Deploy inference service from GitHub Actions

training/train.py

@@ -121,9 +121,9 @@ def run_epoch(model, loader, optimizer, device, training: bool, config: dict):
                 loss.backward()
                 optimizer.step()
             total_loss += loss.item()
-            hb_preds.extend(hb_pred.squeeze(1).cpu().numpy().tolist())
+            hb_preds.extend(hb_pred.detach().squeeze(1).cpu().numpy().tolist())
             hb_trues.extend(hb.cpu().numpy().tolist())
-            cls_preds.extend(torch.softmax(cls_logits, dim=1).cpu().numpy().tolist())
+            cls_preds.extend(torch.softmax(cls_logits.detach(), dim=1).cpu().numpy().tolist())
             cls_trues.extend(cls.cpu().numpy().tolist())
 
     reg_metrics = compute_regression_metrics(np.array(hb_trues), np.array(hb_preds))

training/utils/augmentation.py

@@ -29,8 +29,13 @@ def get_augmentation_pipeline(image_size: int = 380):
             A.Rotate(limit=15, p=0.7),
             A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.6),
             A.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=15, val_shift_limit=10, p=0.4),
-            A.GaussNoise(var_limit=(10, 50), p=0.2),
-            A.CoarseDropout(max_holes=4, max_height=32, max_width=32, p=0.3),
+            A.GaussNoise(std_range=(0.04, 0.12), p=0.2),
+            A.CoarseDropout(
+                num_holes_range=(1, 4),
+                hole_height_range=(0.05, 0.12),
+                hole_width_range=(0.05, 0.12),
+                p=0.3,
+            ),
         ]
     )