ultralytics 8.3.59 Add ability to load any torchvision model as module (#18564)

Signed-off-by: Glenn Jocher <glenn.jocher@ultralytics.com>
Co-authored-by: UltralyticsAssistant <web@ultralytics.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

docs/en/reference/nn/modules/block.md

@@ -189,4 +189,8 @@ keywords: Ultralytics, YOLO, neural networks, block modules, DFL, Proto, HGStem,
 
 ## ::: ultralytics.nn.modules.block.SCDown
 
+<br><br><hr><br>
+
+## ::: ultralytics.nn.modules.block.TorchVision
+
 <br><br>

docs/en/reference/nn/modules/conv.md

@@ -63,6 +63,10 @@ keywords: Ultralytics, convolution modules, Conv, LightConv, GhostConv, YOLO, de
 
 <br><br><hr><br>
 
+## ::: ultralytics.nn.modules.conv.Index
+
+<br><br><hr><br>
+
 ## ::: ultralytics.nn.modules.conv.autopad
 
 <br><br>

ultralytics/__init__.py

@@ -1,6 +1,6 @@
 # Ultralytics YOLO 🚀, AGPL-3.0 license
 
-__version__ = "8.3.58"
+__version__ = "8.3.59"
 
 import os
 

ultralytics/cfg/models/11/yolo11-cls-resnet18.yaml

@@ -0,0 +1,21 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+# YOLO11-cls image classification model. For Usage examples see https://docs.ultralytics.com/tasks/classify
+
+# Parameters
+nc: 10 # number of classes
+scales: # model compound scaling constants, i.e. 'model=yolo11n-cls.yaml' will call yolo11-cls.yaml with scale 'n'
+  # [depth, width, max_channels]
+  n: [0.33, 0.25, 1024]
+  s: [0.33, 0.50, 1024]
+  m: [0.67, 0.75, 1024]
+  l: [1.00, 1.00, 1024]
+  x: [1.00, 1.25, 1024]
+
+# YOLO11n backbone
+backbone:
+  # [from, repeats, module, args]
+  - [-1, 1, TorchVision, [512, "resnet18", "DEFAULT", True, 2]] # truncate two layers from the end
+
+# YOLO11n head
+head:
+  - [-1, 1, Classify, [nc]] # Classify

ultralytics/nn/modules/__init__.py

@@ -56,6 +56,7 @@ from .block import (
     RepVGGDW,
     ResNetLayer,
     SCDown,
+    TorchVision,
 )
 from .conv import (
     CBAM,
@@ -68,6 +69,7 @@ from .conv import (
     DWConvTranspose2d,
     Focus,
     GhostConv,
+    Index,
     LightConv,
     RepConv,
     SpatialAttention,
@@ -156,4 +158,6 @@ __all__ = (
     "C2fCIB",
     "Attention",
     "PSA",
+    "TorchVision",
+    "Index",
 )

ultralytics/nn/modules/block.py

@@ -49,6 +49,7 @@ __all__ = (
     "Attention",
     "PSA",
     "SCDown",
+    "TorchVision",
 )
 
 
@@ -1107,3 +1108,51 @@ class SCDown(nn.Module):
     def forward(self, x):
         """Applies convolution and downsampling to the input tensor in the SCDown module."""
         return self.cv2(self.cv1(x))
+
+
+class TorchVision(nn.Module):
+    """
+    TorchVision module to allow loading any torchvision model.
+
+    This class provides a way to load a model from the torchvision library, optionally load pre-trained weights, and customize the model by truncating or unwrapping layers.
+
+    Attributes:
+        m (nn.Module): The loaded torchvision model, possibly truncated and unwrapped.
+
+    Args:
+        c1 (int): Input channels.
+        c2 (): Output channels.
+        model (str): Name of the torchvision model to load.
+        weights (str, optional): Pre-trained weights to load. Default is "DEFAULT".
+        unwrap (bool, optional): If True, unwraps the model to a sequential containing all but the last `truncate` layers. Default is True.
+        truncate (int, optional): Number of layers to truncate from the end if `unwrap` is True. Default is 2.
+        split (bool, optional): Returns output from intermediate child modules as list. Default is False.
+    """
+
+    def __init__(self, c1, c2, model, weights="DEFAULT", unwrap=True, truncate=2, split=False):
+        """Load the model and weights from torchvision."""
+        import torchvision
+
+        super().__init__()
+        if hasattr(torchvision.models, "get_model"):
+            self.m = torchvision.models.get_model(model, weights=weights)
+        else:
+            self.m = torchvision.models.__dict__[model](pretrained=bool(weights))
+        if unwrap:
+            layers = list(self.m.children())[:-truncate]
+            if isinstance(layers[0], nn.Sequential):  # Second-level for some models like EfficientNet, Swin
+                layers = [*list(layers[0].children()), *layers[1:]]
+            self.m = nn.Sequential(*layers)
+            self.split = split
+        else:
+            self.split = False
+            self.m.head = self.m.heads = nn.Identity()
+
+    def forward(self, x):
+        """Forward pass through the model."""
+        if self.split:
+            y = [x]
+            y.extend(m(y[-1]) for m in self.m)
+        else:
+            y = self.m(x)
+        return y

ultralytics/nn/modules/conv.py

@@ -21,6 +21,7 @@ __all__ = (
     "CBAM",
     "Concat",
     "RepConv",
+    "Index",
 )
 
 
@@ -330,3 +331,20 @@ class Concat(nn.Module):
     def forward(self, x):
         """Forward pass for the YOLOv8 mask Proto module."""
         return torch.cat(x, self.d)
+
+
+class Index(nn.Module):
+    """Returns a particular index of the input."""
+
+    def __init__(self, c1, c2, index=0):
+        """Returns a particular index of the input."""
+        super().__init__()
+        self.index = index
+
+    def forward(self, x):
+        """
+        Forward pass.
+
+        Expects a list of tensors as input.
+        """
+        return x[self.index]

ultralytics/nn/tasks.py

@@ -50,6 +50,7 @@ from ultralytics.nn.modules import (
     HGBlock,
     HGStem,
     ImagePoolingAttn,
+    Index,
     Pose,
     RepC3,
     RepConv,
@@ -59,6 +60,7 @@ from ultralytics.nn.modules import (
     RTDETRDecoder,
     SCDown,
     Segment,
+    TorchVision,
     WorldDetect,
     v10Detect,
 )
@@ -1052,7 +1054,7 @@ def parse_model(d, ch, verbose=True):  # model_dict, input_channels(3)
                 m.legacy = legacy
         elif m is RTDETRDecoder:  # special case, channels arg must be passed in index 1
             args.insert(1, [ch[x] for x in f])
-        elif m is CBLinear:
+        elif m in {CBLinear, TorchVision, Index}:
             c2 = args[0]
             c1 = ch[f]
             args = [c1, c2, *args[1:]]