, in_nc, out_nc, nf, nb, gc=32, scale=2): + super(AssistedRRDBNet, self).__init__() + self.scale = scale + self.conv_first = nn.Conv2d(in_nc, nf, 7, 1, padding=3, bias=True) + + # Set-up the assist-net, which should do feature extraction for us. + self.assistnet = torchvision.models.wide_resnet50_2(pretrained=True) + self.set_enable_assistnet_training(False) + assist_nf = [2, 4, 8, 16] # Fixed for resnet. Re-evaluate if using other networks. + self.assist1 = RRDB(nf + assist_nf[0], gc) + self.assist2 = RRDB(nf + sum(assist_nf[:2]), gc) + self.assist3 = RRDB(nf + sum(assist_nf[:3]), gc) + self.assist4 = RRDB(nf + sum(assist_nf), gc) + nf = nf + sum(assist_nf) + + # After this, it's just a "standard" RRDB net. + RRDB_block_f = functools.partial(RRDB, nf=nf, gc=gc) + self.RRDB_trunk = arch_util.make_layer(RRDB_block_f, nb) + self.trunk_conv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True) + #### upsampling + self.upconv1 = nn.Conv2d(nf, nf, 5, 1, padding=2, bias=True) + self.upconv2 = nn.Conv2d(nf, nf, 5, 1, padding=2, bias=True) + self.HRconv = nn.Conv2d(nf, nf, 5, 1, padding=2, bias=True) + self.conv_last = nn.Conv2d(nf, out_nc, 3, 1, 1, bias=True) + + self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) + + def set_enable_assistnet_training(self, en): + for p in self.assistnet.parameters(): + p.requires_grad = en + + def res_extract(self, x): + x = self.assistnet.conv1(x) + x = self.assistnet.bn1(x) + x = self.assistnet.relu(x) + x = self.assistnet.maxpool(x) + + x = self.assistnet.layer1(x) + l1 = F.pixel_shuffle(x, 4) + x = self.assistnet.layer2(x) + l2 = F.pixel_shuffle(x, 8) + x = self.assistnet.layer3(x) + l3 = F.pixel_shuffle(x, 16) + x = self.assistnet.layer4(x) + l4 = F.pixel_shuffle(x, 32) + return l1, l2, l3, l4 + + def forward(self, x): + # Invoke the assistant net first. + l1, l2, l3, l4 = self.res_extract(x) + + fea = self.conv_first(x) + fea = self.assist1(torch.cat([fea, l4], dim=1)) + fea = self.assist2(torch.cat([fea, l3], dim=1)) + fea = self.assist3(torch.cat([fea, l2], dim=1)) + fea = self.assist4(torch.cat([fea, l1], dim=1)) + + trunk = self.trunk_conv(self.RRDB_trunk(fea)) + fea = fea + trunk + + if self.scale >= 2: + fea = F.interpolate(fea, scale_factor=2, mode='nearest') + fea = self.lrelu(self.upconv1(fea)) + if self.scale >= 4: + fea = F.interpolate(fea, scale_factor=2, mode='nearest') + fea = self.lrelu(self.upconv2(fea)) + out = self.conv_last(self.lrelu(self.HRconv(fea))) + + return (out,) \ No newline at end of file diff --git a/codes/models/networks.py b/codes/models/networks.py index 98c17b46..c66ac347 100644 --- a/codes/models/networks.py +++ b/codes/models/networks.py @@ -5,12 +5,9 @@ import models.archs.DiscriminatorResnet_arch as DiscriminatorResnet_arch import models.archs.DiscriminatorResnet_arch_passthrough as DiscriminatorResnet_arch_passthrough import models.archs.FlatProcessorNetNew_arch as FlatProcessorNetNew_arch import models.archs.RRDBNet_arch as RRDBNet_arch -import models.archs.RRDBNetXL_arch as RRDBNetXL_arch -#import models.archs.EDVR_arch as EDVR_arch import models.archs.HighToLowResNet as HighToLowResNet -import models.archs.FlatProcessorNet_arch as FlatProcessorNet_arch -import models.archs.arch_util as arch_utils import models.archs.ResGen_arch as ResGen_arch +import models.archs.biggan_gen_arch as biggan_arch import math # Generator @@ -27,11 +24,9 @@ def define_G(opt, net_key='network_G'): # RRDB does scaling in two steps, so take the sqrt of the scale we actually want to achieve and feed it to RRDB. netG = RRDBNet_arch.RRDBNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], nf=opt_net['nf'], nb=opt_net['nb'], scale=scale) - elif which_model == 'RRDBNetXL': - scale_per_step = math.sqrt(scale) - netG = RRDBNetXL_arch.RRDBNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], - nf=opt_net['nf'], nb_lo=opt_net['nblo'], nb_med=opt_net['nbmed'], nb_hi=opt_net['nbhi'], - interpolation_scale_factor=scale_per_step) + elif which_model == 'AssistedRRDBNet': + netG = RRDBNet_arch.AssistedRRDBNet(in_nc=opt_net['in_nc'], out_nc=opt_net['out_nc'], + nf=opt_net['nf'], nb=opt_net['nb'], scale=scale) elif which_model == 'ResGen': netG = ResGen_arch.fixup_resnet34(nb_denoiser=opt_net['nb_denoiser'], nb_upsampler=opt_net['nb_upsampler'], upscale_applications=opt_net['upscale_applications'], num_filters=opt_net['nf']) @@ -39,6 +34,8 @@ def define_G(opt, net_key='network_G'): netG = ResGen_arch.fixup_resnet34_v2(nb_denoiser=opt_net['nb_denoiser'], nb_upsampler=opt_net['nb_upsampler'], upscale_applications=opt_net['upscale_applications'], num_filters=opt_net['nf'], inject_noise=opt_net['inject_noise']) + elif which_model == "BigGan": + netG = biggan_arch.biggan_medium(filters=opt_net['nf']) # image corruption elif which_model == 'HighToLowResNet':