ultralytics-opencv-headless 8.3.253__py3-none-any.whl → 8.4.0__py3-none-any.whl

ultralytics/cfg/models/26/yolo26-pose.yaml

@@ -0,0 +1,53 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
+# Model docs: https://docs.ultralytics.com/models/yolo26
+# Task docs: https://docs.ultralytics.com/tasks/pose
+
+# Parameters
+nc: 80 # number of classes
+end2end: True # whether to use end-to-end mode
+reg_max: 1 # DFL bins
+kpt_shape: [17, 3] # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
+scales: # model compound scaling constants, i.e. 'model=YOLO26n.yaml' will call YOLO26.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 363 layers, 3,747,554 parameters, 3,747,554 gradients, 10.7 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 363 layers, 11,870,498 parameters, 11,870,498 gradients, 29.6 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 383 layers, 24,344,482 parameters, 24,344,482 gradients, 85.9 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 495 layers, 28,747,938 parameters, 28,747,938 gradients, 104.3 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 495 layers, 62,914,350 parameters, 62,914,350 gradients, 226.3 GFLOPs
+
+# YOLO26n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
+  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
+  - [-1, 2, C3k2, [256, False, 0.25]]
+  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
+  - [-1, 2, C3k2, [512, False, 0.25]]
+  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
+  - [-1, 2, C3k2, [512, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
+  - [-1, 2, C3k2, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5, 3, True]] # 9
+  - [-1, 2, C2PSA, [1024]] # 10
+
+# YOLO26n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 2, C3k2, [512, True]] # 13
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 2, C3k2, [256, True]] # 16 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 13], 1, Concat, [1]] # cat head P4
+  - [-1, 2, C3k2, [512, True]] # 19 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 10], 1, Concat, [1]] # cat head P5
+  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 22 (P5/32-large)
+
+  - [[16, 19, 22], 1, Pose26, [nc, kpt_shape]] # Detect(P3, P4, P5)

ultralytics/cfg/models/26/yolo26-seg.yaml

@@ -0,0 +1,52 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
+# Model docs: https://docs.ultralytics.com/models/yolo26
+# Task docs: https://docs.ultralytics.com/tasks/segment
+
+# Parameters
+nc: 80 # number of classes
+end2end: True # whether to use end-to-end mode
+reg_max: 1 # DFL bins
+scales: # model compound scaling constants, i.e. 'model=YOLO26n.yaml' will call YOLO26.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 309 layers, 3,126,280 parameters, 3,126,280 gradients, 10.5 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 309 layers, 11,505,800 parameters, 11,505,800 gradients, 37.4 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 329 layers, 27,112,072 parameters, 27,112,072 gradients, 132.5 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 441 layers, 31,515,528 parameters, 31,515,528 gradients, 150.9 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 441 layers, 70,693,800 parameters, 70,693,800 gradients, 337.7 GFLOPs
+
+# YOLO26n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
+  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
+  - [-1, 2, C3k2, [256, False, 0.25]]
+  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
+  - [-1, 2, C3k2, [512, False, 0.25]]
+  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
+  - [-1, 2, C3k2, [512, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
+  - [-1, 2, C3k2, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5, 3, True]] # 9
+  - [-1, 2, C2PSA, [1024]] # 10
+
+# YOLO26n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 2, C3k2, [512, True]] # 13
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 2, C3k2, [256, True]] # 16 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 13], 1, Concat, [1]] # cat head P4
+  - [-1, 2, C3k2, [512, True]] # 19 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 10], 1, Concat, [1]] # cat head P5
+  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 22 (P5/32-large)
+
+  - [[16, 19, 22], 1, Segment26, [nc, 32, 256]] # Segment(P3, P4, P5)

ultralytics/cfg/models/26/yolo26.yaml

@@ -0,0 +1,52 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
+# Model docs: https://docs.ultralytics.com/models/yolo26
+# Task docs: https://docs.ultralytics.com/tasks/detect
+
+# Parameters
+nc: 80 # number of classes
+end2end: True # whether to use end-to-end mode
+reg_max: 1 # DFL bins
+scales: # model compound scaling constants, i.e. 'model=YOLO26n.yaml' will call YOLO26.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 260 layers, 2,572,280 parameters, 2,572,280 gradients, 6.1 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 260 layers, 10,009,784 parameters, 10,009,784 gradients, 22.8 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 280 layers, 21,896,248 parameters, 21,896,248 gradients, 75.4 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 392 layers, 26,299,704 parameters, 26,299,704 gradients, 93.8 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 392 layers, 58,993,368 parameters, 58,993,368 gradients, 209.5 GFLOPs
+
+# YOLO26n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
+  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
+  - [-1, 2, C3k2, [256, False, 0.25]]
+  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
+  - [-1, 2, C3k2, [512, False, 0.25]]
+  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
+  - [-1, 2, C3k2, [512, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
+  - [-1, 2, C3k2, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5, 3, True]] # 9
+  - [-1, 2, C2PSA, [1024]] # 10
+
+# YOLO26n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 2, C3k2, [512, True]] # 13
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 2, C3k2, [256, True]] # 16 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 13], 1, Concat, [1]] # cat head P4
+  - [-1, 2, C3k2, [512, True]] # 19 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 10], 1, Concat, [1]] # cat head P5
+  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 22 (P5/32-large)
+
+  - [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)

ultralytics/cfg/models/26/yoloe-26-seg.yaml

@@ -0,0 +1,53 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
+# Model docs: https://docs.ultralytics.com/models/yolo26
+# Task docs: https://docs.ultralytics.com/tasks/segment
+
+# Parameters
+nc: 80 # number of classes
+end2end: True # whether to use end-to-end mode
+reg_max: 1 # DFL bins
+text_model: mobileclip2:b
+scales: # model compound scaling constants, i.e. 'model=YOLO26n.yaml' will call YOLO26.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 181 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 181 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 231 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 357 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 357 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs
+
+# YOLO26n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
+  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
+  - [-1, 2, C3k2, [256, False, 0.25]]
+  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
+  - [-1, 2, C3k2, [512, False, 0.25]]
+  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
+  - [-1, 2, C3k2, [512, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
+  - [-1, 2, C3k2, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5, 3, True]] # 9
+  - [-1, 2, C2PSA, [1024]] # 10
+
+# YOLO26n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 2, C3k2, [512, True]] # 13
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 2, C3k2, [256, True]] # 16 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 13], 1, Concat, [1]] # cat head P4
+  - [-1, 2, C3k2, [512, True]] # 19 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 10], 1, Concat, [1]] # cat head P5
+  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 22 (P5/32-large)
+
+  - [[16, 19, 22], 1, YOLOESegment26, [nc, 32, 256, 512, True]] # Detect(P3, P4, P5)

ultralytics/cfg/models/26/yoloe-26.yaml

@@ -0,0 +1,53 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26 object detection model with P3/8 - P5/32 outputs
+# Model docs: https://docs.ultralytics.com/models/yolo26
+# Task docs: https://docs.ultralytics.com/tasks/detect
+
+# Parameters
+nc: 80 # number of classes
+end2end: True # whether to use end-to-end mode
+reg_max: 1 # DFL bins
+text_model: mobileclip2:b
+scales: # model compound scaling constants, i.e. 'model=YOLO26n.yaml' will call YOLO26.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 181 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 181 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 231 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 357 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 357 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs
+
+# YOLO26n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
+  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
+  - [-1, 2, C3k2, [256, False, 0.25]]
+  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
+  - [-1, 2, C3k2, [512, False, 0.25]]
+  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
+  - [-1, 2, C3k2, [512, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
+  - [-1, 2, C3k2, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5, 3, True]] # 9
+  - [-1, 2, C2PSA, [1024]] # 10
+
+# YOLO26n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 2, C3k2, [512, True]] # 13
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 2, C3k2, [256, True]] # 16 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 13], 1, Concat, [1]] # cat head P4
+  - [-1, 2, C3k2, [512, True]] # 19 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 10], 1, Concat, [1]] # cat head P5
+  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 22 (P5/32-large)
+
+  - [[16, 19, 22], 1, YOLOEDetect, [nc, 512, True]] # Detect(P3, P4, P5)

ultralytics/data/augment.py

@@ -2062,11 +2062,18 @@ class Format:
             if nl:
                 masks, instances, cls = self._format_segments(instances, cls, w, h)
                 masks = torch.from_numpy(masks)
+                cls_tensor = torch.from_numpy(cls.squeeze(1))
+                if self.mask_overlap:
+                    sem_masks = cls_tensor[masks[0].long() - 1]  # (H, W) from (1, H, W) instance indices
+                else:
+                    sem_masks = (masks * cls_tensor[:, None, None]).max(0).values  # (H, W) from (N, H, W) binary
             else:
                 masks = torch.zeros(
                     1 if self.mask_overlap else nl, img.shape[0] // self.mask_ratio, img.shape[1] // self.mask_ratio
                 )
+                sem_masks = torch.zeros(img.shape[0] // self.mask_ratio, img.shape[1] // self.mask_ratio)
             labels["masks"] = masks
+            labels["sem_masks"] = sem_masks.float()
         labels["img"] = self._format_img(img)
         labels["cls"] = torch.from_numpy(cls) if nl else torch.zeros(nl, 1)
         labels["bboxes"] = torch.from_numpy(instances.bboxes) if nl else torch.zeros((nl, 4))

ultralytics/data/dataset.py

@@ -294,7 +294,7 @@ class YOLODataset(BaseDataset):
         values = list(zip(*[list(b.values()) for b in batch]))
         for i, k in enumerate(keys):
             value = values[i]
-            if k in {"img", "text_feats"}:
+            if k in {"img", "text_feats", "sem_masks"}:
                 value = torch.stack(value, 0)
             elif k == "visuals":
                 value = torch.nn.utils.rnn.pad_sequence(value, batch_first=True)

ultralytics/engine/exporter.py

@@ -503,7 +503,9 @@ class Exporter:
                 m.dynamic = self.args.dynamic
                 m.export = True
                 m.format = self.args.format
-                m.max_det = self.args.max_det
+                # Clamp max_det to anchor count for small image sizes (required for TensorRT compatibility)
+                anchors = sum(int(self.imgsz[0] / s) * int(self.imgsz[1] / s) for s in model.stride.tolist())
+                m.max_det = min(self.args.max_det, anchors)
                 m.xyxy = self.args.nms and not coreml
                 m.shape = None  # reset cached shape for new export input size
                 if hasattr(model, "pe") and hasattr(m, "fuse"):  # for YOLOE models
@@ -551,6 +553,8 @@ class Exporter:
             self.metadata["kpt_shape"] = model.model[-1].kpt_shape
             if hasattr(model, "kpt_names"):
                 self.metadata["kpt_names"] = model.kpt_names
+        if getattr(model.model[-1], "end2end", False):
+            self.metadata["end2end"] = True
 
         LOGGER.info(
             f"\n{colorstr('PyTorch:')} starting from '{file}' with input shape {tuple(im.shape)} BCHW and "
@@ -860,8 +864,11 @@ class Exporter:
     @try_export
     def export_ncnn(self, prefix=colorstr("NCNN:")):
         """Export YOLO model to NCNN format using PNNX https://github.com/pnnx/pnnx."""
-        # use git source for ARM64 due to broken PyPI packages https://github.com/Tencent/ncnn/issues/6509
-        check_requirements("git+https://github.com/Tencent/ncnn.git" if ARM64 else "ncnn", cmds="--no-deps")
+        if ARM64:
+            raise NotImplementedError(
+                "NCNN export is not supported on ARM64"
+            )  # https://github.com/Tencent/ncnn/issues/6509
+        check_requirements("ncnn", cmds="--no-deps")  # no deps to avoid installing opencv-python
         check_requirements("pnnx")
         import ncnn
         import pnnx

ultralytics/engine/model.py

@@ -825,7 +825,7 @@ class Model(torch.nn.Module):
 
             custom = {}  # method defaults
             args = {**self.overrides, **custom, **kwargs, "mode": "train"}  # highest priority args on the right
-            return Tuner(args=args, _callbacks=self.callbacks)(model=self, iterations=iterations)
+            return Tuner(args=args, _callbacks=self.callbacks)(iterations=iterations)
 
     def _apply(self, fn) -> Model:
         """Apply a function to model tensors that are not parameters or registered buffers.

ultralytics/engine/trainer.py

@@ -27,6 +27,7 @@ from ultralytics import __version__
 from ultralytics.cfg import get_cfg, get_save_dir
 from ultralytics.data.utils import check_cls_dataset, check_det_dataset
 from ultralytics.nn.tasks import load_checkpoint
+from ultralytics.optim import MuSGD
 from ultralytics.utils import (
     DEFAULT_CFG,
     GIT,
@@ -464,6 +465,9 @@ class BaseTrainer:
 
                 self.run_callbacks("on_train_batch_end")
 
+            if hasattr(unwrap_model(self.model).criterion, "update"):
+                unwrap_model(self.model).criterion.update()
+
             self.lr = {f"lr/pg{ir}": x["lr"] for ir, x in enumerate(self.optimizer.param_groups)}  # for loggers
 
             self.run_callbacks("on_train_epoch_end")
@@ -930,7 +934,7 @@ class BaseTrainer:
         Returns:
             (torch.optim.Optimizer): The constructed optimizer.
         """
-        g = [], [], []  # optimizer parameter groups
+        g = [{}, {}, {}, {}]  # optimizer parameter groups
         bn = tuple(v for k, v in nn.__dict__.items() if "Norm" in k)  # normalization layers, i.e. BatchNorm2d()
         if name == "auto":
             LOGGER.info(
@@ -940,38 +944,59 @@ class BaseTrainer:
             )
             nc = self.data.get("nc", 10)  # number of classes
             lr_fit = round(0.002 * 5 / (4 + nc), 6)  # lr0 fit equation to 6 decimal places
-            name, lr, momentum = ("SGD", 0.01, 0.9) if iterations > 10000 else ("AdamW", lr_fit, 0.9)
+            name, lr, momentum = ("SGD", 0.01, 0.9) if iterations > 10000 else ("MuSGD", lr_fit, 0.9)
             self.args.warmup_bias_lr = 0.0  # no higher than 0.01 for Adam
 
-        for module_name, module in model.named_modules():
+        use_muon = name == "MuSGD"
+        for module_name, module in unwrap_model(model).named_modules():
             for param_name, param in module.named_parameters(recurse=False):
                 fullname = f"{module_name}.{param_name}" if module_name else param_name
-                if "bias" in fullname:  # bias (no decay)
-                    g[2].append(param)
+                if param.ndim >= 2 and use_muon:
+                    g[3][fullname] = param  # muon params
+                elif "bias" in fullname:  # bias (no decay)
+                    g[2][fullname] = param
                 elif isinstance(module, bn) or "logit_scale" in fullname:  # weight (no decay)
                     # ContrastiveHead and BNContrastiveHead included here with 'logit_scale'
-                    g[1].append(param)
+                    g[1][fullname] = param
                 else:  # weight (with decay)
-                    g[0].append(param)
+                    g[0][fullname] = param
+        if not use_muon:
+            g = [x.values() for x in g[:3]]  # convert to list of params
 
-        optimizers = {"Adam", "Adamax", "AdamW", "NAdam", "RAdam", "RMSProp", "SGD", "auto"}
+        optimizers = {"Adam", "Adamax", "AdamW", "NAdam", "RAdam", "RMSProp", "SGD", "MuSGD", "auto"}
         name = {x.lower(): x for x in optimizers}.get(name.lower())
         if name in {"Adam", "Adamax", "AdamW", "NAdam", "RAdam"}:
-            optimizer = getattr(optim, name, optim.Adam)(g[2], lr=lr, betas=(momentum, 0.999), weight_decay=0.0)
+            optim_args = dict(lr=lr, betas=(momentum, 0.999), weight_decay=0.0)
         elif name == "RMSProp":
-            optimizer = optim.RMSprop(g[2], lr=lr, momentum=momentum)
-        elif name == "SGD":
-            optimizer = optim.SGD(g[2], lr=lr, momentum=momentum, nesterov=True)
+            optim_args = dict(lr=lr, momentum=momentum)
+        elif name == "SGD" or name == "MuSGD":
+            optim_args = dict(lr=lr, momentum=momentum, nesterov=True)
         else:
             raise NotImplementedError(
                 f"Optimizer '{name}' not found in list of available optimizers {optimizers}. "
                 "Request support for addition optimizers at https://github.com/ultralytics/ultralytics."
             )
 
-        optimizer.add_param_group({"params": g[0], "weight_decay": decay})  # add g0 with weight_decay
-        optimizer.add_param_group({"params": g[1], "weight_decay": 0.0})  # add g1 (BatchNorm2d weights)
+        g[2] = {"params": g[2], **optim_args}
+        g[0] = {"params": g[0], **optim_args, "weight_decay": decay}
+        g[1] = {"params": g[1], **optim_args, "weight_decay": 0.0}
+        if name == "MuSGD":
+            g[3] = {"params": g[3], **optim_args, "weight_decay": decay, "use_muon": True}
+            import re
+
+            # higher lr for certain parameters in MuSGD
+            pattern = re.compile(r"(?=.*23)(?=.*cv3)|proto\.semseg|flow_model")
+            g_ = []  # new param groups
+            for x in g:
+                p = x.pop("params")
+                p1 = [v for k, v in p.items() if pattern.search(k)]
+                p2 = [v for k, v in p.items() if not pattern.search(k)]
+                g_.extend([{"params": p1, **x, "lr": lr * 3}, {"params": p2, **x}])
+            g = g_
+        optimizer = getattr(optim, name, MuSGD)(params=g)
+
         LOGGER.info(
             f"{colorstr('optimizer:')} {type(optimizer).__name__}(lr={lr}, momentum={momentum}) with parameter groups "
-            f"{len(g[1])} weight(decay=0.0), {len(g[0])} weight(decay={decay}), {len(g[2])} bias(decay=0.0)"
+            f"{len(g[1]['params'])} weight(decay=0.0), {len(g[0]['params']) if len(g[0]) else len(g[3]['params'])} weight(decay={decay}), {len(g[2]['params'])} bias(decay=0.0)"
         )
         return optimizer

ultralytics/engine/tuner.py

@@ -90,15 +90,15 @@ class Tuner:
         """
         self.space = args.pop("space", None) or {  # key: (min, max, gain(optional))
             # 'optimizer': tune.choice(['SGD', 'Adam', 'AdamW', 'NAdam', 'RAdam', 'RMSProp']),
-            "lr0": (1e-5, 1e-1),  # initial learning rate (i.e. SGD=1E-2, Adam=1E-3)
-            "lrf": (0.0001, 0.1),  # final OneCycleLR learning rate (lr0 * lrf)
+            "lr0": (1e-5, 1e-2),  # initial learning rate (i.e. SGD=1E-2, Adam=1E-3)
+            "lrf": (0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
             "momentum": (0.7, 0.98, 0.3),  # SGD momentum/Adam beta1
             "weight_decay": (0.0, 0.001),  # optimizer weight decay 5e-4
             "warmup_epochs": (0.0, 5.0),  # warmup epochs (fractions ok)
             "warmup_momentum": (0.0, 0.95),  # warmup initial momentum
             "box": (1.0, 20.0),  # box loss gain
             "cls": (0.1, 4.0),  # cls loss gain (scale with pixels)
-            "dfl": (0.4, 6.0),  # dfl loss gain
+            "dfl": (0.4, 12.0),  # dfl loss gain
             "hsv_h": (0.0, 0.1),  # image HSV-Hue augmentation (fraction)
             "hsv_s": (0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
             "hsv_v": (0.0, 0.9),  # image HSV-Value augmentation (fraction)
@@ -254,7 +254,7 @@ class Tuner:
                 f.write(headers)
                 for result in all_results:
                     fitness = result["fitness"]
-                    hyp_values = [result["hyperparameters"][k] for k in self.space.keys()]
+                    hyp_values = [result["hyperparameters"].get(k, self.args.get(k)) for k in self.space.keys()]
                     log_row = [round(fitness, 5), *hyp_values]
                     f.write(",".join(map(str, log_row)) + "\n")
 
@@ -273,6 +273,8 @@ class Tuner:
         parents_mat = np.stack([x[i][1:] for i in idxs], 0)  # (k, ng) strip fitness
         lo, hi = parents_mat.min(0), parents_mat.max(0)
         span = hi - lo
+        # given a small value when span is zero to avoid no mutation
+        span = np.where(span == 0, np.random.uniform(0.01, 0.1, span.shape), span)
         return np.random.uniform(lo - alpha * span, hi + alpha * span)
 
     def _mutate(
@@ -297,7 +299,12 @@ class Tuner:
         if self.mongodb:
             if results := self._get_mongodb_results(n):
                 # MongoDB already sorted by fitness DESC, so results[0] is best
-                x = np.array([[r["fitness"]] + [r["hyperparameters"][k] for k in self.space.keys()] for r in results])
+                x = np.array(
+                    [
+                        [r["fitness"]] + [r["hyperparameters"].get(k, self.args.get(k)) for k in self.space.keys()]
+                        for r in results
+                    ]
+                )
             elif self.collection.name in self.collection.database.list_collection_names():  # Tuner started elsewhere
                 x = np.array([[0.0] + [getattr(self.args, k) for k in self.space.keys()]])
 
@@ -335,10 +342,12 @@ class Tuner:
         # Update types
         if "close_mosaic" in hyp:
             hyp["close_mosaic"] = round(hyp["close_mosaic"])
+        if "epochs" in hyp:
+            hyp["epochs"] = round(hyp["epochs"])
 
         return hyp
 
-    def __call__(self, model=None, iterations: int = 10, cleanup: bool = True):
+    def __call__(self, iterations: int = 10, cleanup: bool = True):
         """Execute the hyperparameter evolution process when the Tuner instance is called.
 
         This method iterates through the specified number of iterations, performing the following steps:
@@ -349,7 +358,6 @@ class Tuner:
         5. Track the best performing configuration across all iterations
 
         Args:
-            model (Model | None, optional): A pre-initialized YOLO model to be used for training.
             iterations (int): The number of generations to run the evolution for.
             cleanup (bool): Whether to delete iteration weights to reduce storage space during tuning.
         """

ultralytics/models/fastsam/predict.py

@@ -63,7 +63,7 @@ class FastSAMPredictor(SegmentationPredictor):
         results = super().postprocess(preds, img, orig_imgs)
         for result in results:
             full_box = torch.tensor(
-                [0, 0, result.orig_shape[1], result.orig_shape[0]], device=preds[0].device, dtype=torch.float32
+                [0, 0, result.orig_shape[1], result.orig_shape[0]], device=result.boxes.data.device, dtype=torch.float32
             )
             boxes = adjust_bboxes_to_image_border(result.boxes.xyxy, result.orig_shape)
             idx = torch.nonzero(box_iou(full_box[None], boxes) > 0.9).flatten()

ultralytics/models/yolo/detect/train.py

@@ -117,10 +117,11 @@ class DetectionTrainer(BaseTrainer):
             if isinstance(v, torch.Tensor):
                 batch[k] = v.to(self.device, non_blocking=self.device.type == "cuda")
         batch["img"] = batch["img"].float() / 255
-        if self.args.multi_scale:
+        multi_scale = self.args.multi_scale
+        if random.random() < multi_scale:
             imgs = batch["img"]
             sz = (
-                random.randrange(int(self.args.imgsz * 0.5), int(self.args.imgsz * 1.5 + self.stride))
+                random.randrange(int(self.args.imgsz * 0.5), int(self.args.imgsz * 1 + self.stride))
                 // self.stride
                 * self.stride
             )  # size

ultralytics/models/yolo/detect/val.py

@@ -494,6 +494,12 @@ class DetectionValidator(BaseValidator):
                     # update mAP50-95 and mAP50
                     stats[f"metrics/mAP50({suffix[i][0]})"] = val.stats_as_dict["AP_50"]
                     stats[f"metrics/mAP50-95({suffix[i][0]})"] = val.stats_as_dict["AP_all"]
+                    # record mAP for small, medium, large objects as well
+                    stats["metrics/mAP_small(B)"] = val.stats_as_dict["AP_small"]
+                    stats["metrics/mAP_medium(B)"] = val.stats_as_dict["AP_medium"]
+                    stats["metrics/mAP_large(B)"] = val.stats_as_dict["AP_large"]
+                    # update fitness
+                    stats["fitness"] = 0.9 * val.stats_as_dict["AP_all"] + 0.1 * val.stats_as_dict["AP_50"]
 
                     if self.is_lvis:
                         stats[f"metrics/APr({suffix[i][0]})"] = val.stats_as_dict["APr"]

ultralytics/models/yolo/model.py

@@ -399,7 +399,7 @@ class YOLOE(Model):
                         "batch": 1,
                         "device": kwargs.get("device", None),
                         "half": kwargs.get("half", False),
-                        "imgsz": kwargs.get("imgsz", self.overrides["imgsz"]),
+                        "imgsz": kwargs.get("imgsz", self.overrides.get("imgsz", 640)),
                     },
                     _callbacks=self.callbacks,
                 )

ultralytics/models/yolo/obb/predict.py

@@ -50,7 +50,7 @@ class OBBPredictor(DetectionPredictor):
             (Results): The result object containing the original image, image path, class names, and oriented bounding
                 boxes.
         """
-        rboxes = ops.regularize_rboxes(torch.cat([pred[:, :4], pred[:, -1:]], dim=-1))
+        rboxes = torch.cat([pred[:, :4], pred[:, -1:]], dim=-1)
         rboxes[:, :4] = ops.scale_boxes(img.shape[2:], rboxes[:, :4], orig_img.shape, xywh=True)
         obb = torch.cat([rboxes, pred[:, 4:6]], dim=-1)
         return Results(orig_img, path=img_path, names=self.model.names, obb=obb)

ultralytics/models/yolo/obb/train.py

@@ -73,7 +73,7 @@ class OBBTrainer(yolo.detect.DetectionTrainer):
 
     def get_validator(self):
         """Return an instance of OBBValidator for validation of YOLO model."""
-        self.loss_names = "box_loss", "cls_loss", "dfl_loss"
+        self.loss_names = "box_loss", "cls_loss", "dfl_loss", "angle_loss"
         return yolo.obb.OBBValidator(
             self.test_loader, save_dir=self.save_dir, args=copy(self.args), _callbacks=self.callbacks
         )

ultralytics/models/yolo/pose/train.py

@@ -90,7 +90,7 @@ class PoseTrainer(yolo.detect.DetectionTrainer):
 
     def get_validator(self):
         """Return an instance of the PoseValidator class for validation."""
-        self.loss_names = "box_loss", "pose_loss", "kobj_loss", "cls_loss", "dfl_loss"
+        self.loss_names = "box_loss", "pose_loss", "kobj_loss", "cls_loss", "dfl_loss", "rle_loss"
         return yolo.pose.PoseValidator(
             self.test_loader, save_dir=self.save_dir, args=copy(self.args), _callbacks=self.callbacks
         )

ultralytics/models/yolo/segment/predict.py

@@ -60,7 +60,7 @@ class SegmentationPredictor(DetectionPredictor):
             >>> results = predictor.postprocess(preds, img, orig_img)
         """
         # Extract protos - tuple if PyTorch model or array if exported
-        protos = preds[1][-1] if isinstance(preds[1], tuple) else preds[1]
+        protos = preds[0][-1] if isinstance(preds[0], tuple) else preds[-1]
         return super().postprocess(preds[0], img, orig_imgs, protos=protos)
 
     def construct_results(self, preds, img, orig_imgs, protos):

ultralytics/models/yolo/segment/train.py

@@ -63,7 +63,7 @@ class SegmentationTrainer(yolo.detect.DetectionTrainer):
 
     def get_validator(self):
         """Return an instance of SegmentationValidator for validation of YOLO model."""
-        self.loss_names = "box_loss", "seg_loss", "cls_loss", "dfl_loss"
+        self.loss_names = "box_loss", "seg_loss", "cls_loss", "dfl_loss", "sem_loss"
         return yolo.segment.SegmentationValidator(
             self.test_loader, save_dir=self.save_dir, args=copy(self.args), _callbacks=self.callbacks
         )

ultralytics/models/yolo/segment/val.py

@@ -99,7 +99,9 @@ class SegmentationValidator(DetectionValidator):
         Returns:
             list[dict[str, torch.Tensor]]: Processed detection predictions with masks.
         """
-        proto = preds[1][-1] if len(preds[1]) == 3 else preds[1]  # second output is len 3 if pt, but only 1 if exported
+        proto = (
+            preds[0][-1] if isinstance(preds[0], tuple) else preds[-1]
+        )  # second output is len 3 if pt, but only 1 if exported
         preds = super().postprocess(preds[0])
         imgsz = [4 * x for x in proto.shape[2:]]  # get image size from proto
         for i, pred in enumerate(preds):

ultralytics/models/yolo/yoloe/train.py

@@ -147,7 +147,12 @@ class YOLOEPETrainer(DetectionTrainer):
         model.model[-1].cv3[0][2] = deepcopy(model.model[-1].cv3[0][2]).requires_grad_(True)
         model.model[-1].cv3[1][2] = deepcopy(model.model[-1].cv3[1][2]).requires_grad_(True)
         model.model[-1].cv3[2][2] = deepcopy(model.model[-1].cv3[2][2]).requires_grad_(True)
-        del model.pe
+
+        if getattr(model.model[-1], "one2one_cv3", None) is not None:
+            model.model[-1].one2one_cv3[0][2] = deepcopy(model.model[-1].cv3[0][2]).requires_grad_(True)
+            model.model[-1].one2one_cv3[1][2] = deepcopy(model.model[-1].cv3[1][2]).requires_grad_(True)
+            model.model[-1].one2one_cv3[2][2] = deepcopy(model.model[-1].cv3[2][2]).requires_grad_(True)
+
         model.train()
 
         return model