Fixup upconv for the next attempt!

codes/models/archs/DiscriminatorResnet_arch.py

@@ -94,7 +94,7 @@ class FixupBottleneck(nn.Module):
 
 class FixupResNet(nn.Module):
 
-    def __init__(self, block, layers, num_filters=64, num_classes=1000):
+    def __init__(self, block, layers, num_filters=64, num_classes=1000, input_img_size=64):
         super(FixupResNet, self).__init__()
         self.num_layers = sum(layers)
         self.inplanes = num_filters
@@ -107,7 +107,8 @@ class FixupResNet(nn.Module):
         self.layer3 = self._make_layer(block, num_filters*4, layers[2], stride=2)
         self.layer4 = self._make_layer(block, num_filters*8, layers[3], stride=2)
         self.bias2 = nn.Parameter(torch.zeros(1))
-        self.fc1 = nn.Linear(num_filters * 8 * 2 * 2, 100)
+        reduced_img_sz = int(input_img_size / 32)
+        self.fc1 = nn.Linear(num_filters * 8 * reduced_img_sz * reduced_img_sz, 100)
         self.fc2 = nn.Linear(100, num_classes)
 
         for m in self.modules():

codes/models/archs/RRDBNet_arch.py

@@ -50,13 +50,13 @@ class RRDBNet(nn.Module):
         super(RRDBNet, self).__init__()
         RRDB_block_f = functools.partial(RRDB, nf=nf, gc=gc)
 
-        self.conv_first = nn.Conv2d(in_nc, nf, 3, 1, 1, bias=True)
+        self.conv_first = nn.Conv2d(in_nc, nf, 7, 1, padding=3, bias=True)
         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, 3, 1, 1, bias=True)
+        self.upconv1 = nn.Conv2d(nf, nf, 5, 1, padding=2, bias=True)
-        self.upconv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
+        self.upconv2 = nn.Conv2d(nf, nf, 5, 1, padding=2, bias=True)
-        self.HRconv = nn.Conv2d(nf, nf, 3, 1, 1, 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)

codes/models/networks.py

@@ -58,7 +58,7 @@ def define_D(opt):
     if which_model == 'discriminator_vgg_128':
         netD = SRGAN_arch.Discriminator_VGG_128(in_nc=opt_net['in_nc'], nf=opt_net['nf'], input_img_factor=img_sz / 128)
     elif which_model == 'discriminator_resnet':
-        netD = DiscriminatorResnet_arch.fixup_resnet34(num_filters=opt_net['nf'], num_classes=1)
+        netD = DiscriminatorResnet_arch.fixup_resnet34(num_filters=opt_net['nf'], num_classes=1, input_img_size=img_sz)
     else:
         raise NotImplementedError('Discriminator model [{:s}] not recognized'.format(which_model))
     return netD

codes/options/train/finetune_ESRGAN_blacked.yml

@@ -1,87 +0,0 @@
-#### general settings
-name: ESRGANx4_blacked_ramped_feat
-use_tb_logger: true
-model: srgan
-distortion: sr
-scale: 4
-gpu_ids: [0]
-amp_opt_level: O1
-
-#### datasets
-datasets:
-  train:
-    name: blacked
-    mode: LQGT
-    dataroot_GT: ../datasets/blacked/train/hr
-    dataroot_LQ: ../datasets/lqprn/train/lr
-    dataroot_PIX: ../datasets/lqprn/train/hr
-
-    use_shuffle: true
-    n_workers: 4  # per GPU
-    batch_size: 12
-    target_size: 256
-    use_flip: false
-    use_rot: false
-    color: RGB
-  val:
-    name: blacked_val
-    mode: LQGT
-    dataroot_GT: ../datasets/vrp/validation/hr
-    dataroot_LQ: ../datasets/vrp/validation/lr
-
-#### network structures
-network_G:
-  which_model_G: RRDBNet
-  in_nc: 3
-  out_nc: 3
-  nf: 64
-  nb: 23
-network_D:
-  which_model_D: discriminator_vgg_128
-  in_nc: 3
-  nf: 64
-
-#### path
-path:
-  pretrain_model_G: ../experiments/blacked_ft_G.pth
-  pretrain_model_D: ../experiments/blacked_ft_D.pth
-  resume_state: ~
-  strict_load: true
-
-#### training settings: learning rate scheme, loss
-train:
-  lr_G: !!float 1e-4
-  weight_decay_G: 0
-  beta1_G: 0.9
-  beta2_G: 0.99
-  lr_D: !!float 1e-4
-  weight_decay_D: 0
-  beta1_D: 0.9
-  beta2_D: 0.99
-  lr_scheme: MultiStepLR
-
-  niter: 400000
-  warmup_iter: -1  # no warm up
-  lr_steps: [10000, 20000, 30000, 40000, 50000]
-  lr_gamma: 0.5
-
-  pixel_criterion: l1
-  pixel_weight: !!float 1e-2
-  feature_criterion: l1
-  feature_weight: !!float 1e0
-  feature_weight_decay: .95
-  feature_weight_decay_steps: 2500
-  feature_weight_minimum: !!float 1e-3
-  gan_type: ragan  # gan | ragan
-  gan_weight: !!float 1e-1
-
-  D_update_ratio: 1
-  D_init_iters: 0
-
-  manual_seed: 10
-  val_freq: !!float 5e2
-
-#### logger
-logger:
-  print_freq: 50
-  save_checkpoint_freq: !!float 5e2

codes/options/train/finetune_ESRGAN_blacked_for_adrianna.yml

@@ -1,85 +0,0 @@
-#### general settings
-name: ESRGANx4_blacked_for_adrianna
-use_tb_logger: true
-model: srgan
-distortion: sr
-scale: 4
-gpu_ids: [0]
-amp_opt_level: O1
-
-#### datasets
-datasets:
-  train:
-    name: vixen
-    mode: LQGT
-    dataroot_GT: K:\4k6k\vixen4k\hr
-    dataroot_LQ: E:\4k6k\mmsr\results\RRDB_ESRGAN_x4\vixen
-    use_shuffle: true
-    n_workers: 4  # per GPU
-    batch_size: 12
-    target_size: 256
-    use_flip: false
-    use_rot: false
-    color: RGB
-  val:
-    name: adrianna_val
-    mode: LQGT
-    dataroot_GT: ../datasets/adrianna/val/hr
-    dataroot_LQ: ../datasets/adrianna/val/lr
-
-#### network structures
-network_G:
-  which_model_G: RRDBNet
-  in_nc: 3
-  out_nc: 3
-  nf: 64
-  nb: 23
-network_D:
-  which_model_D: discriminator_vgg_128
-  in_nc: 3
-  nf: 64
-
-#### path
-path:
-  pretrain_model_G: ../experiments/ESRGANx4_blacked_for_adrianna/models/15500_G.pth
-  pretrain_model_D: ../experiments/ESRGANx4_blacked_for_adrianna/models/15500_D.pth
-  resume_state: ../experiments/ESRGANx4_blacked_for_adrianna/training_state/15500.state
-  strict_load: true
-
-#### training settings: learning rate scheme, loss
-train:
-  lr_G: !!float 1e-4
-  weight_decay_G: 0
-  beta1_G: 0.9
-  beta2_G: 0.99
-  lr_D: !!float 1e-4
-  weight_decay_D: 0
-  beta1_D: 0.9
-  beta2_D: 0.99
-  lr_scheme: MultiStepLR
-
-  niter: 400000
-  warmup_iter: -1  # no warm up
-  lr_steps: [10000, 20000, 30000, 40000, 50000]
-  lr_gamma: 0.5
-
-  pixel_criterion: l1
-  pixel_weight: !!float 1e-2
-  feature_criterion: l1
-  feature_weight: !!float 1e0
-  feature_weight_decay: 1
-  feature_weight_decay_steps: 2500
-  feature_weight_minimum: !!float 1e-3
-  gan_type: ragan  # gan | ragan
-  gan_weight: !!float 1e-1
-
-  D_update_ratio: 1
-  D_init_iters: 0
-
-  manual_seed: 10
-  val_freq: !!float 5e2
-
-#### logger
-logger:
-  print_freq: 50
-  save_checkpoint_freq: !!float 5e2

codes/options/train/train_ESRGAN.yml

@@ -5,18 +5,19 @@ model: srgan
 distortion: sr
 scale: 4
 gpu_ids: [0]
+amp_opt_level: O1
 
 #### datasets
 datasets:
   train:
     name: DIV2K
     mode: LQGT
-    dataroot_GT: ../datasets/div2k/DIV2K800_sub
+    dataroot_GT: E:/4k6k/datasets/div2k/DIV2K800_sub
-    dataroot_LQ: ../datasets/div2k/DIV2K800_sub_bicLRx4
+    dataroot_LQ: E:/4k6k/datasets/div2k/DIV2K800_sub_bicLRx4
 
     use_shuffle: true
     n_workers: 16  # per GPU
-    batch_size: 16
+    batch_size: 32
     target_size: 128
     use_flip: true
     use_rot: true
@@ -24,8 +25,8 @@ datasets:
   val:
     name: div2kval
     mode: LQGT
-    dataroot_GT: ../datasets/div2k/div2k_valid_hr
+    dataroot_GT: E:/4k6k/datasets/div2k/div2k_valid_hr
-    dataroot_LQ: ../datasets/div2k/div2k_valid_lr_bicubic
+    dataroot_LQ: E:/4k6k/datasets/div2k/div2k_valid_lr_bicubic
 
 #### network structures
 network_G:
@@ -35,13 +36,13 @@ network_G:
   nf: 64
   nb: 23
 network_D:
-  which_model_D: discriminator_vgg_128
+  which_model_D: discriminator_resnet
   in_nc: 3
   nf: 64
 
 #### path
 path:
-  pretrain_model_G: ../experiments/RRDB_PSNR_x4.pth
+  pretrain_model_G: ~
   strict_load: true
   resume_state: ~
 
@@ -66,10 +67,13 @@ train:
   pixel_weight: !!float 1e-2
   feature_criterion: l1
   feature_weight: 1
-  gan_type: ragan  # gan | ragan
+  feature_weight_decay: .98
+  feature_weight_decay_steps: 500
+  feature_weight_minimum: .1
+  gan_type: gan  # gan | ragan
   gan_weight: !!float 5e-3
 
-  D_update_ratio: 1
+  D_update_ratio: 2
   D_init_iters: 0
 
   manual_seed: 10

codes/options/train/train_ESRGAN_blacked.yml

@@ -1,58 +1,56 @@
 #### general settings
-name: ESRGANx4_VRP
+name: blacked_fix_and_upconv
 use_tb_logger: true
 model: srgan
 distortion: sr
 scale: 4
 gpu_ids: [0]
+amp_opt_level: O1
 
 #### datasets
 datasets:
   train:
-    name: VRP
+    name: vixcloseup
     mode: LQGT
-    dataroot_GT: ../datasets/vrp/train/hr
+    dataroot_GT: K:\4k6k\4k_closeup\hr
-    dataroot_LQ: ../datasets/vrp/train/lr
+    dataroot_LQ: K:\4k6k\4k_closeup\lr_corrupted
 
     use_shuffle: true
-    n_workers: 0  # per GPU
+    n_workers: 12  # per GPU
-    batch_size: 16
+    batch_size: 12
-    target_size: 128
+    target_size: 256
-    use_flip: true
-    use_rot: true
     color: RGB
   val:
-    name: VRP_val
+    name: adrianna_val
     mode: LQGT
-    dataroot_GT: ../datasets/vrp/validation/hr
+    dataroot_GT: E:\4k6k\datasets\adrianna\val\hhq
-    dataroot_LQ: ../datasets/vrp/validation/lr
+    dataroot_LQ: E:\4k6k\datasets\adrianna\val\hr
 
 #### network structures
 network_G:
   which_model_G: RRDBNet
   in_nc: 3
   out_nc: 3
-  nf: 64
+  nf: 48
   nb: 23
 network_D:
-  which_model_D: discriminator_vgg_128
+  which_model_D: discriminator_resnet
   in_nc: 3
-  nf: 64
+  nf: 48
 
 #### path
 path:
-  pretrain_model_G: ../experiments/div2k_gen_pretrain.pth
+  pretrain_model_G: ~
-  pretrain_model_D: ../experiments/div2k_disc_pretrain.pth
   strict_load: true
   resume_state: ~
 
 #### training settings: learning rate scheme, loss
 train:
-  lr_G: !!float 1e-5
+  lr_G: !!float 1e-4
   weight_decay_G: 0
   beta1_G: 0.9
   beta2_G: 0.99
-  lr_D: !!float 1e-5
+  lr_D: !!float 2e-4
   weight_decay_D: 0
   beta1_D: 0.9
   beta2_D: 0.99
@@ -60,14 +58,17 @@ train:
 
   niter: 400000
   warmup_iter: -1  # no warm up
-  lr_steps: [50000, 100000, 200000, 300000]
+  lr_steps: [20000, 40000, 60000, 80000]
   lr_gamma: 0.5
 
   pixel_criterion: l1
   pixel_weight: !!float 1e-2
   feature_criterion: l1
   feature_weight: 1
-  gan_type: ragan  # gan | ragan
+  feature_weight_decay: .98
+  feature_weight_decay_steps: 500
+  feature_weight_minimum: .1
+  gan_type: gan  # gan | ragan
   gan_weight: !!float 5e-3
 
   D_update_ratio: 1

codes/options/train/train_GAN_blacked_corrupt.yml

@@ -21,6 +21,7 @@ datasets:
     target_size: 64
     use_flip: false
     use_rot: false
+    doCrop: false
     color: RGB
   val:
     name: blacked_val
@@ -48,18 +49,18 @@ network_D:
 
 #### path
 path:
-  pretrain_model_G: ~
+  pretrain_model_G: ../experiments/corruptGAN_4k_lqprn_closeup_flat_net/models/29000_G.pth
-  pretrain_model_D: ~ #../experiments/resnet_corrupt_discriminator_fixup.pth
+  pretrain_model_D: ../experiments/corruptGAN_4k_lqprn_closeup_flat_net/models/29000_D.pth
-  resume_state: ~
+  resume_state: ../experiments/corruptGAN_4k_lqprn_closeup_flat_net/training_state/29000.state
   strict_load: true
 
 #### training settings: learning rate scheme, loss
 train:
-  lr_G: !!float 1e-4
+  lr_G: !!float 5e-5
   weight_decay_G: 0
   beta1_G: 0.9
   beta2_G: 0.99
-  lr_D: !!float 2e-4
+  lr_D: !!float 1e-4
   weight_decay_D: 0
   beta1_D: 0.9
   beta2_D: 0.99
@@ -77,8 +78,8 @@ train:
   gan_type: gan  # gan | ragan
   gan_weight: !!float 1e-1
 
-  D_update_ratio: 2
+  D_update_ratio: 1
-  D_init_iters: 0
+  D_init_iters: -1
 
   manual_seed: 10
   val_freq: !!float 5e2

codes/train.py

@@ -30,7 +30,7 @@ def init_dist(backend='nccl', **kwargs):
 def main():
     #### options
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='options/train/train_GAN_blacked_corrupt.yml')
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='options/train/train_ESRGAN_blacked.yml')
     parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none',
                         help='job launcher')
     parser.add_argument('--local_rank', type=int, default=0)
@@ -191,6 +191,8 @@ def main():
                     idx = 0
                     for val_data in val_loader:
                         idx += 1
+                        if idx >= 20:
+                            break
                         img_name = os.path.splitext(os.path.basename(val_data['LQ_path'][0]))[0]
                         img_dir = os.path.join(opt['path']['val_images'], img_name)
                         util.mkdir(img_dir)