diff --git a/classification/engine.py b/classification/engine.py
index 65eb52e..cc31103 100644
--- a/classification/engine.py
+++ b/classification/engine.py
@@ -68,7 +68,7 @@ def train_one_epoch(model: torch.nn.Module, criterion: DistillationLoss,
 
 
 @torch.no_grad()
-def evaluate(data_loader, model, device):
+def evaluate(data_loader, model, device, fp32=False):
     criterion = torch.nn.CrossEntropyLoss()
 
     metric_logger = utils.MetricLogger(delimiter="  ")
@@ -82,7 +82,7 @@ def evaluate(data_loader, model, device):
         target = target.to(device, non_blocking=True)
 
         # compute output
-        with torch.cuda.amp.autocast():
+        with torch.cuda.amp.autocast(enabled=not fp32):
             output = model(images)
             loss = criterion(output, target)
 
diff --git a/classification/main.py b/classification/main.py
index c4eccbf..f50e591 100644
--- a/classification/main.py
+++ b/classification/main.py
@@ -361,7 +361,7 @@ def main(args):
                 loss_scaler.load_state_dict(checkpoint['scaler'])
 
     if args.eval:
-        test_stats = evaluate(data_loader_val, model, device)
+        test_stats = evaluate(data_loader_val, model, device, fp32=args.fp32_resume)
         print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
         return