mask_rcnn.py 17.4 KB
Newer Older
1
2
3
4
5
6
from collections import OrderedDict

from torch import nn

from torchvision.ops import MultiScaleRoIAlign

7
from ._utils import overwrite_eps
8
9
from ..utils import load_state_dict_from_url

10
from .faster_rcnn import FasterRCNN
11
from .backbone_utils import resnet_fpn_backbone, _validate_trainable_layers
12
13

__all__ = [
14
    "MaskRCNN", "maskrcnn_resnet50_fpn",
15
16
17
18
]


class MaskRCNN(FasterRCNN):
19
20
21
22
23
24
25
26
    """
    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.

27
    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
28
    containing:
29
30
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
31
        - labels (Int64Tensor[N]): the class label for each ground-truth box
32
        - masks (UInt8Tensor[N, H, W]): the segmentation binary masks for each instance
33

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 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:
40
41
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
42
        - labels (Int64Tensor[N]): the predicted labels for each image
43
        - scores (Tensor[N]): the scores or each prediction
44
        - masks (UInt8Tensor[N, 1, H, W]): the predicted masks for each instance, in 0-1 range. In order to
45
46
          obtain the final segmentation masks, the soft masks can be thresholded, generally
          with a value of 0.5 (mask >= 0.5)
47

48
    Args:
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
        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
78
79
        rpn_score_thresh (float): during inference, only return proposals with a classification score
            greater than rpn_score_thresh
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
        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::

107
        >>> import torch
108
109
        >>> import torchvision
        >>> from torchvision.models.detection import MaskRCNN
110
        >>> from torchvision.models.detection.anchor_utils import AnchorGenerator
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
        >>>
        >>> # load a pre-trained model for classification and return
        >>> # only the features
        >>> backbone = torchvision.models.mobilenet_v2(pretrained=True).features
        >>> # 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
132
        >>> # be ['0']. More generally, the backbone should return an
133
134
        >>> # OrderedDict[Tensor], and in featmap_names you can choose which
        >>> # feature maps to use.
135
        >>> roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
136
137
138
        >>>                                                 output_size=7,
        >>>                                                 sampling_ratio=2)
        >>>
139
        >>> mask_roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0'],
140
141
        >>>                                                      output_size=14,
        >>>                                                      sampling_ratio=2)
142
        >>> # put the pieces together inside a MaskRCNN model
143
144
145
146
147
        >>> model = MaskRCNN(backbone,
        >>>                  num_classes=2,
        >>>                  rpn_anchor_generator=anchor_generator,
        >>>                  box_roi_pool=roi_pooler,
        >>>                  mask_roi_pool=mask_roi_pooler)
148
149
150
151
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
    """
152
153
154
155
156
157
158
159
160
161
162
    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,
163
                 rpn_score_thresh=0.0,
164
165
166
167
168
169
170
                 # 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
171
                 mask_roi_pool=None, mask_head=None, mask_predictor=None):
172
173
174
175
176
177
178
179
180
181
182

        assert isinstance(mask_roi_pool, (MultiScaleRoIAlign, type(None)))

        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:
            mask_roi_pool = MultiScaleRoIAlign(
eellison's avatar
eellison committed
183
                featmap_names=['0', '1', '2', '3'],
184
185
186
187
188
189
190
191
192
                output_size=14,
                sampling_ratio=2)

        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:
193
194
195
196
            mask_predictor_in_channels = 256  # == mask_layers[-1]
            mask_dim_reduced = 256
            mask_predictor = MaskRCNNPredictor(mask_predictor_in_channels,
                                               mask_dim_reduced, num_classes)
197
198
199
200
201
202
203
204
205
206
207
208
209

        super(MaskRCNN, 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,
210
            rpn_score_thresh,
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
            # 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.mask_roi_pool = mask_roi_pool
        self.roi_heads.mask_head = mask_head
        self.roi_heads.mask_predictor = mask_predictor


class MaskRCNNHeads(nn.Sequential):
    def __init__(self, in_channels, layers, dilation):
        """
226
        Args:
227
228
229
            in_channels (int): number of input channels
            layers (list): feature dimensions of each FCN layer
            dilation (int): dilation rate of kernel
230
231
232
233
        """
        d = OrderedDict()
        next_feature = in_channels
        for layer_idx, layer_features in enumerate(layers, 1):
234
            d["mask_fcn{}".format(layer_idx)] = nn.Conv2d(
235
236
237
238
239
240
241
242
243
244
245
246
247
                next_feature, layer_features, kernel_size=3,
                stride=1, padding=dilation, dilation=dilation)
            d["relu{}".format(layer_idx)] = nn.ReLU(inplace=True)
            next_feature = layer_features

        super(MaskRCNNHeads, self).__init__(d)
        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)


248
class MaskRCNNPredictor(nn.Sequential):
249
    def __init__(self, in_channels, dim_reduced, num_classes):
250
        super(MaskRCNNPredictor, self).__init__(OrderedDict([
251
            ("conv5_mask", nn.ConvTranspose2d(in_channels, dim_reduced, 2, 2, 0)),
252
            ("relu", nn.ReLU(inplace=True)),
253
            ("mask_fcn_logits", nn.Conv2d(dim_reduced, num_classes, 1, 1, 0)),
254
255
256
257
258
259
260
261
262
        ]))

        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)


263
264
265
266
267
268
269
model_urls = {
    'maskrcnn_resnet50_fpn_coco':
        'https://download.pytorch.org/models/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth',
}


def maskrcnn_resnet50_fpn(pretrained=False, progress=True,
270
                          num_classes=91, pretrained_backbone=True, trainable_backbone_layers=None, **kwargs):
271
272
273
    """
    Constructs a Mask R-CNN model with a ResNet-50-FPN backbone.

274
275
276
277
278
    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.

279
    During training, the model expects both the input tensors, as well as a targets (list of dictionary),
280
    containing:
281

282
283
        - boxes (``FloatTensor[N, 4]``): the ground-truth boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
284
        - labels (``Int64Tensor[N]``): the class label for each ground-truth box
285
        - masks (``UInt8Tensor[N, H, W]``): the segmentation binary masks for each instance
286
287
288
289
290
291

    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
292
    follows, where ``N`` is the number of detected instances:
293

294
295
        - boxes (``FloatTensor[N, 4]``): the predicted boxes in ``[x1, y1, x2, y2]`` format, with
          ``0 <= x1 < x2 <= W`` and ``0 <= y1 < y2 <= H``.
296
297
        - labels (``Int64Tensor[N]``): the predicted labels for each instance
        - scores (``Tensor[N]``): the scores or each instance
298
        - masks (``UInt8Tensor[N, 1, H, W]``): the predicted masks for each instance, in ``0-1`` range. In order to
299
300
301
          obtain the final segmentation masks, the soft masks can be thresholded, generally
          with a value of 0.5 (``mask >= 0.5``)

302
303
    For more details on the output and on how to plot the masks, you may refer to :ref:`instance_seg_output`.

304
305
    Mask R-CNN is exportable to ONNX for a fixed batch size with inputs images of fixed size.

306
307
308
309
310
311
    Example::

        >>> model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)
        >>> model.eval()
        >>> x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
        >>> predictions = model(x)
312
313
314
        >>>
        >>> # optionally, if you want to export the model to ONNX:
        >>> torch.onnx.export(model, x, "mask_rcnn.onnx", opset_version = 11)
315

316
    Args:
317
318
        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
319
        num_classes (int): number of output classes of the model (including the background)
320
        pretrained_backbone (bool): If True, returns a model with backbone pre-trained on Imagenet
321
322
        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.
323
    """
324
325
    trainable_backbone_layers = _validate_trainable_layers(
        pretrained or pretrained_backbone, trainable_backbone_layers, 5, 3)
326

327
328
329
    if pretrained:
        # no need to download the backbone if pretrained is set
        pretrained_backbone = False
330
    backbone = resnet_fpn_backbone('resnet50', pretrained_backbone, trainable_layers=trainable_backbone_layers)
331
332
    model = MaskRCNN(backbone, num_classes, **kwargs)
    if pretrained:
333
334
335
        state_dict = load_state_dict_from_url(model_urls['maskrcnn_resnet50_fpn_coco'],
                                              progress=progress)
        model.load_state_dict(state_dict)
336
        overwrite_eps(model, 0.0)
337
    return model