added early stopping with max, min and percentage mode

Author: niklasrah

tools/train.py

@@ -88,6 +88,13 @@ def make_parser():
                 Implemented loggers include `tensorboard`, `mlflow` and `wandb`.",
         default="tensorboard"
     )
+    parser.add_argument(
+        "--early-stopping",
+        dest="early_stopping",
+        default=False,
+        action="store_true",
+        help="Use early stopping to prevent overfitting.",
+    )
     parser.add_argument(
         "opts",
         help="Modify config options using the command-line",
@@ -115,6 +122,13 @@ def main(exp: Exp, args):
     cudnn.benchmark = True
 
     trainer = exp.get_trainer(args)
+
+    # configure early stopping parameters
+    if args.early_stopping:
+        # requires 1% relative improvement over 10 epochs to reset patience
+        # available modes: "max", "min", "percentage"
+        trainer.early_stopper = exp.get_early_stopping(patience=10, min_delta=0.01, mode="percentage")
+
     trainer.train()
 
 

yolox/core/__init__.py

@@ -4,3 +4,4 @@
 
 from .launch import launch
 from .trainer import Trainer
+from .trainer import EarlyStopping

yolox/core/trainer.py

@@ -33,6 +33,40 @@
     synchronize
 )
 
+class EarlyStopping:
+    def __init__(self, patience: int, min_delta: float, mode="max"):
+        self.patience = patience
+        self.min_delta = min_delta
+        self.mode = mode # "max", "min", "percentage"
+        self.best = None
+        self.counter = 0
+
+    def step(self, value):
+        # Initialize best value on first call
+        if self.best is None:
+            self.best = value
+            return False
+
+        # Compute improvement depending on mode
+        if self.mode == "max":
+            improvement = value - self.best
+        elif self.mode == "min":
+            improvement = self.best - value
+        elif self.mode == "percentage":
+            if self.best == 0:
+                improvement = 0  # avoid division by zero
+            else:
+                improvement = (value - self.best) / abs(self.best)
+
+        # Check if improvement is sufficient
+        if improvement > self.min_delta:
+            self.best = value
+            self.counter = 0
+        else:
+            self.counter += 1
+
+        return self.counter >= self.patience
+
 
 class Trainer:
     def __init__(self, exp: Exp, args):
@@ -234,7 +268,15 @@ def after_epoch(self):
 
         if (self.epoch + 1) % self.exp.eval_interval == 0:
             all_reduce_norm(self.model)
-            self.evaluate_and_save_model()
+            ap50_95 = self.evaluate_and_save_model()
+
+        # Early stopping
+        if self.early_stopper is not None:
+            if self.early_stopper.step(ap50_95):
+                logger.info(f"Early stopping triggered at epoch {self.epoch}. " f"Best AP: {self.early_stopper.best}")
+                # save best checkpoint before exiting
+                self.save_ckpt("best_ckpt")
+                raise SystemExit
 
     def before_iter(self):
         pass
@@ -395,6 +437,7 @@ def evaluate_and_save_model(self):
                 }
             self.mlflow_logger.save_checkpoints(self.args, self.exp, self.file_name, self.epoch,
                                                 metadata, update_best_ckpt)
+        return ap50_95
 
     def save_ckpt(self, ckpt_name, update_best_ckpt=False, ap=None):
         if self.rank == 0: