Merge pull request #380 from yuedongli1/v0.4.0

fix yolov8x-segment bug and fix scale for loss hyps

configs/yolov8/seg/yolov8-seg-base.yaml

@@ -4,6 +4,7 @@ img_size: 640
 conf_thres: 0.001
 iou_thres: 0.7
 conf_free: True
+anchor_base: False
 sync_bn: True
 clip_grad: True
 clip_grad_value: 10.0

configs/yolov8/yolov8-base.yaml

@@ -4,6 +4,7 @@ img_size: 640
 iou_thres: 0.7
 conf_free: True
 sync_bn: True
+anchor_base: False
 opencv_threads_num: 0  # opencv: disable threading optimizations
 
 network:

mindyolo/data/dataset.py

@@ -687,7 +687,7 @@ def random_perspective(
             sample['cls'] = cls
             sample['bboxes'] = targets
             if use_segments:
-                sample['segments'] = segments[i]
+                sample['segments'] = new_segments[i]
 
         sample['img'] = img
 

train.py

@@ -30,6 +30,7 @@ def get_parser_train(parents=None):
     parser.add_argument("--ms_amp_level", type=str, default="O0", help="amp level, O0/O1/O2/O3")
     parser.add_argument("--keep_loss_fp32", type=ast.literal_eval, default=True,
                         help="Whether to maintain loss using fp32/O0-level calculation")
+    parser.add_argument("--anchor_base", type=ast.literal_eval, default=True, help="Anchor-base")
     parser.add_argument("--ms_loss_scaler", type=str, default="static", help="train loss scaler, static/dynamic/none")
     parser.add_argument("--ms_loss_scaler_value", type=float, default=1024.0, help="static loss scale value")
     parser.add_argument("--ms_jit", type=ast.literal_eval, default=True, help="use jit or not")
@@ -188,6 +189,13 @@ def train(args):
     else:
         eval_dataset, eval_dataloader = None, None
 
+    # Scale loss hyps
+    nl = network.model.model[-1].nl
+    args.loss.box *= 3 / nl  # scale to layers
+    args.loss.cls *= args.data.nc / 80 * 3 / nl  # scale to classes and layers
+    if args.anchor_base:
+        args.loss.obj *= (args.img_size / 640) ** 2 * 3 / nl  # scale to image size and layers
+
     # Create Loss
     loss_fn = create_loss(
         **args.loss, anchors=args.network.get("anchors", 1), stride=args.network.stride, nc=args.data.nc