Switch default LR scheduler from cos to linear (ultralytics#6729)

* Switch default LR scheduler from cos to linear

Based on empirical results of training both ways on all YOLOv5 models.

* linear bug fix

train.py

@@ -176,10 +176,10 @@ def train(hyp,  # path/to/hyp.yaml or hyp dictionary
     del g0, g1, g2
 
     # Scheduler
-    if opt.linear_lr:
-        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
-    else:
+    if opt.cos_lr:
         lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
     scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
 
     # EMA
@@ -479,7 +479,7 @@ def parse_opt(known=False):
     parser.add_argument('--name', default='exp', help='save to project/name')
     parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
     parser.add_argument('--quad', action='store_true', help='quad dataloader')
-    parser.add_argument('--linear-lr', action='store_true', help='linear LR')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
     parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
     parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
     parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')