Update config_RGB-NIR.yaml

config_RGB-NIR.yaml

@@ -1,8 +1,8 @@
 Data:
-  train_batch_size: 6
+  train_batch_size: 12
   val_batch_size: 4
   num_workers: 6
-  prefetch_factor: 2
+  prefetch_factor: 4
   dataset_type: SISR_WW
   normalization: normalise_10k
 Model:
@@ -12,14 +12,16 @@ Model:
 Training:
   device: cuda
   gpus:
+  - 0
+  - 1
   - 2
   - 3
   max_epochs: 9999
-  val_check_interval: 0.25
+  val_check_interval: 0.5
   limit_val_batches: 250
   pretrain_g_only: true
-  g_pretrain_steps: 20000
-  adv_loss_ramp_steps: 5000
+  g_pretrain_steps: 25000
+  adv_loss_ramp_steps: 10000
   label_smoothing: true
   EMA:
     enabled: false
@@ -27,7 +29,7 @@ Training:
     update_after_step: 0
     use_num_updates: true
   Losses:
-    adv_loss_beta: 0.0005
+    adv_loss_beta: 0.001
     adv_loss_schedule: cosine
     l1_weight: 1.0
     sam_weight: 0.05
@@ -37,15 +39,20 @@ Training:
     max_val: 1.0
     ssim_win: 11
 Generator:
-  model_type: rrdb
+  model_type: SRResNet
+  block_type: rrdb
   large_kernel_size: 9
   small_kernel_size: 3
   n_channels: 64
-  n_blocks: 16
+  n_blocks: 24
   scaling_factor: 4
+  growth_channels: 32
+  res_scale: 0.2
 Discriminator:
   model_type: standard
   n_blocks: 8
+  base_channels: 64
+  linear_size: 1024
 Optimizers:
   optim_g_lr: 0.0001
   optim_d_lr: 1.0e-06
@@ -60,7 +67,7 @@ Schedulers:
   g_warmup_steps: 1000
   g_warmup_type: cosine
   metric_g: val_metrics/l1
-  metric_d: validation/DISC_adversarial_loss
+  metric_d: discriminator/adversarial_loss
   patience_g: 10
   patience_d: 10
   factor_g: 0.5