retinanet.py 27 KB
Newer Older
1
2
import math
import warnings
3
4
from collections import OrderedDict
from typing import Dict, List, Tuple, Optional
5
6

import torch
7
from torch import nn, Tensor
8

9
from ..._internally_replaced_utils import load_state_dict_from_url
10
11
from ...ops import sigmoid_focal_loss
from ...ops import boxes as box_ops
12
from ...ops import misc as misc_nn_ops
13
from ...ops.feature_pyramid_network import LastLevelP6P7
14
from ...utils import _log_api_usage_once
15
from ..resnet import resnet50
16
from . import _utils as det_utils
17
from ._utils import overwrite_eps
18
from .anchor_utils import AnchorGenerator
19
from .backbone_utils import _resnet_fpn_extractor, _validate_trainable_layers
20
from .transform import GeneralizedRCNNTransform
21
22


23
__all__ = ["RetinaNet", "retinanet_resnet50_fpn"]
24
25
26
27
28
29
30
31
32
33
34
35
36


def _sum(x: List[Tensor]) -> Tensor:
    res = x[0]
    for i in x[1:]:
        res = res + i
    return res


class RetinaNetHead(nn.Module):
    """
    A regression and classification head for use in RetinaNet.

37
    Args:
38
39
40
41
42
43
44
45
46
47
48
49
50
        in_channels (int): number of channels of the input feature
        num_anchors (int): number of anchors to be predicted
        num_classes (int): number of classes to be predicted
    """

    def __init__(self, in_channels, num_anchors, num_classes):
        super().__init__()
        self.classification_head = RetinaNetClassificationHead(in_channels, num_anchors, num_classes)
        self.regression_head = RetinaNetRegressionHead(in_channels, num_anchors)

    def compute_loss(self, targets, head_outputs, anchors, matched_idxs):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor], List[Tensor]) -> Dict[str, Tensor]
        return {
51
52
            "classification": self.classification_head.compute_loss(targets, head_outputs, matched_idxs),
            "bbox_regression": self.regression_head.compute_loss(targets, head_outputs, anchors, matched_idxs),
53
54
55
56
        }

    def forward(self, x):
        # type: (List[Tensor]) -> Dict[str, Tensor]
57
        return {"cls_logits": self.classification_head(x), "bbox_regression": self.regression_head(x)}
58
59
60
61
62
63


class RetinaNetClassificationHead(nn.Module):
    """
    A classification head for use in RetinaNet.

64
    Args:
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
        in_channels (int): number of channels of the input feature
        num_anchors (int): number of anchors to be predicted
        num_classes (int): number of classes to be predicted
    """

    def __init__(self, in_channels, num_anchors, num_classes, prior_probability=0.01):
        super().__init__()

        conv = []
        for _ in range(4):
            conv.append(nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1))
            conv.append(nn.ReLU())
        self.conv = nn.Sequential(*conv)

        for layer in self.conv.children():
            if isinstance(layer, nn.Conv2d):
                torch.nn.init.normal_(layer.weight, std=0.01)
                torch.nn.init.constant_(layer.bias, 0)

        self.cls_logits = nn.Conv2d(in_channels, num_anchors * num_classes, kernel_size=3, stride=1, padding=1)
        torch.nn.init.normal_(self.cls_logits.weight, std=0.01)
        torch.nn.init.constant_(self.cls_logits.bias, -math.log((1 - prior_probability) / prior_probability))

        self.num_classes = num_classes
        self.num_anchors = num_anchors

        # This is to fix using det_utils.Matcher.BETWEEN_THRESHOLDS in TorchScript.
        # TorchScript doesn't support class attributes.
        # https://github.com/pytorch/vision/pull/1697#issuecomment-630255584
        self.BETWEEN_THRESHOLDS = det_utils.Matcher.BETWEEN_THRESHOLDS

    def compute_loss(self, targets, head_outputs, matched_idxs):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor]) -> Tensor
        losses = []

100
        cls_logits = head_outputs["cls_logits"]
101
102
103
104
105

        for targets_per_image, cls_logits_per_image, matched_idxs_per_image in zip(targets, cls_logits, matched_idxs):
            # determine only the foreground
            foreground_idxs_per_image = matched_idxs_per_image >= 0
            num_foreground = foreground_idxs_per_image.sum()
106
107
108
109
110

            # create the target classification
            gt_classes_target = torch.zeros_like(cls_logits_per_image)
            gt_classes_target[
                foreground_idxs_per_image,
111
                targets_per_image["labels"][matched_idxs_per_image[foreground_idxs_per_image]],
112
113
114
115
            ] = 1.0

            # find indices for which anchors should be ignored
            valid_idxs_per_image = matched_idxs_per_image != self.BETWEEN_THRESHOLDS
116
117

            # compute the classification loss
118
119
120
121
122
123
124
125
            losses.append(
                sigmoid_focal_loss(
                    cls_logits_per_image[valid_idxs_per_image],
                    gt_classes_target[valid_idxs_per_image],
                    reduction="sum",
                )
                / max(1, num_foreground)
            )
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151

        return _sum(losses) / len(targets)

    def forward(self, x):
        # type: (List[Tensor]) -> Tensor
        all_cls_logits = []

        for features in x:
            cls_logits = self.conv(features)
            cls_logits = self.cls_logits(cls_logits)

            # Permute classification output from (N, A * K, H, W) to (N, HWA, K).
            N, _, H, W = cls_logits.shape
            cls_logits = cls_logits.view(N, -1, self.num_classes, H, W)
            cls_logits = cls_logits.permute(0, 3, 4, 1, 2)
            cls_logits = cls_logits.reshape(N, -1, self.num_classes)  # Size=(N, HWA, 4)

            all_cls_logits.append(cls_logits)

        return torch.cat(all_cls_logits, dim=1)


class RetinaNetRegressionHead(nn.Module):
    """
    A regression head for use in RetinaNet.

152
    Args:
153
154
155
        in_channels (int): number of channels of the input feature
        num_anchors (int): number of anchors to be predicted
    """
156

157
    __annotations__ = {
158
        "box_coder": det_utils.BoxCoder,
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
    }

    def __init__(self, in_channels, num_anchors):
        super().__init__()

        conv = []
        for _ in range(4):
            conv.append(nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1))
            conv.append(nn.ReLU())
        self.conv = nn.Sequential(*conv)

        self.bbox_reg = nn.Conv2d(in_channels, num_anchors * 4, kernel_size=3, stride=1, padding=1)
        torch.nn.init.normal_(self.bbox_reg.weight, std=0.01)
        torch.nn.init.zeros_(self.bbox_reg.bias)

        for layer in self.conv.children():
            if isinstance(layer, nn.Conv2d):
                torch.nn.init.normal_(layer.weight, std=0.01)
                torch.nn.init.zeros_(layer.bias)

        self.box_coder = det_utils.BoxCoder(weights=(1.0, 1.0, 1.0, 1.0))

    def compute_loss(self, targets, head_outputs, anchors, matched_idxs):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor], List[Tensor]) -> Tensor
        losses = []

185
        bbox_regression = head_outputs["bbox_regression"]
186

187
188
189
        for targets_per_image, bbox_regression_per_image, anchors_per_image, matched_idxs_per_image in zip(
            targets, bbox_regression, anchors, matched_idxs
        ):
190
            # determine only the foreground indices, ignore the rest
191
192
            foreground_idxs_per_image = torch.where(matched_idxs_per_image >= 0)[0]
            num_foreground = foreground_idxs_per_image.numel()
193
194

            # select only the foreground boxes
195
            matched_gt_boxes_per_image = targets_per_image["boxes"][matched_idxs_per_image[foreground_idxs_per_image]]
196
197
198
199
200
201
202
            bbox_regression_per_image = bbox_regression_per_image[foreground_idxs_per_image, :]
            anchors_per_image = anchors_per_image[foreground_idxs_per_image, :]

            # compute the regression targets
            target_regression = self.box_coder.encode_single(matched_gt_boxes_per_image, anchors_per_image)

            # compute the loss
203
204
205
206
            losses.append(
                torch.nn.functional.l1_loss(bbox_regression_per_image, target_regression, reduction="sum")
                / max(1, num_foreground)
            )
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239

        return _sum(losses) / max(1, len(targets))

    def forward(self, x):
        # type: (List[Tensor]) -> Tensor
        all_bbox_regression = []

        for features in x:
            bbox_regression = self.conv(features)
            bbox_regression = self.bbox_reg(bbox_regression)

            # Permute bbox regression output from (N, 4 * A, H, W) to (N, HWA, 4).
            N, _, H, W = bbox_regression.shape
            bbox_regression = bbox_regression.view(N, -1, 4, H, W)
            bbox_regression = bbox_regression.permute(0, 3, 4, 1, 2)
            bbox_regression = bbox_regression.reshape(N, -1, 4)  # Size=(N, HWA, 4)

            all_bbox_regression.append(bbox_regression)

        return torch.cat(all_bbox_regression, dim=1)


class RetinaNet(nn.Module):
    """
    Implements RetinaNet.

    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.

    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
    containing:
240
241
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
242
243
244
245
246
247
248
249
        - labels (Int64Tensor[N]): the class label for each ground-truth box

    The model returns a Dict[Tensor] during training, containing the classification and regression
    losses.

    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:
250
251
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
252
253
254
        - labels (Int64Tensor[N]): the predicted labels for each image
        - scores (Tensor[N]): the scores for each prediction

255
    Args:
256
257
258
259
        backbone (nn.Module): the network used to compute the features for the model.
            It should contain an 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 an OrderedDict[Tensor].
260
        num_classes (int): number of output classes of the model (including the background).
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
        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
        anchor_generator (AnchorGenerator): module that generates the anchors for a set of feature
            maps.
        head (nn.Module): Module run on top of the feature pyramid.
            Defaults to a module containing a classification and regression module.
        score_thresh (float): Score threshold used for postprocessing the detections.
        nms_thresh (float): NMS threshold used for postprocessing the detections.
        detections_per_img (int): Number of best detections to keep after NMS.
        fg_iou_thresh (float): minimum IoU between the anchor and the GT box so that they can be
            considered as positive during training.
        bg_iou_thresh (float): maximum IoU between the anchor and the GT box so that they can be
            considered as negative during training.
279
        topk_candidates (int): Number of best detections to keep before NMS.
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300

    Example:

        >>> import torch
        >>> import torchvision
        >>> from torchvision.models.detection import RetinaNet
        >>> from torchvision.models.detection.anchor_utils import AnchorGenerator
        >>> # load a pre-trained model for classification and return
        >>> # only the features
        >>> backbone = torchvision.models.mobilenet_v2(pretrained=True).features
        >>> # RetinaNet 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 network 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(
301
302
        >>>     sizes=((32, 64, 128, 256, 512),),
        >>>     aspect_ratios=((0.5, 1.0, 2.0),)
303
304
305
306
307
308
309
310
311
312
        >>> )
        >>>
        >>> # put the pieces together inside a RetinaNet model
        >>> model = RetinaNet(backbone,
        >>>                   num_classes=2,
        >>>                   anchor_generator=anchor_generator)
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
    """
313

314
    __annotations__ = {
315
316
        "box_coder": det_utils.BoxCoder,
        "proposal_matcher": det_utils.Matcher,
317
318
    }

319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
    def __init__(
        self,
        backbone,
        num_classes,
        # transform parameters
        min_size=800,
        max_size=1333,
        image_mean=None,
        image_std=None,
        # Anchor parameters
        anchor_generator=None,
        head=None,
        proposal_matcher=None,
        score_thresh=0.05,
        nms_thresh=0.5,
        detections_per_img=300,
        fg_iou_thresh=0.5,
        bg_iou_thresh=0.4,
        topk_candidates=1000,
    ):
339
        super().__init__()
340
        _log_api_usage_once(self)
341
342
343
344
345

        if not hasattr(backbone, "out_channels"):
            raise ValueError(
                "backbone should contain an attribute out_channels "
                "specifying the number of output channels (assumed to be the "
346
347
                "same for all the levels)"
            )
348
349
350
351
352
353
354
        self.backbone = backbone

        assert isinstance(anchor_generator, (AnchorGenerator, type(None)))

        if anchor_generator is None:
            anchor_sizes = tuple((x, int(x * 2 ** (1.0 / 3)), int(x * 2 ** (2.0 / 3))) for x in [32, 64, 128, 256, 512])
            aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)
355
            anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
        self.anchor_generator = anchor_generator

        if head is None:
            head = RetinaNetHead(backbone.out_channels, anchor_generator.num_anchors_per_location()[0], num_classes)
        self.head = head

        if proposal_matcher is None:
            proposal_matcher = det_utils.Matcher(
                fg_iou_thresh,
                bg_iou_thresh,
                allow_low_quality_matches=True,
            )
        self.proposal_matcher = proposal_matcher

        self.box_coder = det_utils.BoxCoder(weights=(1.0, 1.0, 1.0, 1.0))

        if image_mean is None:
            image_mean = [0.485, 0.456, 0.406]
        if image_std is None:
            image_std = [0.229, 0.224, 0.225]
        self.transform = GeneralizedRCNNTransform(min_size, max_size, image_mean, image_std)

        self.score_thresh = score_thresh
        self.nms_thresh = nms_thresh
        self.detections_per_img = detections_per_img
381
        self.topk_candidates = topk_candidates
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397

        # used only on torchscript mode
        self._has_warned = False

    @torch.jit.unused
    def eager_outputs(self, losses, detections):
        # type: (Dict[str, Tensor], List[Dict[str, Tensor]]) -> Tuple[Dict[str, Tensor], List[Dict[str, Tensor]]]
        if self.training:
            return losses

        return detections

    def compute_loss(self, targets, head_outputs, anchors):
        # type: (List[Dict[str, Tensor]], Dict[str, Tensor], List[Tensor]) -> Dict[str, Tensor]
        matched_idxs = []
        for anchors_per_image, targets_per_image in zip(anchors, targets):
398
399
400
401
            if targets_per_image["boxes"].numel() == 0:
                matched_idxs.append(
                    torch.full((anchors_per_image.size(0),), -1, dtype=torch.int64, device=anchors_per_image.device)
                )
402
403
                continue

404
            match_quality_matrix = box_ops.box_iou(targets_per_image["boxes"], anchors_per_image)
405
406
407
408
409
            matched_idxs.append(self.proposal_matcher(match_quality_matrix))

        return self.head.compute_loss(targets, head_outputs, anchors, matched_idxs)

    def postprocess_detections(self, head_outputs, anchors, image_shapes):
410
        # type: (Dict[str, List[Tensor]], List[List[Tensor]], List[Tuple[int, int]]) -> List[Dict[str, Tensor]]
411
412
        class_logits = head_outputs["cls_logits"]
        box_regression = head_outputs["bbox_regression"]
413

414
        num_images = len(image_shapes)
415

416
        detections: List[Dict[str, Tensor]] = []
417

418
419
420
421
        for index in range(num_images):
            box_regression_per_image = [br[index] for br in box_regression]
            logits_per_image = [cl[index] for cl in class_logits]
            anchors_per_image, image_shape = anchors[index], image_shapes[index]
422
423
424
425
426

            image_boxes = []
            image_scores = []
            image_labels = []

427
428
429
            for box_regression_per_level, logits_per_level, anchors_per_level in zip(
                box_regression_per_image, logits_per_image, anchors_per_image
            ):
430
431
                num_classes = logits_per_level.shape[-1]

432
                # remove low scoring boxes
433
434
435
436
                scores_per_level = torch.sigmoid(logits_per_level).flatten()
                keep_idxs = scores_per_level > self.score_thresh
                scores_per_level = scores_per_level[keep_idxs]
                topk_idxs = torch.where(keep_idxs)[0]
437

438
439
440
441
                # keep only topk scoring predictions
                num_topk = min(self.topk_candidates, topk_idxs.size(0))
                scores_per_level, idxs = scores_per_level.topk(num_topk)
                topk_idxs = topk_idxs[idxs]
442

443
                anchor_idxs = torch.div(topk_idxs, num_classes, rounding_mode="floor")
444
                labels_per_level = topk_idxs % num_classes
445

446
447
448
                boxes_per_level = self.box_coder.decode_single(
                    box_regression_per_level[anchor_idxs], anchors_per_level[anchor_idxs]
                )
449
450
451
452
453
                boxes_per_level = box_ops.clip_boxes_to_image(boxes_per_level, image_shape)

                image_boxes.append(boxes_per_level)
                image_scores.append(scores_per_level)
                image_labels.append(labels_per_level)
454

455
456
457
            image_boxes = torch.cat(image_boxes, dim=0)
            image_scores = torch.cat(image_scores, dim=0)
            image_labels = torch.cat(image_labels, dim=0)
458

459
460
            # non-maximum suppression
            keep = box_ops.batched_nms(image_boxes, image_scores, image_labels, self.nms_thresh)
461
462
463
464
465
466
467
468
469
            keep = keep[: self.detections_per_img]

            detections.append(
                {
                    "boxes": image_boxes[keep],
                    "scores": image_scores[keep],
                    "labels": image_labels[keep],
                }
            )
470
471
472
473
474
475

        return detections

    def forward(self, images, targets=None):
        # type: (List[Tensor], Optional[List[Dict[str, Tensor]]]) -> Tuple[Dict[str, Tensor], List[Dict[str, Tensor]]]
        """
476
        Args:
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
            images (list[Tensor]): images to be processed
            targets (list[Dict[Tensor]]): ground-truth boxes present in the image (optional)

        Returns:
            result (list[BoxList] or dict[Tensor]): the output from the model.
                During training, it returns a dict[Tensor] which contains the losses.
                During testing, it returns list[BoxList] contains additional fields
                like `scores`, `labels` and `mask` (for Mask R-CNN models).

        """
        if self.training and targets is None:
            raise ValueError("In training mode, targets should be passed")

        if self.training:
            assert targets is not None
            for target in targets:
                boxes = target["boxes"]
                if isinstance(boxes, torch.Tensor):
                    if len(boxes.shape) != 2 or boxes.shape[-1] != 4:
496
                        raise ValueError(f"Expected target boxes to be a tensor of shape [N, 4], got {boxes.shape}.")
497
                else:
498
                    raise ValueError(f"Expected target boxes to be of type Tensor, got {type(boxes)}.")
499
500

        # get the original image sizes
501
        original_image_sizes: List[Tuple[int, int]] = []
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
        for img in images:
            val = img.shape[-2:]
            assert len(val) == 2
            original_image_sizes.append((val[0], val[1]))

        # transform the input
        images, targets = self.transform(images, targets)

        # Check for degenerate boxes
        # TODO: Move this to a function
        if targets is not None:
            for target_idx, target in enumerate(targets):
                boxes = target["boxes"]
                degenerate_boxes = boxes[:, 2:] <= boxes[:, :2]
                if degenerate_boxes.any():
                    # print the first degenerate box
                    bb_idx = torch.where(degenerate_boxes.any(dim=1))[0][0]
                    degen_bb: List[float] = boxes[bb_idx].tolist()
520
521
                    raise ValueError(
                        "All bounding boxes should have positive height and width."
522
                        f" Found invalid box {degen_bb} for target at index {target_idx}."
523
                    )
524
525
526
527

        # get the features from the backbone
        features = self.backbone(images.tensors)
        if isinstance(features, torch.Tensor):
528
            features = OrderedDict([("0", features)])
529
530
531
532
533
534
535
536
537
538
539

        # TODO: Do we want a list or a dict?
        features = list(features.values())

        # compute the retinanet heads outputs using the features
        head_outputs = self.head(features)

        # create the set of anchors
        anchors = self.anchor_generator(images, features)

        losses = {}
540
        detections: List[Dict[str, Tensor]] = []
541
542
543
544
545
546
        if self.training:
            assert targets is not None

            # compute the losses
            losses = self.compute_loss(targets, head_outputs, anchors)
        else:
547
548
549
550
551
            # recover level sizes
            num_anchors_per_level = [x.size(2) * x.size(3) for x in features]
            HW = 0
            for v in num_anchors_per_level:
                HW += v
552
            HWA = head_outputs["cls_logits"].size(1)
553
554
555
556
557
558
559
560
561
            A = HWA // HW
            num_anchors_per_level = [hw * A for hw in num_anchors_per_level]

            # split outputs per level
            split_head_outputs: Dict[str, List[Tensor]] = {}
            for k in head_outputs:
                split_head_outputs[k] = list(head_outputs[k].split(num_anchors_per_level, dim=1))
            split_anchors = [list(a.split(num_anchors_per_level)) for a in anchors]

562
            # compute the detections
563
            detections = self.postprocess_detections(split_head_outputs, split_anchors, images.image_sizes)
564
565
566
567
568
569
            detections = self.transform.postprocess(detections, images.image_sizes, original_image_sizes)

        if torch.jit.is_scripting():
            if not self._has_warned:
                warnings.warn("RetinaNet always returns a (Losses, Detections) tuple in scripting")
                self._has_warned = True
570
            return losses, detections
571
572
573
574
        return self.eager_outputs(losses, detections)


model_urls = {
575
    "retinanet_resnet50_fpn_coco": "https://download.pytorch.org/models/retinanet_resnet50_fpn_coco-eeacb38b.pth",
576
577
578
}


579
580
581
def retinanet_resnet50_fpn(
    pretrained=False, progress=True, num_classes=91, pretrained_backbone=True, trainable_backbone_layers=None, **kwargs
):
582
583
584
    """
    Constructs a RetinaNet model with a ResNet-50-FPN backbone.

585
586
    Reference: `"Focal Loss for Dense Object Detection" <https://arxiv.org/abs/1708.02002>`_.

587
588
589
590
591
592
593
    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.

    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
    containing:
594

595
596
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
597
598
599
600
601
602
603
        - labels (``Int64Tensor[N]``): the class label for each ground-truth box

    The model returns a ``Dict[Tensor]`` during training, containing the classification and regression
    losses.

    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
604
    follows, where ``N`` is the number of detections:
605

606
607
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
608
609
610
611
        - labels (``Int64Tensor[N]``): the predicted labels for each detection
        - scores (``Tensor[N]``): the scores of each detection

    For more details on the output, you may refer to :ref:`instance_seg_output`.
612
613
614
615
616
617
618
619

    Example::

        >>> model = torchvision.models.detection.retinanet_resnet50_fpn(pretrained=True)
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)

620
    Args:
621
622
        pretrained (bool): If True, returns a model pre-trained on COCO train2017
        progress (bool): If True, displays a progress bar of the download to stderr
623
624
625
626
        num_classes (int): number of output classes of the model (including the background)
        pretrained_backbone (bool): If True, returns a model with backbone pre-trained on Imagenet
        trainable_backbone_layers (int): number of trainable (not frozen) resnet layers starting from final block.
            Valid values are between 0 and 5, with 5 meaning all backbone layers are trainable.
627
    """
628
    trainable_backbone_layers = _validate_trainable_layers(
629
630
        pretrained or pretrained_backbone, trainable_backbone_layers, 5, 3
    )
631

632
633
634
    if pretrained:
        # no need to download the backbone if pretrained is set
        pretrained_backbone = False
635
636

    backbone = resnet50(pretrained=pretrained_backbone, progress=progress, norm_layer=misc_nn_ops.FrozenBatchNorm2d)
637
    # skip P2 because it generates too many anchors (according to their paper)
638
639
    backbone = _resnet_fpn_extractor(
        backbone, trainable_backbone_layers, returned_layers=[2, 3, 4], extra_blocks=LastLevelP6P7(256, 256)
640
    )
641
642
    model = RetinaNet(backbone, num_classes, **kwargs)
    if pretrained:
643
        state_dict = load_state_dict_from_url(model_urls["retinanet_resnet50_fpn_coco"], progress=progress)
644
        model.load_state_dict(state_dict)
645
        overwrite_eps(model, 0.0)
646
    return model