ultralytics-opencv-headless 8.3.253__py3-none-any.whl → 8.4.2__py3-none-any.whl

ultralytics/cfg/models/26/yolo26-p6.yaml

@@ -0,0 +1,62 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26 object detection model with P3/8 - P6/64 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-p6.yaml' will call yolo26-p6.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 349 layers, 4,063,872 parameters, 4,063,872 gradients, 6.0 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 349 layers, 15,876,448 parameters, 15,876,448 gradients, 22.3 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 369 layers, 32,400,096 parameters, 32,400,096 gradients, 77.3 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 523 layers, 39,365,600 parameters, 39,365,600 gradients, 97.0 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 523 layers, 88,330,368 parameters, 88,330,368 gradients, 216.6 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, [768, 3, 2]] # 7-P5/32
+  - [-1, 2, C3k2, [768, True]]
+  - [-1, 1, Conv, [1024, 3, 2]] # 9-P6/64
+  - [-1, 2, C3k2, [1024, True]]
+  - [-1, 1, SPPF, [1024, 5]] # 11
+  - [-1, 2, C2PSA, [1024]] # 12
+
+# YOLO26n head
+head:
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 8], 1, Concat, [1]] # cat backbone P5
+  - [-1, 2, C3k2, [768, True]] # 15
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
+  - [-1, 2, C3k2, [512, True]] # 18
+
+  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
+  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
+  - [-1, 2, C3k2, [256, True]] # 21 (P3/8-small)
+
+  - [-1, 1, Conv, [256, 3, 2]]
+  - [[-1, 18], 1, Concat, [1]] # cat head P4
+  - [-1, 2, C3k2, [512, True]] # 24 (P4/16-medium)
+
+  - [-1, 1, Conv, [512, 3, 2]]
+  - [[-1, 15], 1, Concat, [1]] # cat head P5
+  - [-1, 2, C3k2, [768, True]] # 27 (P5/32-large)
+
+  - [-1, 1, Conv, [768, 3, 2]]
+  - [[-1, 12], 1, Concat, [1]] # cat head P6
+  - [-1, 1, C3k2, [1024, True, 0.5, True]] # 30 (P6/64-large)
+
+  - [[21, 24, 27, 30], 1, Detect, [nc]] # Detect(P3, P4, P5, P6)

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

@@ -0,0 +1,53 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLO26-pose keypoints/pose estimation 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-pose.yaml' will call yolo26-pose.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-seg instance segmentation 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-seg.yaml' will call yolo26-seg.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 YOLOE-26 open-vocabulary instance segmentation 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=yoloe-26n-seg.yaml' will call yoloe-26-seg.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 347 layers, 5,615,540 parameters, 5,615,540 gradients, 11.7 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 347 layers, 15,272,852 parameters, 15,272,852 gradients, 39.3 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 367 layers, 34,922,132 parameters, 34,922,132 gradients, 136.3 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 479 layers, 39,325,588 parameters, 39,325,588 gradients, 154.7 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 479 layers, 85,397,684 parameters, 85,397,684 gradients, 343.3 GFLOPs
+
+# YOLOE26n 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
+
+# YOLOE26n 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]] # YOLOESegment26(P3, P4, P5)

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

@@ -0,0 +1,53 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Ultralytics YOLOE-26 open-vocabulary 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=yoloe-26n.yaml' will call yoloe-26.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.50, 0.25, 1024] # summary: 298 layers, 5,061,540 parameters, 5,061,540 gradients, 7.3 GFLOPs
+  s: [0.50, 0.50, 1024] # summary: 298 layers, 13,776,836 parameters, 13,776,836 gradients, 24.8 GFLOPs
+  m: [0.50, 1.00, 512] # summary: 318 layers, 29,706,308 parameters, 29,706,308 gradients, 79.2 GFLOPs
+  l: [1.00, 1.00, 512] # summary: 430 layers, 34,109,764 parameters, 34,109,764 gradients, 97.6 GFLOPs
+  x: [1.00, 1.50, 512] # summary: 430 layers, 73,697,252 parameters, 73,697,252 gradients, 215.2 GFLOPs
+
+# YOLOE26n 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
+
+# YOLOE26n 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]] # YOLOEDetect(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/converter.py

@@ -749,12 +749,13 @@ def convert_to_multispectral(path: str | Path, n_channels: int = 10, replace: bo
 async def convert_ndjson_to_yolo(ndjson_path: str | Path, output_path: str | Path | None = None) -> Path:
     """Convert NDJSON dataset format to Ultralytics YOLO11 dataset structure.
 
-    This function converts datasets stored in NDJSON (Newline Delimited JSON) format to the standard YOLO format with
-    separate directories for images and labels. It supports parallel processing for efficient conversion of large
-    datasets and can download images from URLs if they don't exist locally.
+    This function converts datasets stored in NDJSON (Newline Delimited JSON) format to the standard YOLO format. For
+    detection/segmentation/pose/obb tasks, it creates separate directories for images and labels. For classification
+    tasks, it creates the ImageNet-style {split}/{class_name}/ folder structure. It supports parallel processing for
+    efficient conversion of large datasets and can download images from URLs.
 
     The NDJSON format consists of:
-    - First line: Dataset metadata with class names and configuration
+    - First line: Dataset metadata with class names, task type, and configuration
     - Subsequent lines: Individual image records with annotations and optional URLs
 
     Args:
@@ -763,7 +764,7 @@ async def convert_ndjson_to_yolo(ndjson_path: str | Path, output_path: str | Pat
             None, uses the parent directory of the NDJSON file. Defaults to None.
 
     Returns:
-        (Path): Path to the generated data.yaml file that can be used for YOLO training.
+        (Path): Path to the generated data.yaml file (detection) or dataset directory (classification).
 
     Examples:
         Convert a local NDJSON file:
@@ -790,36 +791,51 @@ async def convert_ndjson_to_yolo(ndjson_path: str | Path, output_path: str | Pat
     dataset_dir = output_path / ndjson_path.stem
     splits = {record["split"] for record in image_records}
 
-    # Create directories and prepare YAML structure
-    dataset_dir.mkdir(parents=True, exist_ok=True)
-    data_yaml = dict(dataset_record)
-    data_yaml["names"] = {int(k): v for k, v in dataset_record.get("class_names", {}).items()}
-    data_yaml.pop("class_names")
+    # Check if this is a classification dataset
+    is_classification = dataset_record.get("task") == "classify"
+    class_names = {int(k): v for k, v in dataset_record.get("class_names", {}).items()}
 
-    for split in sorted(splits):
-        (dataset_dir / "images" / split).mkdir(parents=True, exist_ok=True)
-        (dataset_dir / "labels" / split).mkdir(parents=True, exist_ok=True)
-        data_yaml[split] = f"images/{split}"
+    # Create base directories
+    dataset_dir.mkdir(parents=True, exist_ok=True)
+    data_yaml = None
+
+    if not is_classification:
+        # Detection/segmentation/pose/obb: prepare YAML and create base structure
+        data_yaml = dict(dataset_record)
+        data_yaml["names"] = class_names
+        data_yaml.pop("class_names", None)
+        data_yaml.pop("type", None)  # Remove NDJSON-specific fields
+        for split in sorted(splits):
+            (dataset_dir / "images" / split).mkdir(parents=True, exist_ok=True)
+            (dataset_dir / "labels" / split).mkdir(parents=True, exist_ok=True)
+            data_yaml[split] = f"images/{split}"
 
     async def process_record(session, semaphore, record):
         """Process single image record with async session."""
         async with semaphore:
             split, original_name = record["split"], record["file"]
-            label_path = dataset_dir / "labels" / split / f"{Path(original_name).stem}.txt"
-            image_path = dataset_dir / "images" / split / original_name
-
             annotations = record.get("annotations", {})
-            lines_to_write = []
-            for key in annotations.keys():
-                lines_to_write = [" ".join(map(str, item)) for item in annotations[key]]
-                break
-            if "classification" in annotations:
-                lines_to_write = [str(cls) for cls in annotations["classification"]]
-
-            label_path.write_text("\n".join(lines_to_write) + "\n" if lines_to_write else "")
 
+            if is_classification:
+                # Classification: place image in {split}/{class_name}/ folder
+                class_ids = annotations.get("classification", [])
+                class_id = class_ids[0] if class_ids else 0
+                class_name = class_names.get(class_id, str(class_id))
+                image_path = dataset_dir / split / class_name / original_name
+            else:
+                # Detection: write label file and place image in images/{split}/
+                image_path = dataset_dir / "images" / split / original_name
+                label_path = dataset_dir / "labels" / split / f"{Path(original_name).stem}.txt"
+                lines_to_write = []
+                for key in annotations.keys():
+                    lines_to_write = [" ".join(map(str, item)) for item in annotations[key]]
+                    break
+                label_path.write_text("\n".join(lines_to_write) + "\n" if lines_to_write else "")
+
+            # Download image if URL provided and file doesn't exist
             if http_url := record.get("url"):
                 if not image_path.exists():
+                    image_path.parent.mkdir(parents=True, exist_ok=True)  # Ensure parent dir exists
                     try:
                         async with session.get(http_url, timeout=aiohttp.ClientTimeout(total=30)) as response:
                             response.raise_for_status()
@@ -848,8 +864,11 @@ async def convert_ndjson_to_yolo(ndjson_path: str | Path, output_path: str | Pat
         await asyncio.gather(*[tracked_process(record) for record in image_records])
         pbar.close()
 
-    # Write data.yaml
-    yaml_path = dataset_dir / "data.yaml"
-    YAML.save(yaml_path, data_yaml)
-
-    return yaml_path
+    if is_classification:
+        # Classification: return dataset directory (check_cls_dataset expects a directory path)
+        return dataset_dir
+    else:
+        # Detection: write data.yaml and return its path
+        yaml_path = dataset_dir / "data.yaml"
+        YAML.save(yaml_path, data_yaml)
+        return yaml_path

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

@@ -463,6 +463,9 @@ class Exporter:
             )
         if tfjs and (ARM64 and LINUX):
             raise SystemError("TF.js exports are not currently supported on ARM64 Linux")
+        if ncnn and hasattr(model.model[-1], "one2one_cv2"):
+            del model.model[-1].one2one_cv2  # Disable end2end branch for NCNN export as it does not support topk
+            LOGGER.warning("NCNN export does not support end2end models, disabling end2end branch.")
         # Recommend OpenVINO if export and Intel CPU
         if SETTINGS.get("openvino_msg"):
             if is_intel():
@@ -503,7 +506,9 @@ 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 +556,8 @@ 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 "
@@ -787,7 +794,6 @@ class Exporter:
                         f".*{head_module_name}/.*/Sub*",
                         f".*{head_module_name}/.*/Mul*",
                         f".*{head_module_name}/.*/Div*",
-                        f".*{head_module_name}\\.dfl.*",
                     ],
                     types=["Sigmoid"],
                 )
@@ -860,8 +866,7 @@ 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")
+        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/results.py

@@ -750,8 +750,8 @@ class Results(SimpleClass, DataExportMixin):
         """Convert inference results to a summarized dictionary with optional normalization for box coordinates.
 
         This method creates a list of detection dictionaries, each containing information about a single detection or
-        classification result. For classification tasks, it returns the top class and its
-        confidence. For detection tasks, it includes class information, bounding box coordinates, and
+        classification result. For classification tasks, it returns the top 5 classes and their
+        confidences. For detection tasks, it includes class information, bounding box coordinates, and
         optionally mask segments and keypoints.
 
         Args:
@@ -772,14 +772,23 @@ class Results(SimpleClass, DataExportMixin):
         # Create list of detection dictionaries
         results = []
         if self.probs is not None:
-            class_id = self.probs.top1
-            results.append(
-                {
-                    "name": self.names[class_id],
-                    "class": class_id,
-                    "confidence": round(self.probs.top1conf.item(), decimals),
-                }
-            )
+            # Return top 5 classification results
+            for class_id, conf in zip(self.probs.top5, self.probs.top5conf.tolist()):
+                class_id = int(class_id)
+                results.append(
+                    {
+                        "name": self.names[class_id],
+                        "class": class_id,
+                        "confidence": round(conf, decimals),
+                    }
+                )
+                results.append(
+                    {
+                        "name": self.names[class_id],
+                        "class": class_id,
+                        "confidence": round(conf, decimals),
+                    }
+                )
             return results
 
         is_obb = self.obb is not None