mask_rcnn.py 26.1 KB
Newer Older
1
from collections import OrderedDict
2
from typing import Any, Callable, Optional
3
4
5
6

from torch import nn
from torchvision.ops import MultiScaleRoIAlign

7
from ...ops import misc as misc_nn_ops
8
from ...transforms._presets import ObjectDetection
9
from .._api import register_model, Weights, WeightsEnum
10
from .._meta import _COCO_CATEGORIES
11
12
from .._utils import _ovewrite_value_param, handle_legacy_interface
from ..resnet import resnet50, ResNet50_Weights
13
from ._utils import overwrite_eps
14
from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers
15
from .faster_rcnn import _default_anchorgen, FasterRCNN, FastRCNNConvFCHead, RPNHead
16

17

18
__all__ = [
19
    "MaskRCNN",
20
    "MaskRCNN_ResNet50_FPN_Weights",
21
    "MaskRCNN_ResNet50_FPN_V2_Weights",
22
    "maskrcnn_resnet50_fpn",
23
    "maskrcnn_resnet50_fpn_v2",
24
25
26
27
]


class MaskRCNN(FasterRCNN):
28
29
30
31
32
33
34
35
    """
    Implements Mask R-CNN.

    The input to the model is expected to be a list of tensors, each of shape [C, H, W], one for each
    image, and should be in 0-1 range. Different images can have different sizes.

    The behavior of the model changes depending if it is in training or evaluation mode.

36
    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
37
    containing:
38
39
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
40
        - labels (Int64Tensor[N]): the class label for each ground-truth box
41
        - masks (UInt8Tensor[N, H, W]): the segmentation binary masks for each instance
42

43
44
45
46
47
48
    The model returns a Dict[Tensor] during training, containing the classification and regression
    losses for both the RPN and the R-CNN, and the mask loss.

    During inference, the model requires only the input tensors, and returns the post-processed
    predictions as a List[Dict[Tensor]], one for each input image. The fields of the Dict are as
    follows:
49
50
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
51
        - labels (Int64Tensor[N]): the predicted labels for each image
52
        - scores (Tensor[N]): the scores or each prediction
53
        - masks (UInt8Tensor[N, 1, H, W]): the predicted masks for each instance, in 0-1 range. In order to
54
55
          obtain the final segmentation masks, the soft masks can be thresholded, generally
          with a value of 0.5 (mask >= 0.5)
56

57
    Args:
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
        backbone (nn.Module): the network used to compute the features for the model.
            It should contain a out_channels attribute, which indicates the number of output
            channels that each feature map has (and it should be the same for all feature maps).
            The backbone should return a single Tensor or and OrderedDict[Tensor].
        num_classes (int): number of output classes of the model (including the background).
            If box_predictor is specified, num_classes should be None.
        min_size (int): minimum size of the image to be rescaled before feeding it to the backbone
        max_size (int): maximum size of the image to be rescaled before feeding it to the backbone
        image_mean (Tuple[float, float, float]): mean values used for input normalization.
            They are generally the mean values of the dataset on which the backbone has been trained
            on
        image_std (Tuple[float, float, float]): std values used for input normalization.
            They are generally the std values of the dataset on which the backbone has been trained on
        rpn_anchor_generator (AnchorGenerator): module that generates the anchors for a set of feature
            maps.
        rpn_head (nn.Module): module that computes the objectness and regression deltas from the RPN
        rpn_pre_nms_top_n_train (int): number of proposals to keep before applying NMS during training
        rpn_pre_nms_top_n_test (int): number of proposals to keep before applying NMS during testing
        rpn_post_nms_top_n_train (int): number of proposals to keep after applying NMS during training
        rpn_post_nms_top_n_test (int): number of proposals to keep after applying NMS during testing
        rpn_nms_thresh (float): NMS threshold used for postprocessing the RPN proposals
        rpn_fg_iou_thresh (float): minimum IoU between the anchor and the GT box so that they can be
            considered as positive during training of the RPN.
        rpn_bg_iou_thresh (float): maximum IoU between the anchor and the GT box so that they can be
            considered as negative during training of the RPN.
        rpn_batch_size_per_image (int): number of anchors that are sampled during training of the RPN
            for computing the loss
        rpn_positive_fraction (float): proportion of positive anchors in a mini-batch during training
            of the RPN
87
88
        rpn_score_thresh (float): during inference, only return proposals with a classification score
            greater than rpn_score_thresh
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
        box_roi_pool (MultiScaleRoIAlign): the module which crops and resizes the feature maps in
            the locations indicated by the bounding boxes
        box_head (nn.Module): module that takes the cropped feature maps as input
        box_predictor (nn.Module): module that takes the output of box_head and returns the
            classification logits and box regression deltas.
        box_score_thresh (float): during inference, only return proposals with a classification score
            greater than box_score_thresh
        box_nms_thresh (float): NMS threshold for the prediction head. Used during inference
        box_detections_per_img (int): maximum number of detections per image, for all classes.
        box_fg_iou_thresh (float): minimum IoU between the proposals and the GT box so that they can be
            considered as positive during training of the classification head
        box_bg_iou_thresh (float): maximum IoU between the proposals and the GT box so that they can be
            considered as negative during training of the classification head
        box_batch_size_per_image (int): number of proposals that are sampled during training of the
            classification head
        box_positive_fraction (float): proportion of positive proposals in a mini-batch during training
            of the classification head
        bbox_reg_weights (Tuple[float, float, float, float]): weights for the encoding/decoding of the
            bounding boxes
        mask_roi_pool (MultiScaleRoIAlign): the module which crops and resizes the feature maps in
             the locations indicated by the bounding boxes, which will be used for the mask head.
        mask_head (nn.Module): module that takes the cropped feature maps as input
        mask_predictor (nn.Module): module that takes the output of the mask_head and returns the
            segmentation mask logits

    Example::

116
        >>> import torch
117
118
        >>> import torchvision
        >>> from torchvision.models.detection import MaskRCNN
119
        >>> from torchvision.models.detection.anchor_utils import AnchorGenerator
120
121
122
        >>>
        >>> # load a pre-trained model for classification and return
        >>> # only the features
123
        >>> backbone = torchvision.models.mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT).features
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
        >>> # MaskRCNN needs to know the number of
        >>> # output channels in a backbone. For mobilenet_v2, it's 1280
        >>> # so we need to add it here
        >>> backbone.out_channels = 1280
        >>>
        >>> # let's make the RPN generate 5 x 3 anchors per spatial
        >>> # location, with 5 different sizes and 3 different aspect
        >>> # ratios. We have a Tuple[Tuple[int]] because each feature
        >>> # map could potentially have different sizes and
        >>> # aspect ratios
        >>> anchor_generator = AnchorGenerator(sizes=((32, 64, 128, 256, 512),),
        >>>                                    aspect_ratios=((0.5, 1.0, 2.0),))
        >>>
        >>> # let's define what are the feature maps that we will
        >>> # use to perform the region of interest cropping, as well as
        >>> # the size of the crop after rescaling.
        >>> # if your backbone returns a Tensor, featmap_names is expected to
141
        >>> # be ['0']. More generally, the backbone should return an
142
143
        >>> # OrderedDict[Tensor], and in featmap_names you can choose which
        >>> # feature maps to use.
144
        >>> roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
145
146
147
        >>>                                                 output_size=7,
        >>>                                                 sampling_ratio=2)
        >>>
148
        >>> mask_roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
149
150
        >>>                                                      output_size=14,
        >>>                                                      sampling_ratio=2)
151
        >>> # put the pieces together inside a MaskRCNN model
152
153
154
155
156
        >>> model = MaskRCNN(backbone,
        >>>                  num_classes=2,
        >>>                  rpn_anchor_generator=anchor_generator,
        >>>                  box_roi_pool=roi_pooler,
        >>>                  mask_roi_pool=mask_roi_pooler)
157
158
159
160
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
    """
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199

    def __init__(
        self,
        backbone,
        num_classes=None,
        # transform parameters
        min_size=800,
        max_size=1333,
        image_mean=None,
        image_std=None,
        # RPN parameters
        rpn_anchor_generator=None,
        rpn_head=None,
        rpn_pre_nms_top_n_train=2000,
        rpn_pre_nms_top_n_test=1000,
        rpn_post_nms_top_n_train=2000,
        rpn_post_nms_top_n_test=1000,
        rpn_nms_thresh=0.7,
        rpn_fg_iou_thresh=0.7,
        rpn_bg_iou_thresh=0.3,
        rpn_batch_size_per_image=256,
        rpn_positive_fraction=0.5,
        rpn_score_thresh=0.0,
        # Box parameters
        box_roi_pool=None,
        box_head=None,
        box_predictor=None,
        box_score_thresh=0.05,
        box_nms_thresh=0.5,
        box_detections_per_img=100,
        box_fg_iou_thresh=0.5,
        box_bg_iou_thresh=0.5,
        box_batch_size_per_image=512,
        box_positive_fraction=0.25,
        bbox_reg_weights=None,
        # Mask parameters
        mask_roi_pool=None,
        mask_head=None,
        mask_predictor=None,
200
        **kwargs,
201
    ):
202

203
204
205
206
        if not isinstance(mask_roi_pool, (MultiScaleRoIAlign, type(None))):
            raise TypeError(
                f"mask_roi_pool should be of type MultiScaleRoIAlign or None instead of {type(mask_roi_pool)}"
            )
207
208
209
210
211
212
213
214

        if num_classes is not None:
            if mask_predictor is not None:
                raise ValueError("num_classes should be None when mask_predictor is specified")

        out_channels = backbone.out_channels

        if mask_roi_pool is None:
215
            mask_roi_pool = MultiScaleRoIAlign(featmap_names=["0", "1", "2", "3"], output_size=14, sampling_ratio=2)
216
217
218
219
220
221
222

        if mask_head is None:
            mask_layers = (256, 256, 256, 256)
            mask_dilation = 1
            mask_head = MaskRCNNHeads(out_channels, mask_layers, mask_dilation)

        if mask_predictor is None:
223
224
            mask_predictor_in_channels = 256  # == mask_layers[-1]
            mask_dim_reduced = 256
225
            mask_predictor = MaskRCNNPredictor(mask_predictor_in_channels, mask_dim_reduced, num_classes)
226

227
        super().__init__(
228
229
            backbone,
            num_classes,
230
            # transform parameters
231
232
233
234
            min_size,
            max_size,
            image_mean,
            image_std,
235
            # RPN-specific parameters
236
237
238
239
240
241
            rpn_anchor_generator,
            rpn_head,
            rpn_pre_nms_top_n_train,
            rpn_pre_nms_top_n_test,
            rpn_post_nms_top_n_train,
            rpn_post_nms_top_n_test,
242
            rpn_nms_thresh,
243
244
245
246
            rpn_fg_iou_thresh,
            rpn_bg_iou_thresh,
            rpn_batch_size_per_image,
            rpn_positive_fraction,
247
            rpn_score_thresh,
248
            # Box parameters
249
250
251
252
253
254
255
256
257
258
259
            box_roi_pool,
            box_head,
            box_predictor,
            box_score_thresh,
            box_nms_thresh,
            box_detections_per_img,
            box_fg_iou_thresh,
            box_bg_iou_thresh,
            box_batch_size_per_image,
            box_positive_fraction,
            bbox_reg_weights,
260
            **kwargs,
261
        )
262
263
264
265
266
267
268

        self.roi_heads.mask_roi_pool = mask_roi_pool
        self.roi_heads.mask_head = mask_head
        self.roi_heads.mask_predictor = mask_predictor


class MaskRCNNHeads(nn.Sequential):
269
270
271
    _version = 2

    def __init__(self, in_channels, layers, dilation, norm_layer: Optional[Callable[..., nn.Module]] = None):
272
        """
273
        Args:
274
275
276
            in_channels (int): number of input channels
            layers (list): feature dimensions of each FCN layer
            dilation (int): dilation rate of kernel
277
            norm_layer (callable, optional): Module specifying the normalization layer to use. Default: None
278
        """
279
        blocks = []
280
        next_feature = in_channels
281
282
283
284
285
286
287
288
289
290
291
        for layer_features in layers:
            blocks.append(
                misc_nn_ops.Conv2dNormActivation(
                    next_feature,
                    layer_features,
                    kernel_size=3,
                    stride=1,
                    padding=dilation,
                    dilation=dilation,
                    norm_layer=norm_layer,
                )
292
            )
293
294
            next_feature = layer_features

295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
        super().__init__(*blocks)
        for layer in self.modules():
            if isinstance(layer, nn.Conv2d):
                nn.init.kaiming_normal_(layer.weight, mode="fan_out", nonlinearity="relu")
                if layer.bias is not None:
                    nn.init.zeros_(layer.bias)

    def _load_from_state_dict(
        self,
        state_dict,
        prefix,
        local_metadata,
        strict,
        missing_keys,
        unexpected_keys,
        error_msgs,
    ):
        version = local_metadata.get("version", None)

        if version is None or version < 2:
            num_blocks = len(self)
            for i in range(num_blocks):
                for type in ["weight", "bias"]:
                    old_key = f"{prefix}mask_fcn{i+1}.{type}"
                    new_key = f"{prefix}{i}.0.{type}"
320
321
                    if old_key in state_dict:
                        state_dict[new_key] = state_dict.pop(old_key)
322
323
324
325
326
327
328
329
330
331

        super()._load_from_state_dict(
            state_dict,
            prefix,
            local_metadata,
            strict,
            missing_keys,
            unexpected_keys,
            error_msgs,
        )
332
333


334
class MaskRCNNPredictor(nn.Sequential):
335
    def __init__(self, in_channels, dim_reduced, num_classes):
336
        super().__init__(
337
338
339
340
341
342
343
344
            OrderedDict(
                [
                    ("conv5_mask", nn.ConvTranspose2d(in_channels, dim_reduced, 2, 2, 0)),
                    ("relu", nn.ReLU(inplace=True)),
                    ("mask_fcn_logits", nn.Conv2d(dim_reduced, num_classes, 1, 1, 0)),
                ]
            )
        )
345
346
347
348
349
350
351
352

        for name, param in self.named_parameters():
            if "weight" in name:
                nn.init.kaiming_normal_(param, mode="fan_out", nonlinearity="relu")
            # elif "bias" in name:
            #     nn.init.constant_(param, 0)


353
354
_COMMON_META = {
    "categories": _COCO_CATEGORIES,
355
    "min_size": (1, 1),
356
357
358
}


359
360
361
362
363
class MaskRCNN_ResNet50_FPN_Weights(WeightsEnum):
    COCO_V1 = Weights(
        url="https://download.pytorch.org/models/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth",
        transforms=ObjectDetection,
        meta={
364
            **_COMMON_META,
365
366
            "num_params": 44401393,
            "recipe": "https://github.com/pytorch/vision/tree/main/references/detection#mask-r-cnn",
367
368
369
370
371
            "_metrics": {
                "COCO-val2017": {
                    "box_map": 37.9,
                    "mask_map": 34.6,
                }
372
            },
373
374
            "_ops": 134.38,
            "_weight_size": 169.84,
375
            "_docs": """These weights were produced by following a similar training recipe as on the paper.""",
376
377
378
379
380
        },
    )
    DEFAULT = COCO_V1


381
class MaskRCNN_ResNet50_FPN_V2_Weights(WeightsEnum):
382
383
384
385
386
387
388
    COCO_V1 = Weights(
        url="https://download.pytorch.org/models/maskrcnn_resnet50_fpn_v2_coco-73cbd019.pth",
        transforms=ObjectDetection,
        meta={
            **_COMMON_META,
            "num_params": 46359409,
            "recipe": "https://github.com/pytorch/vision/pull/5773",
389
390
391
392
393
            "_metrics": {
                "COCO-val2017": {
                    "box_map": 47.4,
                    "mask_map": 41.8,
                }
394
            },
395
396
            "_ops": 333.577,
            "_weight_size": 177.219,
397
            "_docs": """These weights were produced using an enhanced training recipe to boost the model accuracy.""",
398
399
400
        },
    )
    DEFAULT = COCO_V1
401
402


403
@register_model()
404
405
406
407
@handle_legacy_interface(
    weights=("pretrained", MaskRCNN_ResNet50_FPN_Weights.COCO_V1),
    weights_backbone=("pretrained_backbone", ResNet50_Weights.IMAGENET1K_V1),
)
408
def maskrcnn_resnet50_fpn(
409
410
411
412
413
414
415
416
    *,
    weights: Optional[MaskRCNN_ResNet50_FPN_Weights] = None,
    progress: bool = True,
    num_classes: Optional[int] = None,
    weights_backbone: Optional[ResNet50_Weights] = ResNet50_Weights.IMAGENET1K_V1,
    trainable_backbone_layers: Optional[int] = None,
    **kwargs: Any,
) -> MaskRCNN:
417
418
    """Mask R-CNN model with a ResNet-50-FPN backbone from the `Mask R-CNN
    <https://arxiv.org/abs/1703.06870>`_ paper.
419

420
421
    .. betastatus:: detection module

422
423
424
425
426
    The input to the model is expected to be a list of tensors, each of shape ``[C, H, W]``, one for each
    image, and should be in ``0-1`` range. Different images can have different sizes.

    The behavior of the model changes depending if it is in training or evaluation mode.

427
    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
428
    containing:
429

430
431
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
432
        - labels (``Int64Tensor[N]``): the class label for each ground-truth box
433
        - masks (``UInt8Tensor[N, H, W]``): the segmentation binary masks for each instance
434
435
436
437
438
439

    The model returns a ``Dict[Tensor]`` during training, containing the classification and regression
    losses for both the RPN and the R-CNN, and the mask loss.

    During inference, the model requires only the input tensors, and returns the post-processed
    predictions as a ``List[Dict[Tensor]]``, one for each input image. The fields of the ``Dict`` are as
440
    follows, where ``N`` is the number of detected instances:
441

442
443
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
444
445
        - labels (``Int64Tensor[N]``): the predicted labels for each instance
        - scores (``Tensor[N]``): the scores or each instance
446
        - masks (``UInt8Tensor[N, 1, H, W]``): the predicted masks for each instance, in ``0-1`` range. In order to
447
448
449
          obtain the final segmentation masks, the soft masks can be thresholded, generally
          with a value of 0.5 (``mask >= 0.5``)

450
451
    For more details on the output and on how to plot the masks, you may refer to :ref:`instance_seg_output`.

452
453
    Mask R-CNN is exportable to ONNX for a fixed batch size with inputs images of fixed size.

454
455
    Example::

456
        >>> model = torchvision.models.detection.maskrcnn_resnet50_fpn(weights=MaskRCNN_ResNet50_FPN_Weights.DEFAULT)
457
458
459
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
460
461
462
        >>>
        >>> # optionally, if you want to export the model to ONNX:
        >>> torch.onnx.export(model, x, "mask_rcnn.onnx", opset_version = 11)
463

464
    Args:
465
466
467
468
469
470
471
        weights (:class:`~torchvision.models.detection.MaskRCNN_ResNet50_FPN_Weights`, optional): The
            pretrained weights to use. See
            :class:`~torchvision.models.detection.MaskRCNN_ResNet50_FPN_Weights` below for
            more details, and possible values. By default, no pre-trained
            weights are used.
        progress (bool, optional): If True, displays a progress bar of the
            download to stderr. Default is True.
472
        num_classes (int, optional): number of output classes of the model (including the background)
473
474
475
476
477
478
479
480
481
482
483
484
        weights_backbone (:class:`~torchvision.models.ResNet50_Weights`, optional): The
            pretrained weights for the backbone.
        trainable_backbone_layers (int, optional): number of trainable (not frozen) layers starting from
            final block. Valid values are between 0 and 5, with 5 meaning all backbone layers are
            trainable. If ``None`` is passed (the default) this value is set to 3.
        **kwargs: parameters passed to the ``torchvision.models.detection.mask_rcnn.MaskRCNN``
            base class. Please refer to the `source code
            <https://github.com/pytorch/vision/blob/main/torchvision/models/detection/mask_rcnn.py>`_
            for more details about this class.

    .. autoclass:: torchvision.models.detection.MaskRCNN_ResNet50_FPN_Weights
        :members:
485
    """
486
487
488
489
490
    weights = MaskRCNN_ResNet50_FPN_Weights.verify(weights)
    weights_backbone = ResNet50_Weights.verify(weights_backbone)

    if weights is not None:
        weights_backbone = None
491
        num_classes = _ovewrite_value_param("num_classes", num_classes, len(weights.meta["categories"]))
492
493
494
495
    elif num_classes is None:
        num_classes = 91

    is_trained = weights is not None or weights_backbone is not None
496
497
    trainable_backbone_layers = _validate_trainable_layers(is_trained, trainable_backbone_layers, 5, 3)
    norm_layer = misc_nn_ops.FrozenBatchNorm2d if is_trained else nn.BatchNorm2d
498

499
    backbone = resnet50(weights=weights_backbone, progress=progress, norm_layer=norm_layer)
500
    backbone = _resnet_fpn_extractor(backbone, trainable_backbone_layers)
501
502
503
504
505
506
507
    model = MaskRCNN(backbone, num_classes=num_classes, **kwargs)

    if weights is not None:
        model.load_state_dict(weights.get_state_dict(progress=progress))
        if weights == MaskRCNN_ResNet50_FPN_Weights.COCO_V1:
            overwrite_eps(model, 0.0)

508
    return model
509
510


511
@register_model()
512
513
514
515
@handle_legacy_interface(
    weights=("pretrained", MaskRCNN_ResNet50_FPN_V2_Weights.COCO_V1),
    weights_backbone=("pretrained_backbone", ResNet50_Weights.IMAGENET1K_V1),
)
516
517
518
519
520
521
522
523
524
def maskrcnn_resnet50_fpn_v2(
    *,
    weights: Optional[MaskRCNN_ResNet50_FPN_V2_Weights] = None,
    progress: bool = True,
    num_classes: Optional[int] = None,
    weights_backbone: Optional[ResNet50_Weights] = None,
    trainable_backbone_layers: Optional[int] = None,
    **kwargs: Any,
) -> MaskRCNN:
525
526
    """Improved Mask R-CNN model with a ResNet-50-FPN backbone from the `Benchmarking Detection Transfer
    Learning with Vision Transformers <https://arxiv.org/abs/2111.11429>`_ paper.
527

528
529
    .. betastatus:: detection module

530
531
532
    :func:`~torchvision.models.detection.maskrcnn_resnet50_fpn` for more details.

    Args:
533
534
535
536
537
538
539
        weights (:class:`~torchvision.models.detection.MaskRCNN_ResNet50_FPN_V2_Weights`, optional): The
            pretrained weights to use. See
            :class:`~torchvision.models.detection.MaskRCNN_ResNet50_FPN_V2_Weights` below for
            more details, and possible values. By default, no pre-trained
            weights are used.
        progress (bool, optional): If True, displays a progress bar of the
            download to stderr. Default is True.
540
        num_classes (int, optional): number of output classes of the model (including the background)
541
542
543
544
545
546
547
548
549
550
551
552
        weights_backbone (:class:`~torchvision.models.ResNet50_Weights`, optional): The
            pretrained weights for the backbone.
        trainable_backbone_layers (int, optional): number of trainable (not frozen) layers starting from
            final block. Valid values are between 0 and 5, with 5 meaning all backbone layers are
            trainable. If ``None`` is passed (the default) this value is set to 3.
        **kwargs: parameters passed to the ``torchvision.models.detection.mask_rcnn.MaskRCNN``
            base class. Please refer to the `source code
            <https://github.com/pytorch/vision/blob/main/torchvision/models/detection/mask_rcnn.py>`_
            for more details about this class.

    .. autoclass:: torchvision.models.detection.MaskRCNN_ResNet50_FPN_V2_Weights
        :members:
553
554
555
556
557
558
    """
    weights = MaskRCNN_ResNet50_FPN_V2_Weights.verify(weights)
    weights_backbone = ResNet50_Weights.verify(weights_backbone)

    if weights is not None:
        weights_backbone = None
559
        num_classes = _ovewrite_value_param("num_classes", num_classes, len(weights.meta["categories"]))
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
    elif num_classes is None:
        num_classes = 91

    is_trained = weights is not None or weights_backbone is not None
    trainable_backbone_layers = _validate_trainable_layers(is_trained, trainable_backbone_layers, 5, 3)

    backbone = resnet50(weights=weights_backbone, progress=progress)
    backbone = _resnet_fpn_extractor(backbone, trainable_backbone_layers, norm_layer=nn.BatchNorm2d)
    rpn_anchor_generator = _default_anchorgen()
    rpn_head = RPNHead(backbone.out_channels, rpn_anchor_generator.num_anchors_per_location()[0], conv_depth=2)
    box_head = FastRCNNConvFCHead(
        (backbone.out_channels, 7, 7), [256, 256, 256, 256], [1024], norm_layer=nn.BatchNorm2d
    )
    mask_head = MaskRCNNHeads(backbone.out_channels, [256, 256, 256, 256], 1, norm_layer=nn.BatchNorm2d)
    model = MaskRCNN(
        backbone,
        num_classes=num_classes,
        rpn_anchor_generator=rpn_anchor_generator,
        rpn_head=rpn_head,
        box_head=box_head,
        mask_head=mask_head,
        **kwargs,
    )

    if weights is not None:
        model.load_state_dict(weights.get_state_dict(progress=progress))

    return model
588
589
590
591
592
593
594
595
596
597
598


# The dictionary below is internal implementation detail and will be removed in v0.15
from .._utils import _ModelURLs


model_urls = _ModelURLs(
    {
        "maskrcnn_resnet50_fpn_coco": MaskRCNN_ResNet50_FPN_Weights.COCO_V1.url,
    }
)