🚸 [Add] try-except in loading cache files

yolo/config/task/validation.yaml

@@ -7,7 +7,7 @@ data:
   shuffle: False
   pin_memory: True
   data_augment: {}
-  dynamic_shape: True
+  dynamic_shape: False
 nms:
   min_confidence: 0.0001
   min_iou: 0.7

yolo/tools/data_loader.py

@@ -56,7 +56,15 @@ class YoloDataset(Dataset):
             data = self.filter_data(dataset_path, phase_name, self.dynamic_shape)
             torch.save(data, cache_path)
         else:
-            data = torch.load(cache_path, weights_only=False)
+            try:
+                data = torch.load(cache_path, weights_only=False)
+            except Exception as e:
+                logger.error(
+                    f":rotating_light: Failed to load the cache at '{cache_path}'.\n"
+                    ":rotating_light: This may be caused by using cache from different other YOLO.\n"
+                    ":rotating_light: Please clean the cache and try running again."
+                )
+                raise e
             logger.info(f":package: Loaded {phase_name} cache")
         return data
 

yolo/tools/solver.py

@@ -56,7 +56,6 @@ class ValidateModel(BaseModel):
                 "map": batch_metrics["map"],
                 "map_50": batch_metrics["map_50"],
             },
-            on_step=True,
             batch_size=batch_size,
         )
         return predicts
@@ -102,9 +101,10 @@ class TrainModel(ValidateModel):
             prog_bar=True,
             on_epoch=True,
             batch_size=batch_size,
+            sync_dist=True,
             rank_zero_only=True,
         )
-        self.log_dict(lr_dict, prog_bar=False, logger=True, on_epoch=False, rank_zero_only=True)
+        self.log_dict(lr_dict, prog_bar=False, logger=True, on_epoch=False, sync_dist=True, rank_zero_only=True)
         return loss * batch_size
 
     def configure_optimizers(self):

yolo/utils/dataset_utils.py

@@ -115,7 +115,14 @@ def scale_segmentation(
 
 
 def tensorlize(data):
-    img_paths, bboxes, img_ratios = zip(*data)
+    try:
+        img_paths, bboxes, img_ratios = zip(*data)
+    except ValueError as e:
+        logger.error(
+            ":rotating_light: This may be caused by using old cache or another version of YOLO's cache.\n"
+            ":rotating_light: Please clean the cache and try running again."
+        )
+        raise e
     max_box = max(bbox.size(0) for bbox in bboxes)
     padded_bbox_list = []
     for bbox in bboxes:

yolo/utils/model_utils.py

@@ -47,6 +47,8 @@ class EMA(Callback):
     def setup(self, trainer, pl_module, stage):
         pl_module.ema = deepcopy(pl_module.model)
         self.ema_parameters = [param.clone().detach().to(pl_module.device) for param in pl_module.parameters()]
+        for param in pl_module.ema.parameters():
+            param.requires_grad = False
 
     def on_validation_start(self, trainer: "Trainer", pl_module: "LightningModule"):
         for param, ema_param in zip(pl_module.ema.parameters(), self.ema_parameters):