No LR Scheduler + 10 Epochs + Unfreeze all layers + No Training Blur

evaluator/model.py

@@ -32,9 +32,6 @@ class ZEWDPCModel(torch.nn.Module):
         self.base_model = torchvision.models.efficientnet_b4(
             pretrained=self.use_pretrained,
         )
-        # Freeze feature extration layers
-        for param in self.base_model.features.parameters():
-            param.requires_grad = False
 
         # Replace the final FC layer to support
         # the required number of classes

evaluator/trainer.py

@@ -25,7 +25,6 @@ from evaluator.evaluation_metrics import get_zew_dpc_metrics
 import torch
 import torchvision
 from torchvision import transforms as T
-from torch.optim.lr_scheduler import ReduceLROnPlateau, CosineAnnealingLR
 
 import torchmetrics
 
@@ -171,8 +170,6 @@ class ZEWDPCTrainer:
         lr = self.hparams["learning_rate"]
         optimizer = torch.optim.Adam(params=self.model.parameters(), lr=lr)
 
-        lr_sched = CosineAnnealingLR(optimizer, num_epochs, eta_min=1e-5, last_epoch=-1, verbose=True)
-
         # Setup Metric Meters
         val_loss_avg_meter = AverageMeter()
         train_loss_avg_meter = AverageMeter()
@@ -264,8 +261,6 @@ class ZEWDPCTrainer:
                         val_loss_avg_meter.update(loss.item(), image.shape[0])
                         tqdm_iter.set_postfix(avg_val_loss=val_loss_avg_meter.avg)
 
-                lr_sched.step()
-
                 print(
                     "Epoch %d - Average Val Loss: %.5f \t Val F1: %.5f \t Learning Rate %0.5f"
                     % (
@@ -279,8 +274,6 @@ class ZEWDPCTrainer:
                     best_val_loss = val_loss_avg_meter.avg
                     self.best_model = copy.deepcopy(self.model)
 
-                    
-
             train_metrics = {"f1": train_f1.compute().item()}
             val_metrics = {"f1": val_f1.compute().item()}
             info = {"learning_rate": optimizer.param_groups[0]["lr"]}

local_evaluation.py

@@ -131,7 +131,7 @@ else:
 
 trainer = ZEWDPCTrainer(num_classes=6, use_pretrained=True)
 trainer.train(
-    aggregated_dataset, num_epochs=25, validation_percentage=0.1, batch_size=64
+    aggregated_dataset, num_epochs=10, validation_percentage=0.1, batch_size=32
 )
 
 y_pred = trainer.predict(val_dataset)