Refactor the backbone builders of detection (#4656)

* Refactoring resnet_fpn backbone building.

* Passing the change to *_rcnn and retinanet.

* Applying for faster_rcnn + mobilenetv3

* Applying for ssdlite + mobilenetv3

* Applying for ssd + vgg16

* Update the expected file of retinanet_resnet50_fpn to fix order of initialization.

* Adding full model weights for the VGG16 features.

Author: datumbox

test/expect/ModelTester.test_retinanet_resnet50_fpn_expect.pkl

@@ -1,5 +1,5 @@
 import warnings
-from typing import Callable, Dict, Optional, List
+from typing import Callable, Dict, Optional, List, Union
 
 from torch import nn, Tensor
 from torchvision.ops import misc as misc_nn_ops
@@ -100,14 +100,14 @@ def resnet_fpn_backbone(
             default a ``LastLevelMaxPool`` is used.
     """
     backbone = resnet.__dict__[backbone_name](pretrained=pretrained, norm_layer=norm_layer)
-    return _resnet_backbone_config(backbone, trainable_layers, returned_layers, extra_blocks)
+    return _resnet_fpn_extractor(backbone, trainable_layers, returned_layers, extra_blocks)
 
 
-def _resnet_backbone_config(
+def _resnet_fpn_extractor(
     backbone: resnet.ResNet,
     trainable_layers: int,
-    returned_layers: Optional[List[int]],
-    extra_blocks: Optional[ExtraFPNBlock],
+    returned_layers: Optional[List[int]] = None,
+    extra_blocks: Optional[ExtraFPNBlock] = None,
 ) -> BackboneWithFPN:
 
     # select layers that wont be frozen
@@ -165,9 +165,18 @@ def mobilenet_backbone(
     returned_layers: Optional[List[int]] = None,
     extra_blocks: Optional[ExtraFPNBlock] = None,
 ) -> nn.Module:
+    backbone = mobilenet.__dict__[backbone_name](pretrained=pretrained, norm_layer=norm_layer)
+    return _mobilenet_extractor(backbone, fpn, trainable_layers, returned_layers, extra_blocks)
 
-    backbone = mobilenet.__dict__[backbone_name](pretrained=pretrained, norm_layer=norm_layer).features
 
+def _mobilenet_extractor(
+    backbone: Union[mobilenet.MobileNetV2, mobilenet.MobileNetV3],
+    fpn: bool,
+    trainable_layers,
+    returned_layers: Optional[List[int]] = None,
+    extra_blocks: Optional[ExtraFPNBlock] = None,
+) -> nn.Module:
+    backbone = backbone.features
     # Gather the indices of blocks which are strided. These are the locations of C1, ..., Cn-1 blocks.
     # The first and last blocks are always included because they are the C0 (conv1) and Cn.
     stage_indices = [0] + [i for i, b in enumerate(backbone) if getattr(b, "_is_cn", False)] + [len(backbone) - 1]

torchvision/models/detection/backbone_utils.py

@@ -3,9 +3,12 @@
 from torchvision.ops import MultiScaleRoIAlign
 
 from ..._internally_replaced_utils import load_state_dict_from_url
+from ...ops import misc as misc_nn_ops
+from ..mobilenetv3 import mobilenet_v3_large
+from ..resnet import resnet50
 from ._utils import overwrite_eps
 from .anchor_utils import AnchorGenerator
-from .backbone_utils import resnet_fpn_backbone, _validate_trainable_layers, mobilenet_backbone
+from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers, _mobilenet_extractor
 from .generalized_rcnn import GeneralizedRCNN
 from .roi_heads import RoIHeads
 from .rpn import RPNHead, RegionProposalNetwork
@@ -385,7 +388,9 @@ def fasterrcnn_resnet50_fpn(
     if pretrained:
         # no need to download the backbone if pretrained is set
         pretrained_backbone = False
-    backbone = resnet_fpn_backbone("resnet50", pretrained_backbone, trainable_layers=trainable_backbone_layers)
+
+    backbone = resnet50(pretrained=pretrained_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d)
+    backbone = _resnet_fpn_extractor(backbone, trainable_backbone_layers)
     model = FasterRCNN(backbone, num_classes, **kwargs)
     if pretrained:
         state_dict = load_state_dict_from_url(model_urls["fasterrcnn_resnet50_fpn_coco"], progress=progress)
@@ -409,9 +414,11 @@ def _fasterrcnn_mobilenet_v3_large_fpn(
 
     if pretrained:
         pretrained_backbone = False
-    backbone = mobilenet_backbone(
-        "mobilenet_v3_large", pretrained_backbone, True, trainable_layers=trainable_backbone_layers
+
+    backbone = mobilenet_v3_large(
+        pretrained=pretrained_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d
     )
+    backbone = _mobilenet_extractor(backbone, True, trainable_backbone_layers)
 
     anchor_sizes = (
         (

torchvision/models/detection/faster_rcnn.py

@@ -3,8 +3,10 @@
 from torchvision.ops import MultiScaleRoIAlign
 
 from ..._internally_replaced_utils import load_state_dict_from_url
+from ...ops import misc as misc_nn_ops
+from ..resnet import resnet50
 from ._utils import overwrite_eps
-from .backbone_utils import resnet_fpn_backbone, _validate_trainable_layers
+from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers
 from .faster_rcnn import FasterRCNN
 
 
@@ -367,7 +369,9 @@ def keypointrcnn_resnet50_fpn(
     if pretrained:
         # no need to download the backbone if pretrained is set
         pretrained_backbone = False
-    backbone = resnet_fpn_backbone("resnet50", pretrained_backbone, trainable_layers=trainable_backbone_layers)
+
+    backbone = resnet50(pretrained=pretrained_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d)
+    backbone = _resnet_fpn_extractor(backbone, trainable_backbone_layers)
     model = KeypointRCNN(backbone, num_classes, num_keypoints=num_keypoints, **kwargs)
     if pretrained:
         key = "keypointrcnn_resnet50_fpn_coco"

torchvision/models/detection/keypoint_rcnn.py

@@ -4,8 +4,10 @@
 from torchvision.ops import MultiScaleRoIAlign
 
 from ..._internally_replaced_utils import load_state_dict_from_url
+from ...ops import misc as misc_nn_ops
+from ..resnet import resnet50
 from ._utils import overwrite_eps
-from .backbone_utils import resnet_fpn_backbone, _validate_trainable_layers
+from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers
 from .faster_rcnn import FasterRCNN
 
 __all__ = [
@@ -364,7 +366,9 @@ def maskrcnn_resnet50_fpn(
     if pretrained:
         # no need to download the backbone if pretrained is set
         pretrained_backbone = False
-    backbone = resnet_fpn_backbone("resnet50", pretrained_backbone, trainable_layers=trainable_backbone_layers)
+
+    backbone = resnet50(pretrained=pretrained_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d)
+    backbone = _resnet_fpn_extractor(backbone, trainable_backbone_layers)
     model = MaskRCNN(backbone, num_classes, **kwargs)
     if pretrained:
         state_dict = load_state_dict_from_url(model_urls["maskrcnn_resnet50_fpn_coco"], progress=progress)

torchvision/models/detection/mask_rcnn.py

@@ -9,11 +9,13 @@
 from ..._internally_replaced_utils import load_state_dict_from_url
 from ...ops import sigmoid_focal_loss
 from ...ops import boxes as box_ops
+from ...ops import misc as misc_nn_ops
 from ...ops.feature_pyramid_network import LastLevelP6P7
+from ..resnet import resnet50
 from . import _utils as det_utils
 from ._utils import overwrite_eps
 from .anchor_utils import AnchorGenerator
-from .backbone_utils import resnet_fpn_backbone, _validate_trainable_layers
+from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers
 from .transform import GeneralizedRCNNTransform
 
 
@@ -630,13 +632,11 @@ def retinanet_resnet50_fpn(
     if pretrained:
         # no need to download the backbone if pretrained is set
         pretrained_backbone = False
+
+    backbone = resnet50(pretrained=pretrained_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d)
     # skip P2 because it generates too many anchors (according to their paper)
-    backbone = resnet_fpn_backbone(
-        "resnet50",
-        pretrained_backbone,
-        returned_layers=[2, 3, 4],
-        extra_blocks=LastLevelP6P7(256, 256),
-        trainable_layers=trainable_backbone_layers,
+    backbone = _resnet_fpn_extractor(
+        backbone, trainable_backbone_layers, returned_layers=[2, 3, 4], extra_blocks=LastLevelP6P7(256, 256)
     )
     model = RetinaNet(backbone, num_classes, **kwargs)
     if pretrained:

torchvision/models/detection/retinanet.py

@@ -23,7 +23,8 @@
 backbone_urls = {
     # We port the features of a VGG16 backbone trained by amdegroot because unlike the one on TorchVision, it uses the
     # same input standardization method as the paper. Ref: https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
-    "vgg16_features": "https://download.pytorch.org/models/vgg16_features-amdegroot.pth"
+    # Only the `features` weights have proper values, those on the `classifier` module are filled with nans.
+    "vgg16_features": "https://download.pytorch.org/models/vgg16_features-amdegroot-88682ab5.pth"
 }
 
 
@@ -519,18 +520,8 @@ def forward(self, x: Tensor) -> Dict[str, Tensor]:
         return OrderedDict([(str(i), v) for i, v in enumerate(output)])
 
 
-def _vgg_extractor(backbone_name: str, highres: bool, progress: bool, pretrained: bool, trainable_layers: int):
-    if backbone_name in backbone_urls:
-        # Use custom backbones more appropriate for SSD
-        arch = backbone_name.split("_")[0]
-        backbone = vgg.__dict__[arch](pretrained=False, progress=progress).features
-        if pretrained:
-            state_dict = load_state_dict_from_url(backbone_urls[backbone_name], progress=progress)
-            backbone.load_state_dict(state_dict)
-    else:
-        # Use standard backbones from TorchVision
-        backbone = vgg.__dict__[backbone_name](pretrained=pretrained, progress=progress).features
-
+def _vgg_extractor(backbone: vgg.VGG, highres: bool, trainable_layers: int):
+    backbone = backbone.features
     # Gather the indices of maxpools. These are the locations of output blocks.
     stage_indices = [0] + [i for i, b in enumerate(backbone) if isinstance(b, nn.MaxPool2d)][:-1]
     num_stages = len(stage_indices)
@@ -609,7 +600,13 @@ def ssd300_vgg16(
         # no need to download the backbone if pretrained is set
         pretrained_backbone = False
 
-    backbone = _vgg_extractor("vgg16_features", False, progress, pretrained_backbone, trainable_backbone_layers)
+    # Use custom backbones more appropriate for SSD
+    backbone = vgg.vgg16(pretrained=False, progress=progress)
+    if pretrained_backbone:
+        state_dict = load_state_dict_from_url(backbone_urls["vgg16_features"], progress=progress)
+        backbone.load_state_dict(state_dict)
+
+    backbone = _vgg_extractor(backbone, False, trainable_backbone_layers)
     anchor_generator = DefaultBoxGenerator(
         [[2], [2, 3], [2, 3], [2, 3], [2], [2]],
         scales=[0.07, 0.15, 0.33, 0.51, 0.69, 0.87, 1.05],

torchvision/models/detection/ssd.py

@@ -1,7 +1,7 @@
 import warnings
 from collections import OrderedDict
 from functools import partial
-from typing import Any, Callable, Dict, List, Optional
+from typing import Any, Callable, Dict, List, Optional, Union
 
 import torch
 from torch import nn, Tensor
@@ -117,7 +117,6 @@ def __init__(
         norm_layer: Callable[..., nn.Module],
         width_mult: float = 1.0,
         min_depth: int = 16,
-        **kwargs: Any,
     ):
         super().__init__()
 
@@ -156,20 +155,11 @@ def forward(self, x: Tensor) -> Dict[str, Tensor]:
 
 
 def _mobilenet_extractor(
-    backbone_name: str,
-    progress: bool,
-    pretrained: bool,
+    backbone: Union[mobilenet.MobileNetV2, mobilenet.MobileNetV3],
     trainable_layers: int,
     norm_layer: Callable[..., nn.Module],
-    **kwargs: Any,
 ):
-    backbone = mobilenet.__dict__[backbone_name](
-        pretrained=pretrained, progress=progress, norm_layer=norm_layer, **kwargs
-    ).features
-    if not pretrained:
-        # Change the default initialization scheme if not pretrained
-        _normal_init(backbone)
-
+    backbone = backbone.features
     # Gather the indices of blocks which are strided. These are the locations of C1, ..., Cn-1 blocks.
     # The first and last blocks are always included because they are the C0 (conv1) and Cn.
     stage_indices = [0] + [i for i, b in enumerate(backbone) if getattr(b, "_is_cn", False)] + [len(backbone) - 1]
@@ -183,7 +173,7 @@ def _mobilenet_extractor(
         for parameter in b.parameters():
             parameter.requires_grad_(False)
 
-    return SSDLiteFeatureExtractorMobileNet(backbone, stage_indices[-2], norm_layer, **kwargs)
+    return SSDLiteFeatureExtractorMobileNet(backbone, stage_indices[-2], norm_layer)
 
 
 def ssdlite320_mobilenet_v3_large(
@@ -235,14 +225,16 @@ def ssdlite320_mobilenet_v3_large(
     if norm_layer is None:
         norm_layer = partial(nn.BatchNorm2d, eps=0.001, momentum=0.03)
 
+    backbone = mobilenet.mobilenet_v3_large(
+        pretrained=pretrained_backbone, progress=progress, norm_layer=norm_layer, reduced_tail=reduce_tail, **kwargs
+    )
+    if not pretrained_backbone:
+        # Change the default initialization scheme if not pretrained
+        _normal_init(backbone)
     backbone = _mobilenet_extractor(
-        "mobilenet_v3_large",
-        progress,
-        pretrained_backbone,
+        backbone,
         trainable_backbone_layers,
         norm_layer,
-        reduced_tail=reduce_tail,
-        **kwargs,
     )
 
     size = (320, 320)

torchvision/models/detection/ssdlite.py

@@ -1,12 +1,17 @@
 import warnings
 from typing import Any, Optional
 
-from ....models.detection.faster_rcnn import FasterRCNN, overwrite_eps, _validate_trainable_layers
+from ....models.detection.faster_rcnn import (
+    _validate_trainable_layers,
+    _resnet_fpn_extractor,
+    FasterRCNN,
+    misc_nn_ops,
+    overwrite_eps,
+)
 from ...transforms.presets import CocoEval
 from .._api import Weights, WeightEntry
 from .._meta import _COCO_CATEGORIES
-from ..resnet import ResNet50Weights
-from .backbone_utils import resnet_fpn_backbone
+from ..resnet import ResNet50Weights, resnet50
 
 
 __all__ = ["FasterRCNN", "FasterRCNNResNet50FPNWeights", "fasterrcnn_resnet50_fpn"]
@@ -49,7 +54,8 @@ def fasterrcnn_resnet50_fpn(
         weights is not None or weights_backbone is not None, trainable_backbone_layers, 5, 3
     )
 
-    backbone = resnet_fpn_backbone("resnet50", weights_backbone, trainable_layers=trainable_backbone_layers)
+    backbone = resnet50(weights=weights_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d)
+    backbone = _resnet_fpn_extractor(backbone, trainable_backbone_layers)
     model = FasterRCNN(backbone, num_classes=num_classes, **kwargs)
 
     if weights is not None: