ultralytics-opencv-headless 8.3.251__py3-none-any.whl → 8.4.1__py3-none-any.whl

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

ultralytics/models/yolo/yoloe/train_seg.py

@@ -104,7 +104,12 @@ class YOLOEPESegTrainer(SegmentationTrainer):
         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

ultralytics/nn/autobackend.py

@@ -221,6 +221,7 @@ class AutoBackend(nn.Module):
             for p in model.parameters():
                 p.requires_grad = False
             self.model = model  # explicitly assign for to(), cpu(), cuda(), half()
+            end2end = getattr(model, "end2end", False)
 
         # TorchScript
         elif jit:
@@ -497,11 +498,11 @@ class AutoBackend(nn.Module):
         elif paddle:
             LOGGER.info(f"Loading {w} for PaddlePaddle inference...")
             check_requirements(
-                "paddlepaddle-gpu"
+                "paddlepaddle-gpu>=3.0.0,!=3.3.0"  # exclude 3.3.0 https://github.com/PaddlePaddle/Paddle/issues/77340
                 if torch.cuda.is_available()
                 else "paddlepaddle==3.0.0"  # pin 3.0.0 for ARM64
                 if ARM64
-                else "paddlepaddle>=3.0.0"
+                else "paddlepaddle>=3.0.0,!=3.3.0"  # exclude 3.3.0 https://github.com/PaddlePaddle/Paddle/issues/77340
             )
             import paddle.inference as pdi
 
@@ -545,11 +546,16 @@ class AutoBackend(nn.Module):
         # NCNN
         elif ncnn:
             LOGGER.info(f"Loading {w} for NCNN inference...")
-            check_requirements("git+https://github.com/Tencent/ncnn.git" if ARM64 else "ncnn", cmds="--no-deps")
+            check_requirements("ncnn", cmds="--no-deps")
             import ncnn as pyncnn
 
             net = pyncnn.Net()
-            net.opt.use_vulkan_compute = cuda
+            if isinstance(cuda, torch.device):
+                net.opt.use_vulkan_compute = cuda
+            elif isinstance(device, str) and device.startswith("vulkan"):
+                net.opt.use_vulkan_compute = True
+                net.set_vulkan_device(int(device.split(":")[1]))
+                device = torch.device("cpu")
             w = Path(w)
             if not w.is_file():  # if not *.param
                 w = next(w.glob("*.param"))  # get *.param file from *_ncnn_model dir
@@ -651,7 +657,7 @@ class AutoBackend(nn.Module):
             names = metadata["names"]
             kpt_shape = metadata.get("kpt_shape")
             kpt_names = metadata.get("kpt_names")
-            end2end = metadata.get("args", {}).get("nms", False)
+            end2end = metadata.get("end2end", False) or metadata.get("args", {}).get("nms", False)
             dynamic = metadata.get("args", {}).get("dynamic", dynamic)
             ch = metadata.get("channels", 3)
         elif not (pt or triton or nn_module):

ultralytics/nn/modules/__init__.py

@@ -78,15 +78,19 @@ from .conv import (
 )
 from .head import (
     OBB,
+    OBB26,
     Classify,
     Detect,
     LRPCHead,
     Pose,
+    Pose26,
     RTDETRDecoder,
     Segment,
+    Segment26,
     WorldDetect,
     YOLOEDetect,
     YOLOESegment,
+    YOLOESegment26,
     v10Detect,
 )
 from .transformer import (
@@ -115,6 +119,7 @@ __all__ = (
     "ELAN1",
     "MLP",
     "OBB",
+    "OBB26",
     "PSA",
     "SPP",
     "SPPELAN",
@@ -161,6 +166,7 @@ __all__ = (
     "MSDeformAttn",
     "MaxSigmoidAttnBlock",
     "Pose",
+    "Pose26",
     "Proto",
     "RTDETRDecoder",
     "RepC3",
@@ -170,6 +176,7 @@ __all__ = (
     "ResNetLayer",
     "SCDown",
     "Segment",
+    "Segment26",
     "SpatialAttention",
     "TorchVision",
     "TransformerBlock",
@@ -178,5 +185,6 @@ __all__ = (
     "WorldDetect",
     "YOLOEDetect",
     "YOLOESegment",
+    "YOLOESegment26",
     "v10Detect",
 )

ultralytics/nn/modules/block.py

@@ -208,28 +208,33 @@ class SPP(nn.Module):
 class SPPF(nn.Module):
     """Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher."""
 
-    def __init__(self, c1: int, c2: int, k: int = 5):
+    def __init__(self, c1: int, c2: int, k: int = 5, n: int = 3, shortcut: bool = False):
         """Initialize the SPPF layer with given input/output channels and kernel size.
 
         Args:
             c1 (int): Input channels.
             c2 (int): Output channels.
             k (int): Kernel size.
+            n (int): Number of pooling iterations.
+            shortcut (bool): Whether to use shortcut connection.
 
         Notes:
             This module is equivalent to SPP(k=(5, 9, 13)).
         """
         super().__init__()
         c_ = c1 // 2  # hidden channels
-        self.cv1 = Conv(c1, c_, 1, 1)
-        self.cv2 = Conv(c_ * 4, c2, 1, 1)
+        self.cv1 = Conv(c1, c_, 1, 1, act=False)
+        self.cv2 = Conv(c_ * (n + 1), c2, 1, 1)
         self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
+        self.n = n
+        self.add = shortcut and c1 == c2
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         """Apply sequential pooling operations to input and return concatenated feature maps."""
         y = [self.cv1(x)]
-        y.extend(self.m(y[-1]) for _ in range(3))
-        return self.cv2(torch.cat(y, 1))
+        y.extend(self.m(y[-1]) for _ in range(getattr(self, "n", 3)))
+        y = self.cv2(torch.cat(y, 1))
+        return y + x if getattr(self, "add", False) else y
 
 
 class C1(nn.Module):
@@ -1065,7 +1070,15 @@ class C3k2(C2f):
     """Faster Implementation of CSP Bottleneck with 2 convolutions."""
 
     def __init__(
-        self, c1: int, c2: int, n: int = 1, c3k: bool = False, e: float = 0.5, g: int = 1, shortcut: bool = True
+        self,
+        c1: int,
+        c2: int,
+        n: int = 1,
+        c3k: bool = False,
+        e: float = 0.5,
+        attn: bool = False,
+        g: int = 1,
+        shortcut: bool = True,
     ):
         """Initialize C3k2 module.
 
@@ -1075,12 +1088,21 @@ class C3k2(C2f):
             n (int): Number of blocks.
             c3k (bool): Whether to use C3k blocks.
             e (float): Expansion ratio.
+            attn (bool): Whether to use attention blocks.
             g (int): Groups for convolutions.
             shortcut (bool): Whether to use shortcut connections.
         """
         super().__init__(c1, c2, n, shortcut, g, e)
         self.m = nn.ModuleList(
-            C3k(self.c, self.c, 2, shortcut, g) if c3k else Bottleneck(self.c, self.c, shortcut, g) for _ in range(n)
+            nn.Sequential(
+                Bottleneck(self.c, self.c, shortcut, g),
+                PSABlock(self.c, attn_ratio=0.5, num_heads=max(self.c // 64, 1)),
+            )
+            if attn
+            else C3k(self.c, self.c, 2, shortcut, g)
+            if c3k
+            else Bottleneck(self.c, self.c, shortcut, g)
+            for _ in range(n)
         )
 
 
@@ -1148,6 +1170,8 @@ class RepVGGDW(torch.nn.Module):
 
         This method fuses the convolutional layers and updates the weights and biases accordingly.
         """
+        if not hasattr(self, "conv1"):
+            return  # already fused
         conv = fuse_conv_and_bn(self.conv.conv, self.conv.bn)
         conv1 = fuse_conv_and_bn(self.conv1.conv, self.conv1.bn)
 
@@ -1391,7 +1415,7 @@ class PSA(nn.Module):
         self.cv1 = Conv(c1, 2 * self.c, 1, 1)
         self.cv2 = Conv(2 * self.c, c1, 1)
 
-        self.attn = Attention(self.c, attn_ratio=0.5, num_heads=self.c // 64)
+        self.attn = Attention(self.c, attn_ratio=0.5, num_heads=max(self.c // 64, 1))
         self.ffn = nn.Sequential(Conv(self.c, self.c * 2, 1), Conv(self.c * 2, self.c, 1, act=False))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -1945,3 +1969,99 @@ class SAVPE(nn.Module):
         aggregated = score.transpose(-2, -3) @ x.reshape(B, self.c, C // self.c, -1).transpose(-1, -2)
 
         return F.normalize(aggregated.transpose(-2, -3).reshape(B, Q, -1), dim=-1, p=2)
+
+
+class Proto26(Proto):
+    """Ultralytics YOLO26 models mask Proto module for segmentation models."""
+
+    def __init__(self, ch: tuple = (), c_: int = 256, c2: int = 32, nc: int = 80):
+        """Initialize the Ultralytics YOLO models mask Proto module with specified number of protos and masks.
+
+        Args:
+            ch (tuple): Tuple of channel sizes from backbone feature maps.
+            c_ (int): Intermediate channels.
+            c2 (int): Output channels (number of protos).
+            nc (int): Number of classes for semantic segmentation.
+        """
+        super().__init__(c_, c_, c2)
+        self.feat_refine = nn.ModuleList(Conv(x, ch[0], k=1) for x in ch[1:])
+        self.feat_fuse = Conv(ch[0], c_, k=3)
+        self.semseg = nn.Sequential(Conv(ch[0], c_, k=3), Conv(c_, c_, k=3), nn.Conv2d(c_, nc, 1))
+
+    def forward(self, x: torch.Tensor, return_semseg: bool = True) -> torch.Tensor:
+        """Perform a forward pass through layers using an upsampled input image."""
+        feat = x[0]
+        for i, f in enumerate(self.feat_refine):
+            up_feat = f(x[i + 1])
+            up_feat = F.interpolate(up_feat, size=feat.shape[2:], mode="nearest")
+            feat = feat + up_feat
+        p = super().forward(self.feat_fuse(feat))
+        if self.training and return_semseg:
+            semseg = self.semseg(feat)
+            return (p, semseg)
+        return p
+
+    def fuse(self):
+        """Fuse the model for inference by removing the semantic segmentation head."""
+        self.semseg = None
+
+
+class RealNVP(nn.Module):
+    """RealNVP: a flow-based generative model.
+
+    References:
+        https://arxiv.org/abs/1605.08803
+        https://github.com/open-mmlab/mmpose/blob/main/mmpose/models/utils/realnvp.py
+    """
+
+    @staticmethod
+    def nets():
+        """Get the scale model in a single invertable mapping."""
+        return nn.Sequential(nn.Linear(2, 64), nn.SiLU(), nn.Linear(64, 64), nn.SiLU(), nn.Linear(64, 2), nn.Tanh())
+
+    @staticmethod
+    def nett():
+        """Get the translation model in a single invertable mapping."""
+        return nn.Sequential(nn.Linear(2, 64), nn.SiLU(), nn.Linear(64, 64), nn.SiLU(), nn.Linear(64, 2))
+
+    @property
+    def prior(self):
+        """The prior distribution."""
+        return torch.distributions.MultivariateNormal(self.loc, self.cov)
+
+    def __init__(self):
+        super().__init__()
+
+        self.register_buffer("loc", torch.zeros(2))
+        self.register_buffer("cov", torch.eye(2))
+        self.register_buffer("mask", torch.tensor([[0, 1], [1, 0]] * 3, dtype=torch.float32))
+
+        self.s = torch.nn.ModuleList([self.nets() for _ in range(len(self.mask))])
+        self.t = torch.nn.ModuleList([self.nett() for _ in range(len(self.mask))])
+        self.init_weights()
+
+    def init_weights(self):
+        """Initialization model weights."""
+        for m in self.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.xavier_uniform_(m.weight, gain=0.01)
+
+    def backward_p(self, x):
+        """Apply mapping form the data space to the latent space and calculate the log determinant of the Jacobian
+        matrix.
+        """
+        log_det_jacob, z = x.new_zeros(x.shape[0]), x
+        for i in reversed(range(len(self.t))):
+            z_ = self.mask[i] * z
+            s = self.s[i](z_) * (1 - self.mask[i])
+            t = self.t[i](z_) * (1 - self.mask[i])
+            z = (1 - self.mask[i]) * (z - t) * torch.exp(-s) + z_
+            log_det_jacob -= s.sum(dim=1)
+        return z, log_det_jacob
+
+    def log_prob(self, x):
+        """Calculate the log probability of given sample in data space."""
+        if x.dtype == torch.float32 and self.s[0][0].weight.dtype != torch.float32:
+            self.float()
+        z, log_det = self.backward_p(x)
+        return self.prior.log_prob(z) + log_det