Update config_RGB-NIR.yaml

config_RGB-NIR.yaml

@@ -1,6 +1,6 @@
 Data:
-  train_batch_size: 12
-  val_batch_size: 8
+  train_batch_size: 6
+  val_batch_size: 4
   num_workers: 6
   prefetch_factor: 2
   dataset_type: SISR_WW
@@ -10,54 +10,65 @@ Model:
   continue_training: false
   load_checkpoint: false
 Training:
+  device: cuda
   gpus:
-  - 0
-  - 1
   - 2
   - 3
   max_epochs: 9999
-  val_check_interval: 1.0
+  val_check_interval: 0.25
   limit_val_batches: 250
   pretrain_g_only: true
-  g_pretrain_steps: 15000
-  adv_loss_ramp_steps: 2500
+  g_pretrain_steps: 20000
+  adv_loss_ramp_steps: 5000
   label_smoothing: true
   EMA:
-    enabled: true
+    enabled: false
     decay: 0.999
     update_after_step: 0
     use_num_updates: true
   Losses:
-    adv_loss_beta: 0.001
+    adv_loss_beta: 0.0005
     adv_loss_schedule: cosine
     l1_weight: 1.0
     sam_weight: 0.05
-    perceptual_weight: 0.1
+    perceptual_weight: 0.2
     perceptual_metric: vgg
     tv_weight: 0.0
     max_val: 1.0
     ssim_win: 11
 Generator:
-  model_type: SRResNet
+  model_type: rrdb
   large_kernel_size: 9
   small_kernel_size: 3
-  n_channels: 96
+  n_channels: 64
   n_blocks: 16
   scaling_factor: 4
 Discriminator:
-  model_type: patchgan
-  n_blocks: 3
+  model_type: standard
+  n_blocks: 8
 Optimizers:
   optim_g_lr: 0.0001
-  optim_d_lr: 0.0001
+  optim_d_lr: 1.0e-06
+  gradient_clip_val: 1.0
+  betas:
+  - 0.0
+  - 0.99
+  eps: 1.0e-07
+  weight_decay_g: 0.0
+  weight_decay_d: 0.0
 Schedulers:
-  g_warmup_steps: 5000
+  g_warmup_steps: 1000
   g_warmup_type: cosine
-  metric: val_metrics/l1
-  patience_g: 100
-  patience_d: 100
+  metric_g: val_metrics/l1
+  metric_d: validation/DISC_adversarial_loss
+  patience_g: 10
+  patience_d: 10
   factor_g: 0.5
   factor_d: 0.5
   verbose: true
 Logging:
   num_val_images: 5
+  wandb:
+    enabled: true
+    entity: opensr
+    project: SRGAN_10m