eschmidbauer/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Enable HighToLowResNet to do a 1:1 transform

Browse Source
This commit is contained in:
James Betker 2020-04-25 21:36:32 -06:00
parent 35bd1ecae4
commit 02ff4a57fd
2 changed files with 42 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
codes/models/archs/HighToLowResNet.py
View File

@ -27,14 +27,18 @@ class HighToLowResNet(nn.Module):
self.trunk_lores = arch_util.make_layer(basic_block2, nb - 15) self.trunk_lores = arch_util.make_layer(basic_block2, nb - 15)
# downsampling # downsampling
if self.downscale == 4: if self.downscale == 4 or self.downscale == 1:
self.downconv1 = nn.Conv2d(nf, nf, 3, stride=2, padding=1, bias=True) self.downconv1 = nn.Conv2d(nf, nf, 3, stride=2, padding=1, bias=True)
self.downconv2 = nn.Conv2d(nf, nf*2, 3, stride=2, padding=1, bias=True) self.downconv2 = nn.Conv2d(nf, nf*2, 3, stride=2, padding=1, bias=True)
else: else:
raise EnvironmentError("Requested downscale not supported: %i" % (downscale,)) raise EnvironmentError("Requested downscale not supported: %i" % (downscale,))
self.HRconv = nn.Conv2d(nf*2, nf*2, 3, stride=1, padding=1, bias=True) self.HRconv = nn.Conv2d(nf*2, nf*2, 3, stride=1, padding=1, bias=True)
self.conv_last = nn.Conv2d(nf*2, out_nc, 3, stride=1, padding=1, bias=True) if self.downscale == 4:
self.conv_last = nn.Conv2d(nf*2, out_nc, 3, stride=1, padding=1, bias=True)
else:
self.pixel_shuffle = nn.PixelShuffle(4)
self.conv_last = nn.Conv2d(int(nf/8), out_nc, 3, stride=1, padding=1, bias=True)
# activation function # activation function
self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True) self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
@ -51,13 +55,22 @@ class HighToLowResNet(nn.Module):
out = self.lrelu(self.conv_first(out)) out = self.lrelu(self.conv_first(out))
out = self.trunk_hires(out) out = self.trunk_hires(out)
if self.downscale == 4: if self.downscale == 4 or self.downscale == 1:
out = self.lrelu(self.downconv1(out)) out = self.lrelu(self.downconv1(out))
out = self.trunk_medres(out) out = self.trunk_medres(out)
out = self.lrelu(self.downconv2(out)) out = self.lrelu(self.downconv2(out))
out = self.trunk_lores(out) out = self.trunk_lores(out)
out = self.conv_last(self.lrelu(self.HRconv(out))) if self.downscale == 1:
base = F.interpolate(x, scale_factor=1/self.downscale, mode='bilinear', align_corners=False) out = self.lrelu(self.pixel_shuffle(self.HRconv(out)))
out = self.conv_last(out)
else:
out = self.conv_last(self.lrelu(self.HRconv(out)))
if self.downscale == 1:
base = x
else:
base = F.interpolate(x, scale_factor=1/self.downscale, mode='bilinear', align_corners=False)
out += base out += base
return out return out

47
codes/options/train/train_DownsampleGAN_blacked.yml → codes/options/train/train_GAN_blacked_corrupt_via_transform.yml
View File

@ -1,9 +1,9 @@
#### general settings #### general settings
name: downsample_GAN_blacked name: corruptGAN_4k_lqprn_closeup_hq_to_hq
use_tb_logger: true use_tb_logger: true
model: srgan model: corruptgan
distortion: sr distortion: downsample
scale: .25 scale: 1
gpu_ids: [0] gpu_ids: [0]
amp_opt_level: O1 amp_opt_level: O1
@ -11,41 +11,41 @@ amp_opt_level: O1
datasets: datasets:
train: train:
name: blacked name: blacked
mode: GTLQ mode: downsample
dataroot_GT: ../datasets/blacked/train/hr dataroot_GT: K:\\4k6k\\4k_closeup\\hr
dataroot_LQ: ../datasets/lqprn/train/lr dataroot_LQ: E:\\4k6k\\adrianna\\for_training\\hr
mismatched_Data_OK: true
use_shuffle: true use_shuffle: true
n_workers: 0 # per GPU n_workers: 4 # per GPU
batch_size: 7 batch_size: 16
target_size: 64 target_size: 256
use_flip: false use_flip: false
use_rot: false use_rot: false
color: RGB color: RGB
val: val:
name: blacked_val name: blacked_val
mode: LQGT mode: downsample
dataroot_GT: ../datasets/vrp/validation/hr target_size: 256
dataroot_LQ: ../datasets/vrp/validation/lr dataroot_GT: ../datasets/blacked/val/hr
dataroot_LQ: ../datasets/blacked/val/hr
#### network structures #### network structures
network_G: network_G:
which_model_G: RRDBNet which_model_G: HighToLowResNet
in_nc: 3 in_nc: 3
out_nc: 3 out_nc: 3
nf: 64 nf: 64
nb: 23 nb: 56
network_D: network_D:
which_model_D: discriminator_vgg_128 which_model_D: discriminator_vgg_128
in_nc: 3 in_nc: 3
nf: 64 nf: 128
#### path #### path
path: path:
#pretrain_model_G: ../experiments/blacked_gen_20000_epochs.pth pretrain_model_G: ~
#pretrain_model_D: ../experiments/blacked_disc_20000_epochs.pth
strict_load: true
resume_state: ~ resume_state: ~
strict_load: true
#### training settings: learning rate scheme, loss #### training settings: learning rate scheme, loss
train: train:
@ -61,16 +61,17 @@ train:
niter: 400000 niter: 400000
warmup_iter: -1 # no warm up warmup_iter: -1 # no warm up
lr_steps: [10000, 30000, 50000, 70000] lr_steps: [4000, 8000, 12000, 15000, 20000]
lr_gamma: 0.5 lr_gamma: 0.5
pixel_criterion: l1 pixel_criterion: l1
pixel_weight: !!float 1e-2 pixel_weight: !!float 1e-2
feature_criterion: l1
feature_weight: 0 feature_weight: 0
gan_type: ragan # gan | ragan gan_type: ragan # gan | ragan
gan_weight: 1 gan_weight: !!float 1e-1
D_update_ratio: 1 D_update_ratio: 2
D_init_iters: 0 D_init_iters: 0
manual_seed: 10 manual_seed: 10