EarlyStopper updates (ultralytics#4679)

train.py

@@ -344,7 +344,7 @@ def train(hyp,  # path/to/hyp.yaml or hyp dictionary
             # mAP
             callbacks.on_train_epoch_end(epoch=epoch)
             ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
-            final_epoch = epoch + 1 == epochs
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
             if not noval or final_epoch:  # Calculate mAP
                 results, maps, _ = val.run(data_dict,
                                            batch_size=batch_size // WORLD_SIZE * 2,
@@ -384,7 +384,7 @@ def train(hyp,  # path/to/hyp.yaml or hyp dictionary
                 callbacks.on_model_save(last, epoch, final_epoch, best_fitness, fi)
 
             # Stop Single-GPU
-            if stopper(epoch=epoch, fitness=fi):
+            if RANK == -1 and stopper(epoch=epoch, fitness=fi):
                 break
 
             # Stop DDP TODO: known issues shttps://github.com/ultralytics/yolov5/pull/4576
@@ -462,7 +462,7 @@ def parse_opt(known=False):
     parser.add_argument('--artifact_alias', type=str, default="latest", help='version of dataset artifact to be used')
     parser.add_argument('--local_rank', type=int, default=-1, help='DDP parameter, do not modify')
     parser.add_argument('--freeze', type=int, default=0, help='Number of layers to freeze. backbone=10, all=24')
-    parser.add_argument('--patience', type=int, default=30, help='EarlyStopping patience (epochs)')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
     opt = parser.parse_known_args()[0] if known else parser.parse_args()
     return opt
 

utils/torch_utils.py

@@ -298,13 +298,16 @@ class EarlyStopping:
     def __init__(self, patience=30):
         self.best_fitness = 0.0  # i.e. mAP
         self.best_epoch = 0
-        self.patience = patience  # epochs to wait after fitness stops improving to stop
+        self.patience = patience or float('inf')  # epochs to wait after fitness stops improving to stop
+        self.possible_stop = False  # possible stop may occur next epoch
 
     def __call__(self, epoch, fitness):
         if fitness >= self.best_fitness:  # >= 0 to allow for early zero-fitness stage of training
             self.best_epoch = epoch
             self.best_fitness = fitness
-        stop = (epoch - self.best_epoch) >= self.patience  # stop training if patience exceeded
+        delta = epoch - self.best_epoch  # epochs without improvement
+        self.possible_stop = delta >= (self.patience - 1)  # possible stop may occur next epoch
+        stop = delta >= self.patience  # stop training if patience exceeded
         if stop:
             LOGGER.info(f'EarlyStopping patience {self.patience} exceeded, stopping training.')
         return stop