keypoint_rcnn.py 17.4 KB
Newer Older
1
2
3
4
5
import torch
from torch import nn

from torchvision.ops import MultiScaleRoIAlign

6
7
from ..utils import load_state_dict_from_url

8
9
10
11
12
13
14
15
16
17
from .faster_rcnn import FasterRCNN
from .backbone_utils import resnet_fpn_backbone


__all__ = [
    "KeypointRCNN", "keypointrcnn_resnet50_fpn"
]


class KeypointRCNN(FasterRCNN):
18
19
20
21
22
23
24
25
    """
    Implements Keypoint 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.

26
    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
27
    containing:
28
29
        - boxes (FloatTensor[N, 4]): the ground-truth boxes in [x1, y1, x2, y2] format, with values of x
          between 0 and W and values of y between 0 and H
30
31
        - labels (Int64Tensor[N]): the class label for each ground-truth box
        - keypoints (FloatTensor[N, K, 3]): the K keypoints location for each of the N instances, in the
32
33
          format [x, y, visibility], where visibility=0 means that the keypoint is not visible.

34
35
36
37
38
39
    The model returns a Dict[Tensor] during training, containing the classification and regression
    losses for both the RPN and the R-CNN, and the keypoint 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:
40
41
        - boxes (FloatTensor[N, 4]): the predicted boxes in [x1, y1, x2, y2] format, with values of x
          between 0 and W and values of y between 0 and H
42
        - labels (Int64Tensor[N]): the predicted labels for each image
43
        - scores (Tensor[N]): the scores or each prediction
44
        - keypoints (FloatTensor[N, K, 3]): the locations of the predicted keypoints, in [x, y, v] format.
45
46
47
48
49
50
51
52
53
54
55
56
57
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
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102

    Arguments:
        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
        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
        keypoint_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 keypoint head.
        keypoint_head (nn.Module): module that takes the cropped feature maps as input
        keypoint_predictor (nn.Module): module that takes the output of the keypoint_head and returns the
            heatmap logits

    Example::

103
        >>> import torch
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
        >>> import torchvision
        >>> from torchvision.models.detection import KeypointRCNN
        >>> from torchvision.models.detection.rpn import AnchorGenerator
        >>>
        >>> # load a pre-trained model for classification and return
        >>> # only the features
        >>> backbone = torchvision.models.mobilenet_v2(pretrained=True).features
        >>> # KeypointRCNN 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
128
        >>> # be ['0']. More generally, the backbone should return an
129
130
        >>> # OrderedDict[Tensor], and in featmap_names you can choose which
        >>> # feature maps to use.
131
        >>> roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
132
133
134
        >>>                                                 output_size=7,
        >>>                                                 sampling_ratio=2)
        >>>
135
        >>> keypoint_roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
136
137
        >>>                                                          output_size=14,
        >>>                                                          sampling_ratio=2)
138
        >>> # put the pieces together inside a KeypointRCNN model
139
140
141
142
143
144
        >>> model = KeypointRCNN(backbone,
        >>>                      num_classes=2,
        >>>                      rpn_anchor_generator=anchor_generator,
        >>>                      box_roi_pool=roi_pooler,
        >>>                      keypoint_roi_pool=keypoint_roi_pooler)
        >>> model.eval()
145
146
147
148
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
    """
149
150
    def __init__(self, backbone, num_classes=None,
                 # transform parameters
151
                 min_size=None, max_size=1333,
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
                 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,
                 # 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,
                 # keypoint parameters
                 keypoint_roi_pool=None, keypoint_head=None, keypoint_predictor=None,
                 num_keypoints=17):

        assert isinstance(keypoint_roi_pool, (MultiScaleRoIAlign, type(None)))
171
172
        if min_size is None:
            min_size = (640, 672, 704, 736, 768, 800)
173
174
175
176
177
178
179
180
181

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

        out_channels = backbone.out_channels

        if keypoint_roi_pool is None:
            keypoint_roi_pool = MultiScaleRoIAlign(
eellison's avatar
eellison committed
182
                featmap_names=['0', '1', '2', '3'],
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
                output_size=14,
                sampling_ratio=2)

        if keypoint_head is None:
            keypoint_layers = tuple(512 for _ in range(8))
            keypoint_head = KeypointRCNNHeads(out_channels, keypoint_layers)

        if keypoint_predictor is None:
            keypoint_dim_reduced = 512  # == keypoint_layers[-1]
            keypoint_predictor = KeypointRCNNPredictor(keypoint_dim_reduced, num_keypoints)

        super(KeypointRCNN, self).__init__(
            backbone, num_classes,
            # transform parameters
            min_size, max_size,
            image_mean, image_std,
            # RPN-specific parameters
            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,
            rpn_nms_thresh,
            rpn_fg_iou_thresh, rpn_bg_iou_thresh,
            rpn_batch_size_per_image, rpn_positive_fraction,
            # Box parameters
            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)

        self.roi_heads.keypoint_roi_pool = keypoint_roi_pool
        self.roi_heads.keypoint_head = keypoint_head
        self.roi_heads.keypoint_predictor = keypoint_predictor


class KeypointRCNNHeads(nn.Sequential):
    def __init__(self, in_channels, layers):
        d = []
        next_feature = in_channels
Francisco Massa's avatar
Francisco Massa committed
222
        for out_channels in layers:
223
            d.append(nn.Conv2d(next_feature, out_channels, 3, stride=1, padding=1))
224
            d.append(nn.ReLU(inplace=True))
Francisco Massa's avatar
Francisco Massa committed
225
            next_feature = out_channels
226
227
        super(KeypointRCNNHeads, self).__init__(*d)
        for m in self.children():
228
            if isinstance(m, nn.Conv2d):
229
230
231
232
233
234
235
236
237
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
                nn.init.constant_(m.bias, 0)


class KeypointRCNNPredictor(nn.Module):
    def __init__(self, in_channels, num_keypoints):
        super(KeypointRCNNPredictor, self).__init__()
        input_features = in_channels
        deconv_kernel = 4
238
        self.kps_score_lowres = nn.ConvTranspose2d(
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
            input_features,
            num_keypoints,
            deconv_kernel,
            stride=2,
            padding=deconv_kernel // 2 - 1,
        )
        nn.init.kaiming_normal_(
            self.kps_score_lowres.weight, mode="fan_out", nonlinearity="relu"
        )
        nn.init.constant_(self.kps_score_lowres.bias, 0)
        self.up_scale = 2
        self.out_channels = num_keypoints

    def forward(self, x):
        x = self.kps_score_lowres(x)
254
        x = torch.nn.functional.interpolate(
eellison's avatar
eellison committed
255
            x, scale_factor=float(self.up_scale), mode="bilinear", align_corners=False
256
257
258
259
        )
        return x


260
model_urls = {
261
262
    # legacy model for BC reasons, see https://github.com/pytorch/vision/issues/1606
    'keypointrcnn_resnet50_fpn_coco_legacy':
263
        'https://download.pytorch.org/models/keypointrcnn_resnet50_fpn_coco-9f466800.pth',
264
265
    'keypointrcnn_resnet50_fpn_coco':
        'https://download.pytorch.org/models/keypointrcnn_resnet50_fpn_coco-fc266e95.pth',
266
267
268
269
270
}


def keypointrcnn_resnet50_fpn(pretrained=False, progress=True,
                              num_classes=2, num_keypoints=17,
271
                              pretrained_backbone=True, trainable_backbone_layers=3, **kwargs):
272
273
274
    """
    Constructs a Keypoint R-CNN model with a ResNet-50-FPN backbone.

275
276
277
278
279
    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.

280
    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
281
    containing:
282
283
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with values of ``x``
          between ``0`` and ``W`` and values of ``y`` between ``0`` and ``H``
284
285
        - labels (``Int64Tensor[N]``): the class label for each ground-truth box
        - keypoints (``FloatTensor[N, K, 3]``): the ``K`` keypoints location for each of the ``N`` instances, in the
286
287
288
289
290
291
292
293
          format ``[x, y, visibility]``, where ``visibility=0`` means that the keypoint is not visible.

    The model returns a ``Dict[Tensor]`` during training, containing the classification and regression
    losses for both the RPN and the R-CNN, and the keypoint 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:
294
295
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format,  with values of ``x``
          between ``0`` and ``W`` and values of ``y`` between ``0`` and ``H``
296
        - labels (``Int64Tensor[N]``): the predicted labels for each image
297
        - scores (``Tensor[N]``): the scores or each prediction
298
        - keypoints (``FloatTensor[N, K, 3]``): the locations of the predicted keypoints, in ``[x, y, v]`` format.
299

300
301
    Keypoint R-CNN is exportable to ONNX for a fixed batch size with inputs images of fixed size.

302
303
304
305
306
307
    Example::

        >>> model = torchvision.models.detection.keypointrcnn_resnet50_fpn(pretrained=True)
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
308
309
310
        >>>
        >>> # optionally, if you want to export the model to ONNX:
        >>> torch.onnx.export(model, x, "keypoint_rcnn.onnx", opset_version = 11)
311

312
313
314
    Arguments:
        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
315
316
317
318
        pretrained_backbone (bool): If True, returns a model with backbone pre-trained on Imagenet
        num_classes (int): number of output classes of the model (including the background)
        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.
319
    """
320
321
322
323
    assert trainable_backbone_layers <= 5 and trainable_backbone_layers >= 0
    # dont freeze any layers if pretrained model or backbone is not used
    if not (pretrained or pretrained_backbone):
        trainable_backbone_layers = 5
324
325
326
    if pretrained:
        # no need to download the backbone if pretrained is set
        pretrained_backbone = False
327
    backbone = resnet_fpn_backbone('resnet50', pretrained_backbone, trainable_layers=trainable_backbone_layers)
328
329
    model = KeypointRCNN(backbone, num_classes, num_keypoints=num_keypoints, **kwargs)
    if pretrained:
330
331
332
333
        key = 'keypointrcnn_resnet50_fpn_coco'
        if pretrained == 'legacy':
            key += '_legacy'
        state_dict = load_state_dict_from_url(model_urls[key],
334
335
                                              progress=progress)
        model.load_state_dict(state_dict)
336
    return model