detection.py 123.5 KB
Newer Older
1
#  Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
2 3 4 5 6
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
7
#    http://www.apache.org/licenses/LICENSE-2.0
8 9 10 11 12 13 14 15 16 17
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
All layers just related to the detection neural network.
"""

18 19
from __future__ import print_function

20 21
from .layer_function_generator import generate_layer_fn
from .layer_function_generator import autodoc, templatedoc
22
from ..layer_helper import LayerHelper
D
dengkaipeng 已提交
23
from ..framework import Variable
24 25
from . import tensor
from . import nn
26
from . import ops
M
minqiyang 已提交
27
from ... import compat as cpt
C
chengduoZH 已提交
28
import math
M
minqiyang 已提交
29
import six
30
import numpy
31
from functools import reduce
32

C
chengduoZH 已提交
33
__all__ = [
34 35 36 37 38 39 40 41
    'prior_box',
    'density_prior_box',
    'multi_box_head',
    'bipartite_match',
    'target_assign',
    'detection_output',
    'ssd_loss',
    'rpn_target_assign',
42
    'retinanet_target_assign',
43
    'sigmoid_focal_loss',
44 45 46 47
    'anchor_generator',
    'roi_perspective_transform',
    'generate_proposal_labels',
    'generate_proposals',
48
    'generate_mask_labels',
49 50 51 52
    'iou_similarity',
    'box_coder',
    'polygon_box_transform',
    'yolov3_loss',
D
dengkaipeng 已提交
53
    'yolo_box',
54
    'box_clip',
J
jerrywgz 已提交
55
    'multiclass_nms',
56
    'retinanet_detection_output',
57
    'distribute_fpn_proposals',
58
    'box_decoder_and_assign',
59
    'collect_fpn_proposals',
C
chengduoZH 已提交
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220
def retinanet_target_assign(bbox_pred,
                            cls_logits,
                            anchor_box,
                            anchor_var,
                            gt_boxes,
                            gt_labels,
                            is_crowd,
                            im_info,
                            num_classes=1,
                            positive_overlap=0.5,
                            negative_overlap=0.4):
    """
    **Target Assign Layer for Retinanet .**

    This layer can be, for given the Intersection-over-Union (IoU) overlap
    between anchors and ground truth boxes, to assign classification and
    regression targets to each anchor, these target labels are used for training
    retinanet. Every anchor is assigned with a length :attr:`num_classes`
    one-hot vector of classification targets, and a 4-vector of box regression
    targets. The assignment rules are as followed:
    
    1. Anchors are assigned to ground-truth boxes when: (i) it has the highest
    IoU overlap with a ground-truth box, or (ii) it has an IoU overlap higher
    than positive_overlap(0.5) with any ground-truth box.
    
    2. Anchors are assigned to background when its IoU ratio is lower than
    negative_overlap (0.4) for all ground-truth boxes.
    
    When an anchor is assigned with a ground-truth box which is the i-th category,
    the i-th entry in its C vector of targets is set to 1 and all other entries
    are set to 0. When an anchor is assigned with background, all entries are set
    to 0. Anchors that are not assigned do not contribute to the training
    objective. The regression targets are the encoded ground-truth boxes
    associated with the assigned anchors.
 
    Args:
        bbox_pred(Variable): A 3-D Tensor with shape [N, M, 4] represents the
            predicted locations of M bounding bboxes. N is the batch size,
            and each bounding box has four coordinate values and the layout
            is [xmin, ymin, xmax, ymax].
        cls_logits(Variable): A 3-D Tensor with shape [N, M, C] represents the
            predicted confidence predictions. N is the batch size, C is the
            number of classes (excluding background), M is number of bounding boxes.
        anchor_box(Variable): A 2-D Tensor with shape [M, 4] holds M boxes,
            each box is represented as [xmin, ymin, xmax, ymax],
            [xmin, ymin] is the left top coordinate of the anchor box,
            if the input is image feature map, they are close to the origin
            of the coordinate system. [xmax, ymax] is the right bottom
            coordinate of the anchor box.
        anchor_var(Variable): A 2-D Tensor with shape [M,4] holds expanded 
            variances of anchors.
        gt_boxes(Variable): The ground-truth bounding boxes (bboxes) are a 2D
            LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth
            bboxes of mini-batch input.
        gt_labels(variable): The ground-truth labels are a 2D LoDTensor with
            shape [Ng, 1], Ng is the total number of ground-truth labels of
            mini-batch input.
        is_crowd(Variable): A 1-D LoDTensor which indicates ground-truth is crowd.
        im_info(Variable): A 2-D LoDTensor with shape [N, 3]. N is the batch size,
            3 is the height, width and scale.
        num_classes(int32): The number of classes.
        positive_overlap(float): Minimum overlap required between an anchor
            and ground-truth box for the (anchor, gt box) pair to be a positive
            example.
        negative_overlap(float): Maximum overlap allowed between an anchor
            and ground-truth box for the (anchor, gt box) pair to be a negative
            examples.

    Returns:
        tuple:
               A tuple(predicted_scores, predicted_location, target_label,
               target_bbox, bbox_inside_weight, fg_num) is returned. The
               predicted_scores and predicted_location are the predicted result
               of the retinanet.The target_label and target_bbox are the ground
               truth, respectively. The predicted_location is a 2D Tensor with
               shape [F, 4], and the shape of target_bbox is same as the shape of
               the predicted_location, F is the number of the foreground
               anchors. The predicted_scores is a 2D Tensor with shape
               [F + B, C], and the shape of target_label is [F + B, 1], B is the
               number of the background anchors, the F and B is depends on the
               input of this operator. Bbox_inside_weight represents whether the
               predicted location is fake foreground or not and the shape is [F, 4].
               Fg_num is the foreground number (including fake foreground) which
               is needed by focal loss.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          bbox_pred = layers.data(name='bbox_pred', shape=[1, 100, 4],
                            append_batch_size=False, dtype='float32')
          cls_logits = layers.data(name='cls_logits', shape=[1, 100, 10],
                            append_batch_size=False, dtype='float32')
          anchor_box = layers.data(name='anchor_box', shape=[100, 4],
                            append_batch_size=False, dtype='float32')
          anchor_var = layers.data(name='anchor_var', shape=[100, 4],
                            append_batch_size=False, dtype='float32')
          gt_boxes = layers.data(name='gt_boxes', shape=[10, 4],
                            append_batch_size=False, dtype='float32')
          gt_labels = layers.data(name='gt_labels', shape=[10, 1],
                            append_batch_size=False, dtype='float32')
          is_crowd = fluid.layers.data(name='is_crowd', shape=[1],
                            append_batch_size=False, dtype='float32')
          im_info = fluid.layers.data(name='im_infoss', shape=[1, 3],
                            append_batch_size=False, dtype='float32')
          loc_pred, score_pred, loc_target, score_target, bbox_inside_weight, fg_num =
                fluid.layers.retinanet_target_assign(bbox_pred, cls_logits, anchor_box,
                anchor_var, gt_boxes, gt_labels, is_crowd, im_info, 10)

    """

    helper = LayerHelper('retinanet_target_assign', **locals())
    # Assign target label to anchors
    loc_index = helper.create_variable_for_type_inference(dtype='int32')
    score_index = helper.create_variable_for_type_inference(dtype='int32')
    target_label = helper.create_variable_for_type_inference(dtype='int32')
    target_bbox = helper.create_variable_for_type_inference(
        dtype=anchor_box.dtype)
    bbox_inside_weight = helper.create_variable_for_type_inference(
        dtype=anchor_box.dtype)
    fg_num = helper.create_variable_for_type_inference(dtype='int32')
    helper.append_op(
        type="retinanet_target_assign",
        inputs={
            'Anchor': anchor_box,
            'GtBoxes': gt_boxes,
            'GtLabels': gt_labels,
            'IsCrowd': is_crowd,
            'ImInfo': im_info
        },
        outputs={
            'LocationIndex': loc_index,
            'ScoreIndex': score_index,
            'TargetLabel': target_label,
            'TargetBBox': target_bbox,
            'BBoxInsideWeight': bbox_inside_weight,
            'ForegroundNumber': fg_num
        },
        attrs={
            'positive_overlap': positive_overlap,
            'negative_overlap': negative_overlap
        })

    loc_index.stop_gradient = True
    score_index.stop_gradient = True
    target_label.stop_gradient = True
    target_bbox.stop_gradient = True
    bbox_inside_weight.stop_gradient = True
    fg_num.stop_gradient = True

    cls_logits = nn.reshape(x=cls_logits, shape=(-1, num_classes))
    bbox_pred = nn.reshape(x=bbox_pred, shape=(-1, 4))
    predicted_cls_logits = nn.gather(cls_logits, score_index)
    predicted_bbox_pred = nn.gather(bbox_pred, loc_index)

    return predicted_cls_logits, predicted_bbox_pred, target_label, target_bbox, bbox_inside_weight, fg_num


221 222
def rpn_target_assign(bbox_pred,
                      cls_logits,
Y
Yuan Gao 已提交
223
                      anchor_box,
224
                      anchor_var,
225 226 227
                      gt_boxes,
                      is_crowd,
                      im_info,
Y
Yuan Gao 已提交
228
                      rpn_batch_size_per_im=256,
229 230
                      rpn_straddle_thresh=0.0,
                      rpn_fg_fraction=0.5,
Y
Yuan Gao 已提交
231
                      rpn_positive_overlap=0.7,
232 233
                      rpn_negative_overlap=0.3,
                      use_random=True):
Y
Yuan Gao 已提交
234
    """
H
haowang101779990 已提交
235
    **Target Assign Layer for region proposal network (RPN) in Faster-RCNN detection.**
Y
Yuan Gao 已提交
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252

    This layer can be, for given the  Intersection-over-Union (IoU) overlap
    between anchors and ground truth boxes, to assign classification and
    regression targets to each each anchor, these target labels are used for
    train RPN. The classification targets is a binary class label (of being
    an object or not). Following the paper of Faster-RCNN, the positive labels
    are two kinds of anchors: (i) the anchor/anchors with the highest IoU
    overlap with a ground-truth box, or (ii) an anchor that has an IoU overlap
    higher than rpn_positive_overlap(0.7) with any ground-truth box. Note
    that a single ground-truth box may assign positive labels to multiple
    anchors. A non-positive anchor is when its IoU ratio is lower than
    rpn_negative_overlap (0.3) for all ground-truth boxes. Anchors that are
    neither positive nor negative do not contribute to the training objective.
    The regression targets are the encoded ground-truth boxes associated with
    the positive anchors.

    Args:
253
        bbox_pred(Variable): A 3-D Tensor with shape [N, M, 4] represents the
Y
Yuan Gao 已提交
254 255 256
            predicted locations of M bounding bboxes. N is the batch size,
            and each bounding box has four coordinate values and the layout
            is [xmin, ymin, xmax, ymax].
257 258 259
        cls_logits(Variable): A 3-D Tensor with shape [N, M, 1] represents the
            predicted confidence predictions. N is the batch size, 1 is the
            frontground and background sigmoid, M is number of bounding boxes.
Y
Yuan Gao 已提交
260 261 262 263 264 265
        anchor_box(Variable): A 2-D Tensor with shape [M, 4] holds M boxes,
            each box is represented as [xmin, ymin, xmax, ymax],
            [xmin, ymin] is the left top coordinate of the anchor box,
            if the input is image feature map, they are close to the origin
            of the coordinate system. [xmax, ymax] is the right bottom
            coordinate of the anchor box.
266 267
        anchor_var(Variable): A 2-D Tensor with shape [M,4] holds expanded 
            variances of anchors.
翟飞跃 已提交
268
        gt_boxes (Variable): The ground-truth bounding boxes (bboxes) are a 2D
Y
Yuan Gao 已提交
269 270
            LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth
            bboxes of mini-batch input.
271 272 273
        is_crowd (Variable): A 1-D LoDTensor which indicates groud-truth is crowd.
        im_info (Variable): A 2-D LoDTensor with shape [N, 3]. N is the batch size,
        3 is the height, width and scale.
Y
Yuan Gao 已提交
274
        rpn_batch_size_per_im(int): Total number of RPN examples per image.
275 276 277
        rpn_straddle_thresh(float): Remove RPN anchors that go outside the image
            by straddle_thresh pixels.
        rpn_fg_fraction(float): Target fraction of RoI minibatch that is labeled
Y
Yuan Gao 已提交
278 279 280 281 282 283 284 285 286
            foreground (i.e. class > 0), 0-th class is background.
        rpn_positive_overlap(float): Minimum overlap required between an anchor
            and ground-truth box for the (anchor, gt box) pair to be a positive
            example.
        rpn_negative_overlap(float): Maximum overlap allowed between an anchor
            and ground-truth box for the (anchor, gt box) pair to be a negative
            examples.

    Returns:
M
minqiyang 已提交
287
        tuple:
Y
Yuan Gao 已提交
288
               A tuple(predicted_scores, predicted_location, target_label,
J
jerrywgz 已提交
289 290
               target_bbox, bbox_inside_weight) is returned. The predicted_scores 
               and predicted_location is the predicted result of the RPN.
Y
Yuan Gao 已提交
291 292 293 294 295 296 297
               The target_label and target_bbox is the ground truth,
               respectively. The predicted_location is a 2D Tensor with shape
               [F, 4], and the shape of target_bbox is same as the shape of
               the predicted_location, F is the number of the foreground
               anchors. The predicted_scores is a 2D Tensor with shape
               [F + B, 1], and the shape of target_label is same as the shape
               of the predicted_scores, B is the number of the background
M
minqiyang 已提交
298
               anchors, the F and B is depends on the input of this operator.
J
jerrywgz 已提交
299 300
               Bbox_inside_weight represents whether the predicted loc is fake_fg
               or not and the shape is [F, 4].
Y
Yuan Gao 已提交
301 302 303 304

    Examples:
        .. code-block:: python

B
Bai Yifan 已提交
305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322
            import paddle.fluid as fluid
            bbox_pred = fluid.layers.data(name='bbox_pred', shape=[100, 4],
                            append_batch_size=False, dtype='float32')
            cls_logits = fluid.layers.data(name='cls_logits', shape=[100, 1],
                            append_batch_size=False, dtype='float32')
            anchor_box = fluid.layers.data(name='anchor_box', shape=[20, 4],
                            append_batch_size=False, dtype='float32')
            anchor_var = fluid.layers.data(name='anchor_var', shape=[20, 4],
                            append_batch_size=False, dtype='float32')
            gt_boxes = fluid.layers.data(name='gt_boxes', shape=[10, 4],
                            append_batch_size=False, dtype='float32')
            is_crowd = fluid.layers.data(name='is_crowd', shape=[1],
                            append_batch_size=False, dtype='float32')
            im_info = fluid.layers.data(name='im_infoss', shape=[1, 3],
                            append_batch_size=False, dtype='float32')
            loc_pred, score_pred, loc_target, score_target, bbox_inside_weight=
                fluid.layers.rpn_target_assign(bbox_pred, cls_logits,
                anchor_box, anchor_var, gt_boxes, is_crowd, im_info)
H
haowang101779990 已提交
323

Y
Yuan Gao 已提交
324 325 326
    """

    helper = LayerHelper('rpn_target_assign', **locals())
327
    # Assign target label to anchors
J
jerrywgz 已提交
328 329 330 331 332 333 334
    loc_index = helper.create_variable_for_type_inference(dtype='int32')
    score_index = helper.create_variable_for_type_inference(dtype='int32')
    target_label = helper.create_variable_for_type_inference(dtype='int32')
    target_bbox = helper.create_variable_for_type_inference(
        dtype=anchor_box.dtype)
    bbox_inside_weight = helper.create_variable_for_type_inference(
        dtype=anchor_box.dtype)
Y
Yuan Gao 已提交
335 336
    helper.append_op(
        type="rpn_target_assign",
337 338 339 340 341 342
        inputs={
            'Anchor': anchor_box,
            'GtBoxes': gt_boxes,
            'IsCrowd': is_crowd,
            'ImInfo': im_info
        },
Y
Yuan Gao 已提交
343 344 345
        outputs={
            'LocationIndex': loc_index,
            'ScoreIndex': score_index,
346
            'TargetLabel': target_label,
J
jerrywgz 已提交
347
            'TargetBBox': target_bbox,
J
jerrywgz 已提交
348
            'BBoxInsideWeight': bbox_inside_weight
Y
Yuan Gao 已提交
349 350 351
        },
        attrs={
            'rpn_batch_size_per_im': rpn_batch_size_per_im,
352
            'rpn_straddle_thresh': rpn_straddle_thresh,
Y
Yuan Gao 已提交
353 354
            'rpn_positive_overlap': rpn_positive_overlap,
            'rpn_negative_overlap': rpn_negative_overlap,
355 356
            'rpn_fg_fraction': rpn_fg_fraction,
            'use_random': use_random
Y
Yuan Gao 已提交
357 358
        })

359 360 361 362
    loc_index.stop_gradient = True
    score_index.stop_gradient = True
    target_label.stop_gradient = True
    target_bbox.stop_gradient = True
J
jerrywgz 已提交
363
    bbox_inside_weight.stop_gradient = True
Y
Yuan Gao 已提交
364

365 366 367 368
    cls_logits = nn.reshape(x=cls_logits, shape=(-1, 1))
    bbox_pred = nn.reshape(x=bbox_pred, shape=(-1, 4))
    predicted_cls_logits = nn.gather(cls_logits, score_index)
    predicted_bbox_pred = nn.gather(bbox_pred, loc_index)
369

J
jerrywgz 已提交
370
    return predicted_cls_logits, predicted_bbox_pred, target_label, target_bbox, bbox_inside_weight
Y
Yuan Gao 已提交
371 372


373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
def sigmoid_focal_loss(x, label, fg_num, gamma=2, alpha=0.25):
    """
    **Sigmoid Focal Loss Operator.**

    Focal loss is used to address the foreground-background class imbalance existed
    on the training phase of one-stage detectors. This operator computes the sigmoid
    value for each element in the input tensor, after which focal loss is measured.
    
    The focal loss is given as followed:

    .. math::
        loss_j = (-label_j * alpha * {(1 - \\sigma(x_j))}^{gamma} * \\log(\\sigma(x_j)) -
        (1 - labels_j) * (1 - alpha) * {(\sigma(x_j)}^{ gamma} * \\log(1 - \\sigma(x_j)))
        / fg\_num, j = 1,...,K

    We know that
    
    .. math::
        \\sigma(x_j) = \\frac{1}{1 + \\exp(-x_j)}

    Args:
        x(Variable): A 2-D tensor with shape [N, D], where N is the batch size and D is the number
            of classes (excluding background). This input is a tensor of logits computed by the
            previous operator.
        label(Variable): A 2-D tensor with shape [N, 1], which is the probabilistic labels.
        fg_num(Variable): A 1-D tensor with shape [1], which is the number of foreground.

        gamma(float): Hyper-parameter to balance the easy and hard examples. Default value is
            set to 2.0.
        alpha(float): Hyper-parameter to balance the positive and negative example. Default value
            is set to 0.25.

    Returns:
        out(Variable): A 2-D tensor with shape [N, D], which is the focal loss.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            input = fluid.layers.data(
                name='data', shape=[10,80], append_batch_size=False, dtype='float32')
            label = fluid.layers.data(
                name='label', shape=[10,1], append_batch_size=False, dtype='int32')
            fg_num = fluid.layers.data(
                name='fg_num', shape=[1], append_batch_size=False, dtype='int32')
            loss = fluid.layers.sigmoid_focal_loss(x=input,
                                                   label=label,
                                                   fg_num=fg_num,
                                                   gamma=2.,
                                                   alpha=0.25)
    """

    helper = LayerHelper("sigmoid_focal_loss", **locals())

    out = helper.create_variable_for_type_inference(dtype=x.dtype)

    helper.append_op(
        type="sigmoid_focal_loss",
        inputs={"X": x,
                "Label": label,
                "FgNum": fg_num},
        attrs={"gamma": gamma,
               'alpha': alpha},
        outputs={"Out": out})
    return out


Y
Yuan Gao 已提交
441 442
def detection_output(loc,
                     scores,
443 444 445 446 447 448 449 450 451
                     prior_box,
                     prior_box_var,
                     background_label=0,
                     nms_threshold=0.3,
                     nms_top_k=400,
                     keep_top_k=200,
                     score_threshold=0.01,
                     nms_eta=1.0):
    """
452
    **Detection Output Layer for Single Shot Multibox Detector (SSD).**
453

454 455
    This operation is to get the detection results by performing following
    two steps:
C
caoying03 已提交
456

457 458 459 460 461 462
    1. Decode input bounding box predictions according to the prior boxes.
    2. Get the final detection results by applying multi-class non maximum
       suppression (NMS).

    Please note, this operation doesn't clip the final output bounding boxes
    to the image window.
463 464 465 466 467 468

    Args:
        loc(Variable): A 3-D Tensor with shape [N, M, 4] represents the
            predicted locations of M bounding bboxes. N is the batch size,
            and each bounding box has four coordinate values and the layout
            is [xmin, ymin, xmax, ymax].
Y
Yuan Gao 已提交
469 470 471 472
        scores(Variable): A 3-D Tensor with shape [N, M, C] represents the
            predicted confidence predictions. N is the batch size, C is the
            class number, M is number of bounding boxes. For each category
            there are total M scores which corresponding M bounding boxes.
473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
        prior_box(Variable): A 2-D Tensor with shape [M, 4] holds M boxes,
            each box is represented as [xmin, ymin, xmax, ymax],
            [xmin, ymin] is the left top coordinate of the anchor box,
            if the input is image feature map, they are close to the origin
            of the coordinate system. [xmax, ymax] is the right bottom
            coordinate of the anchor box.
        prior_box_var(Variable): A 2-D Tensor with shape [M, 4] holds M group
            of variance.
        background_label(float): The index of background label,
            the background label will be ignored. If set to -1, then all
            categories will be considered.
        nms_threshold(float): The threshold to be used in NMS.
        nms_top_k(int): Maximum number of detections to be kept according
            to the confidences aftern the filtering detections based on
            score_threshold.
        keep_top_k(int): Number of total bboxes to be kept per image after
            NMS step. -1 means keeping all bboxes after NMS step.
        score_threshold(float): Threshold to filter out bounding boxes with
            low confidence score. If not provided, consider all boxes.
        nms_eta(float): The parameter for adaptive NMS.

    Returns:
M
minqiyang 已提交
495 496
        Variable:

497
            The detection outputs is a LoDTensor with shape [No, 6].
498 499 500 501 502 503
            Each row has six values: [label, confidence, xmin, ymin, xmax, ymax].
            `No` is the total number of detections in this mini-batch. For each
            instance, the offsets in first dimension are called LoD, the offset
            number is N + 1, N is the batch size. The i-th image has
            `LoD[i + 1] - LoD[i]` detected results, if it is 0, the i-th image
            has no detected results. If all images have not detected results,
J
jerrywgz 已提交
504
            LoD will be set to {1}, and output tensor only contains one
505
            value, which is -1.
J
jerrywgz 已提交
506 507
            (After version 1.3, when no boxes detected, the lod is changed
             from {0} to {1}.)
508 509 510 511

    Examples:
        .. code-block:: python

512 513 514
            import paddle.fluid as fluid

            pb = fluid.layers.data(name='prior_box', shape=[10, 4],
515
                         append_batch_size=False, dtype='float32')
516
            pbv = fluid.layers.data(name='prior_box_var', shape=[10, 4],
517
                          append_batch_size=False, dtype='float32')
518
            loc = fluid.layers.data(name='target_box', shape=[2, 21, 4],
519
                          append_batch_size=False, dtype='float32')
520
            scores = fluid.layers.data(name='scores', shape=[2, 21, 10],
521
                          append_batch_size=False, dtype='float32')
522
            nmsed_outs = fluid.layers.detection_output(scores=scores,
523 524 525 526 527
                                       loc=loc,
                                       prior_box=pb,
                                       prior_box_var=pbv)
    """
    helper = LayerHelper("detection_output", **locals())
528 529 530 531 532
    decoded_box = box_coder(
        prior_box=prior_box,
        prior_box_var=prior_box_var,
        target_box=loc,
        code_type='decode_center_size')
533
    scores = nn.softmax(input=scores)
Y
Yuan Gao 已提交
534
    scores = nn.transpose(scores, perm=[0, 2, 1])
535
    scores.stop_gradient = True
X
Xin Pan 已提交
536 537
    nmsed_outs = helper.create_variable_for_type_inference(
        dtype=decoded_box.dtype)
538 539 540 541 542 543 544 545 546 547 548 549 550
    helper.append_op(
        type="multiclass_nms",
        inputs={'Scores': scores,
                'BBoxes': decoded_box},
        outputs={'Out': nmsed_outs},
        attrs={
            'background_label': 0,
            'nms_threshold': nms_threshold,
            'nms_top_k': nms_top_k,
            'keep_top_k': keep_top_k,
            'score_threshold': score_threshold,
            'nms_eta': 1.0
        })
551
    nmsed_outs.stop_gradient = True
552
    return nmsed_outs
C
chengduoZH 已提交
553 554


X
Xin Pan 已提交
555 556 557 558 559 560 561 562 563 564 565
@templatedoc()
def iou_similarity(x, y, name=None):
    """
    ${comment}

    Args:
        x(${x_type}): ${x_comment}
        y(${y_type}): ${y_comment}

    Returns:
        out(${out_type}): ${out_comment}
566 567 568 569 570 571 572 573 574

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            x = fluid.layers.data(name='x', shape=[4], dtype='float32')
            y = fluid.layers.data(name='y', shape=[4], dtype='float32')
            iou = fluid.layers.iou_similarity(x=x, y=y)
X
Xin Pan 已提交
575 576 577
    """
    helper = LayerHelper("iou_similarity", **locals())
    if name is None:
X
Xin Pan 已提交
578
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597
    else:
        out = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

    helper.append_op(
        type="iou_similarity",
        inputs={"X": x,
                "Y": y},
        attrs={},
        outputs={"Out": out})
    return out


@templatedoc()
def box_coder(prior_box,
              prior_box_var,
              target_box,
              code_type="encode_center_size",
              box_normalized=True,
598 599
              name=None,
              axis=0):
X
Xin Pan 已提交
600
    """
601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
    **Box Coder Layer**

    Encode/Decode the target bounding box with the priorbox information.
    
    The Encoding schema described below:

    .. math::

        ox = (tx - px) / pw / pxv

        oy = (ty - py) / ph / pyv

        ow = \log(\abs(tw / pw)) / pwv 

        oh = \log(\abs(th / ph)) / phv 

    The Decoding schema described below:
    
    .. math::
  
        ox = (pw * pxv * tx * + px) - tw / 2

        oy = (ph * pyv * ty * + py) - th / 2

        ow = \exp(pwv * tw) * pw + tw / 2

        oh = \exp(phv * th) * ph + th / 2   

    where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, 
    width and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote 
    the priorbox's (anchor) center coordinates, width and height. `pxv`, 
    `pyv`, `pwv`, `phv` denote the variance of the priorbox and `ox`, `oy`, 
    `ow`, `oh` denote the encoded/decoded coordinates, width and height. 

    During Box Decoding, two modes for broadcast are supported. Say target 
    box has shape [N, M, 4], and the shape of prior box can be [N, 4] or 
    [M, 4]. Then prior box will broadcast to target box along the 
    assigned axis. 
X
Xin Pan 已提交
639 640

    Args:
641 642 643 644 645 646 647
        prior_box(Variable): Box list prior_box is a 2-D Tensor with shape 
                             [M, 4] holds M boxes, each box is represented as
                             [xmin, ymin, xmax, ymax], [xmin, ymin] is the 
                             left top coordinate of the anchor box, if the 
                             input is image feature map, they are close to 
                             the origin of the coordinate system. [xmax, ymax]
                             is the right bottom coordinate of the anchor box.       
648 649 650 651
        prior_box_var(Variable|list|None): prior_box_var supports two types 
                              of input. One is variable with shape [M, 4] 
                              holds M group. The other one is list consist of 
                              4 elements shared by all boxes. 
652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669
        target_box(Variable): This input can be a 2-D LoDTensor with shape 
                              [N, 4] when code_type is 'encode_center_size'. 
                              This input also can be a 3-D Tensor with shape 
                              [N, M, 4] when code_type is 'decode_center_size'. 
                              Each box is represented as  
                              [xmin, ymin, xmax, ymax]. This tensor can 
                              contain LoD information to represent a batch 
                              of inputs. 
        code_type(string): The code type used with the target box. It can be
                           encode_center_size or decode_center_size
        box_normalized(int): Whether treat the priorbox as a noramlized box.
                             Set true by default.
        name(string): The name of box coder.
        axis(int): Which axis in PriorBox to broadcast for box decode, 
                   for example, if axis is 0 and TargetBox has shape
                   [N, M, 4] and PriorBox has shape [M, 4], then PriorBox
                   will broadcast to [N, M, 4] for decoding. It is only valid
                   when code type is decode_center_size. Set 0 by default. 
X
Xin Pan 已提交
670 671

    Returns:
672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698
        output_box(Variable): When code_type is 'encode_center_size', the 
                              output tensor of box_coder_op with shape 
                              [N, M, 4] representing the result of N target 
                              boxes encoded with M Prior boxes and variances. 
                              When code_type is 'decode_center_size', 
                              N represents the batch size and M represents 
                              the number of deocded boxes.

    Examples:
 
        .. code-block:: python
 
            prior_box = fluid.layers.data(name='prior_box', 
                                          shape=[512, 4], 
                                          dtype='float32',
                                          append_batch_size=False)
            target_box = fluid.layers.data(name='target_box',
                                           shape=[512,81,4],
                                           dtype='float32',
                                           append_batch_size=False)
            output = fluid.layers.box_coder(prior_box=prior_box,
                                            prior_box_var=[0.1,0.1,0.2,0.2],
                                            target_box=target_box,
                                            code_type="decode_center_size",
                                            box_normalized=False,
                                            axis=1)

X
Xin Pan 已提交
699 700 701 702
    """
    helper = LayerHelper("box_coder", **locals())

    if name is None:
X
Xin Pan 已提交
703 704
        output_box = helper.create_variable_for_type_inference(
            dtype=prior_box.dtype)
X
Xin Pan 已提交
705 706 707 708
    else:
        output_box = helper.create_variable(
            name=name, dtype=prior_box.dtype, persistable=False)

709 710 711 712 713 714 715 716 717 718 719 720
    inputs = {"PriorBox": prior_box, "TargetBox": target_box}
    attrs = {
        "code_type": code_type,
        "box_normalized": box_normalized,
        "axis": axis
    }
    if isinstance(prior_box_var, Variable):
        inputs['PriorBoxVar'] = prior_box_var
    elif isinstance(prior_box_var, list):
        attrs['variance'] = prior_box_var
    else:
        raise TypeError("Input variance of box_coder must be Variable or lisz")
X
Xin Pan 已提交
721 722
    helper.append_op(
        type="box_coder",
723 724
        inputs=inputs,
        attrs=attrs,
X
Xin Pan 已提交
725 726 727 728 729 730 731 732 733 734 735 736 737 738
        outputs={"OutputBox": output_box})
    return output_box


@templatedoc()
def polygon_box_transform(input, name=None):
    """
    ${comment}

    Args:
        input(${input_type}): ${input_comment}

    Returns:
        output(${output_type}): ${output_comment}
B
Bai Yifan 已提交
739 740 741 742 743 744 745 746

    Examples:
        .. code-block:: python
            
            import paddle.fluid as fluid
            input = fluid.layers.data(name='input', shape=[4, 10, 5, 5],
                                      append_batch_size=False, dtype='float32')
            out = fluid.layers.polygon_box_transform(input)
X
Xin Pan 已提交
747 748 749
    """
    helper = LayerHelper("polygon_box_transform", **locals())
    if name is None:
X
Xin Pan 已提交
750
        output = helper.create_variable_for_type_inference(dtype=input.dtype)
X
Xin Pan 已提交
751 752 753 754 755 756 757 758 759 760 761 762
    else:
        output = helper.create_variable(
            name=name, dtype=prior_box.input, persistable=False)

    helper.append_op(
        type="polygon_box_transform",
        inputs={"Input": input},
        attrs={},
        outputs={"Output": output})
    return output


D
dengkaipeng 已提交
763 764
@templatedoc(op_type="yolov3_loss")
def yolov3_loss(x,
765 766
                gt_box,
                gt_label,
D
dengkaipeng 已提交
767
                anchors,
768
                anchor_mask,
D
dengkaipeng 已提交
769 770
                class_num,
                ignore_thresh,
771
                downsample_ratio,
772
                gt_score=None,
D
dengkaipeng 已提交
773
                use_label_smooth=True,
D
dengkaipeng 已提交
774 775 776 777 778 779
                name=None):
    """
    ${comment}

    Args:
        x (Variable): ${x_comment}
780
        gt_box (Variable): groud truth boxes, should be in shape of [N, B, 4],
781 782 783 784
                          in the third dimenstion, x, y, w, h should be stored. 
                          x,y is the center cordinate of boxes, w, h are the
                          width and height, x, y, w, h should be divided by 
                          input image height to scale to [0, 1].
D
dengkaipeng 已提交
785 786
                          N is the batch number and B is the max box number in 
                          an image.
787
        gt_label (Variable): class id of ground truth boxes, shoud be in shape
D
dengkaipeng 已提交
788
                            of [N, B].
D
dengkaipeng 已提交
789
        anchors (list|tuple): ${anchors_comment}
790
        anchor_mask (list|tuple): ${anchor_mask_comment}
D
dengkaipeng 已提交
791 792
        class_num (int): ${class_num_comment}
        ignore_thresh (float): ${ignore_thresh_comment}
793
        downsample_ratio (int): ${downsample_ratio_comment}
794
        name (string): the name of yolov3 loss. Default None.
795
        gt_score (Variable): mixup score of ground truth boxes, shoud be in shape
796
                            of [N, B]. Default None.
797
        use_label_smooth (bool): ${use_label_smooth_comment}
D
dengkaipeng 已提交
798 799

    Returns:
800
        Variable: A 1-D tensor with shape [N], the value of yolov3 loss
D
dengkaipeng 已提交
801 802 803

    Raises:
        TypeError: Input x of yolov3_loss must be Variable
D
dengkaipeng 已提交
804 805
        TypeError: Input gtbox of yolov3_loss must be Variable
        TypeError: Input gtlabel of yolov3_loss must be Variable
D
dengkaipeng 已提交
806
        TypeError: Input gtscore of yolov3_loss must be None or Variable
D
dengkaipeng 已提交
807 808 809
        TypeError: Attr anchors of yolov3_loss must be list or tuple
        TypeError: Attr class_num of yolov3_loss must be an integer
        TypeError: Attr ignore_thresh of yolov3_loss must be a float number
810
        TypeError: Attr use_label_smooth of yolov3_loss must be a bool value
D
dengkaipeng 已提交
811 812

    Examples:
813 814 815
      .. code-block:: python

          x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32')
816 817 818
          gt_box = fluid.layers.data(name='gt_box', shape=[6, 4], dtype='float32')
          gt_label = fluid.layers.data(name='gt_label', shape=[6], dtype='int32')
          gt_score = fluid.layers.data(name='gt_score', shape=[6], dtype='float32')
819 820
          anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]
          anchor_mask = [0, 1, 2]
821 822
          loss = fluid.layers.yolov3_loss(x=x, gt_box=gt_box, gt_label=gt_label,
                                          gt_score=gt_score, anchors=anchors, 
823 824
                                          anchor_mask=anchor_mask, class_num=80,
                                          ignore_thresh=0.7, downsample_ratio=32)
D
dengkaipeng 已提交
825 826 827 828 829
    """
    helper = LayerHelper('yolov3_loss', **locals())

    if not isinstance(x, Variable):
        raise TypeError("Input x of yolov3_loss must be Variable")
830
    if not isinstance(gt_box, Variable):
D
dengkaipeng 已提交
831
        raise TypeError("Input gtbox of yolov3_loss must be Variable")
832
    if not isinstance(gt_label, Variable):
D
dengkaipeng 已提交
833
        raise TypeError("Input gtlabel of yolov3_loss must be Variable")
834
    if gt_score is not None and not isinstance(gt_score, Variable):
835
        raise TypeError("Input gtscore of yolov3_loss must be Variable")
D
dengkaipeng 已提交
836 837
    if not isinstance(anchors, list) and not isinstance(anchors, tuple):
        raise TypeError("Attr anchors of yolov3_loss must be list or tuple")
838 839
    if not isinstance(anchor_mask, list) and not isinstance(anchor_mask, tuple):
        raise TypeError("Attr anchor_mask of yolov3_loss must be list or tuple")
D
dengkaipeng 已提交
840 841 842 843 844
    if not isinstance(class_num, int):
        raise TypeError("Attr class_num of yolov3_loss must be an integer")
    if not isinstance(ignore_thresh, float):
        raise TypeError(
            "Attr ignore_thresh of yolov3_loss must be a float number")
845 846 847
    if not isinstance(use_label_smooth, bool):
        raise TypeError(
            "Attr use_label_smooth of yolov3_loss must be a bool value")
D
dengkaipeng 已提交
848 849 850 851 852 853 854

    if name is None:
        loss = helper.create_variable_for_type_inference(dtype=x.dtype)
    else:
        loss = helper.create_variable(
            name=name, dtype=x.dtype, persistable=False)

855 856 857
    objectness_mask = helper.create_variable_for_type_inference(dtype='int32')
    gt_match_mask = helper.create_variable_for_type_inference(dtype='int32')

858 859
    inputs = {
        "X": x,
860 861
        "GTBox": gt_box,
        "GTLabel": gt_label,
862
    }
863
    if gt_score:
864
        inputs["GTScore"] = gt_score
865

D
dengkaipeng 已提交
866 867
    attrs = {
        "anchors": anchors,
868
        "anchor_mask": anchor_mask,
D
dengkaipeng 已提交
869 870
        "class_num": class_num,
        "ignore_thresh": ignore_thresh,
871
        "downsample_ratio": downsample_ratio,
872
        "use_label_smooth": use_label_smooth,
D
dengkaipeng 已提交
873 874 875 876
    }

    helper.append_op(
        type='yolov3_loss',
877
        inputs=inputs,
878 879 880 881 882
        outputs={
            'Loss': loss,
            'ObjectnessMask': objectness_mask,
            'GTMatchMask': gt_match_mask
        },
D
dengkaipeng 已提交
883 884 885 886
        attrs=attrs)
    return loss


D
dengkaipeng 已提交
887
@templatedoc(op_type="yolo_box")
888 889 890 891 892 893 894
def yolo_box(x,
             img_size,
             anchors,
             class_num,
             conf_thresh,
             downsample_ratio,
             name=None):
D
dengkaipeng 已提交
895 896 897 898 899
    """
    ${comment}

    Args:
        x (Variable): ${x_comment}
900
        img_size (Variable): ${img_size_comment}
D
dengkaipeng 已提交
901 902 903 904
        anchors (list|tuple): ${anchors_comment}
        class_num (int): ${class_num_comment}
        conf_thresh (float): ${conf_thresh_comment}
        downsample_ratio (int): ${downsample_ratio_comment}
905
        name (string): the name of yolo box layer. Default None.
D
dengkaipeng 已提交
906 907

    Returns:
D
dengkaipeng 已提交
908
        Variable: A 3-D tensor with shape [N, M, 4], the coordinates of boxes,
D
dengkaipeng 已提交
909 910
        and a 3-D tensor with shape [N, M, :attr:`class_num`], the classification 
        scores of boxes.
D
dengkaipeng 已提交
911 912 913 914 915 916 917 918

    Raises:
        TypeError: Input x of yolov_box must be Variable
        TypeError: Attr anchors of yolo box must be list or tuple
        TypeError: Attr class_num of yolo box must be an integer
        TypeError: Attr conf_thresh of yolo box must be a float number

    Examples:
D
dengkaipeng 已提交
919

D
dengkaipeng 已提交
920 921
    .. code-block:: python

X
xiaoting 已提交
922
        import paddle.fluid as fluid
D
dengkaipeng 已提交
923
        x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32')
924
        img_size = fluid.layers.data(name='img_size',shape=[2],dtype='int64')
D
dengkaipeng 已提交
925
        anchors = [10, 13, 16, 30, 33, 23]
926
        loss = fluid.layers.yolo_box(x=x, img_size=img_size, class_num=80, anchors=anchors, 
D
dengkaipeng 已提交
927 928 929 930 931
                                        conf_thresh=0.01, downsample_ratio=32)
    """
    helper = LayerHelper('yolo_box', **locals())

    if not isinstance(x, Variable):
932 933 934
        raise TypeError("Input x of yolo_box must be Variable")
    if not isinstance(img_size, Variable):
        raise TypeError("Input img_size of yolo_box must be Variable")
D
dengkaipeng 已提交
935
    if not isinstance(anchors, list) and not isinstance(anchors, tuple):
936
        raise TypeError("Attr anchors of yolo_box must be list or tuple")
D
dengkaipeng 已提交
937
    if not isinstance(class_num, int):
938
        raise TypeError("Attr class_num of yolo_box must be an integer")
D
dengkaipeng 已提交
939
    if not isinstance(conf_thresh, float):
940
        raise TypeError("Attr ignore_thresh of yolo_box must be a float number")
D
dengkaipeng 已提交
941 942 943 944 945 946 947

    boxes = helper.create_variable_for_type_inference(dtype=x.dtype)
    scores = helper.create_variable_for_type_inference(dtype=x.dtype)

    attrs = {
        "anchors": anchors,
        "class_num": class_num,
D
dengkaipeng 已提交
948
        "conf_thresh": conf_thresh,
D
dengkaipeng 已提交
949 950 951 952 953
        "downsample_ratio": downsample_ratio,
    }

    helper.append_op(
        type='yolo_box',
954 955 956 957
        inputs={
            "X": x,
            "ImgSize": img_size,
        },
D
dengkaipeng 已提交
958 959 960 961 962 963 964 965
        outputs={
            'Boxes': boxes,
            'Scores': scores,
        },
        attrs=attrs)
    return boxes, scores


X
Xin Pan 已提交
966
@templatedoc()
967 968
def detection_map(detect_res,
                  label,
969 970
                  class_num,
                  background_label=0,
971 972
                  overlap_threshold=0.3,
                  evaluate_difficult=True,
973 974 975 976
                  has_state=None,
                  input_states=None,
                  out_states=None,
                  ap_version='integral'):
X
Xin Pan 已提交
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004
    """
    ${comment}

    Args:
        detect_res: ${detect_res_comment}
        label:  ${label_comment}
        class_num: ${class_num_comment}
        background_label: ${background_label_comment}
        overlap_threshold: ${overlap_threshold_comment}
        evaluate_difficult: ${evaluate_difficult_comment}
        has_state: ${has_state_comment}
        input_states: If not None, It contains 3 elements:
            1. pos_count ${pos_count_comment}.
            2. true_pos ${true_pos_comment}.
            3. false_pos ${false_pos_comment}.
        out_states: If not None, it contains 3 elements.
            1. accum_pos_count ${accum_pos_count_comment}.
            2. accum_true_pos ${accum_true_pos_comment}.
            3. accum_false_pos ${accum_false_pos_comment}.
        ap_version: ${ap_type_comment}

    Returns:
        ${map_comment}


    Examples:
          .. code-block:: python

1005
            from fluid.layers import detection
X
Xin Pan 已提交
1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
            detect_res = fluid.layers.data(
                name='detect_res',
                shape=[10, 6],
                append_batch_size=False,
                dtype='float32')
            label = fluid.layers.data(
                name='label',
                shape=[10, 6],
                append_batch_size=False,
                dtype='float32')

1017
            map_out = detection.detection_map(detect_res, label, 21)
X
Xin Pan 已提交
1018
    """
1019 1020
    helper = LayerHelper("detection_map", **locals())

1021
    def __create_var(type):
X
Xin Pan 已提交
1022
        return helper.create_variable_for_type_inference(dtype=type)
1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034

    map_out = __create_var('float32')
    accum_pos_count_out = out_states[0] if out_states else __create_var('int32')
    accum_true_pos_out = out_states[1] if out_states else __create_var(
        'float32')
    accum_false_pos_out = out_states[2] if out_states else __create_var(
        'float32')

    pos_count = input_states[0] if input_states else None
    true_pos = input_states[1] if input_states else None
    false_pos = input_states[2] if input_states else None

1035 1036 1037 1038 1039
    helper.append_op(
        type="detection_map",
        inputs={
            'Label': label,
            'DetectRes': detect_res,
1040
            'HasState': has_state,
1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053
            'PosCount': pos_count,
            'TruePos': true_pos,
            'FalsePos': false_pos
        },
        outputs={
            'MAP': map_out,
            'AccumPosCount': accum_pos_count_out,
            'AccumTruePos': accum_true_pos_out,
            'AccumFalsePos': accum_false_pos_out
        },
        attrs={
            'overlap_threshold': overlap_threshold,
            'evaluate_difficult': evaluate_difficult,
1054 1055
            'ap_type': ap_version,
            'class_num': class_num,
1056
        })
1057
    return map_out
1058 1059


1060 1061 1062 1063
def bipartite_match(dist_matrix,
                    match_type=None,
                    dist_threshold=None,
                    name=None):
1064
    """
Y
yuyang18 已提交
1065 1066
    This operator implements a greedy bipartite matching algorithm, which is
    used to obtain the matching with the maximum distance based on the input
1067
    distance matrix. For input 2D matrix, the bipartite matching algorithm can
Y
yuyang18 已提交
1068 1069 1070 1071 1072 1073 1074 1075
    find the matched column for each row (matched means the largest distance),
    also can find the matched row for each column. And this operator only
    calculate matched indices from column to row. For each instance,
    the number of matched indices is the column number of the input distance
    matrix.

    There are two outputs, matched indices and distance.
    A simple description, this algorithm matched the best (maximum distance)
1076 1077 1078
    row entity to the column entity and the matched indices are not duplicated
    in each row of ColToRowMatchIndices. If the column entity is not matched
    any row entity, set -1 in ColToRowMatchIndices.
C
chengduoZH 已提交
1079

Y
yuyang18 已提交
1080
    NOTE: the input DistMat can be LoDTensor (with LoD) or Tensor.
1081 1082 1083
    If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size.
    If Tensor, the height of ColToRowMatchIndices is 1.

Y
yuyang18 已提交
1084 1085 1086
    NOTE: This API is a very low level API. It is used by :code:`ssd_loss`
    layer. Please consider to use :code:`ssd_loss` instead.

1087 1088 1089 1090 1091
    Args:
        dist_matrix(Variable): This input is a 2-D LoDTensor with shape
            [K, M]. It is pair-wise distance matrix between the entities
            represented by each row and each column. For example, assumed one
            entity is A with shape [K], another entity is B with shape [M]. The
Y
yuyang18 已提交
1092 1093 1094 1095 1096 1097
            dist_matrix[i][j] is the distance between A[i] and B[j]. The bigger
            the distance is, the better matching the pairs are.

            NOTE: This tensor can contain LoD information to represent a batch
            of inputs. One instance of this batch can contain different numbers
            of entities.
1098
        match_type(string|None): The type of matching method, should be
Y
yuyang18 已提交
1099
           'bipartite' or 'per_prediction'. [default 'bipartite'].
1100 1101
        dist_threshold(float|None): If `match_type` is 'per_prediction',
            this threshold is to determine the extra matching bboxes based
Y
yuyang18 已提交
1102
            on the maximum distance, 0.5 by default.
1103
    Returns:
Y
yuyang18 已提交
1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126
        tuple: a tuple with two elements is returned. The first is
        matched_indices, the second is matched_distance.

        The matched_indices is a 2-D Tensor with shape [N, M] in int type.
        N is the batch size. If match_indices[i][j] is -1, it
        means B[j] does not match any entity in i-th instance.
        Otherwise, it means B[j] is matched to row
        match_indices[i][j] in i-th instance. The row number of
        i-th instance is saved in match_indices[i][j].

        The matched_distance is a 2-D Tensor with shape [N, M] in float type
        . N is batch size. If match_indices[i][j] is -1,
        match_distance[i][j] is also -1.0. Otherwise, assumed
        match_distance[i][j] = d, and the row offsets of each instance
        are called LoD. Then match_distance[i][j] =
        dist_matrix[d+LoD[i]][j].

    Examples:

        >>> x = fluid.layers.data(name='x', shape=[4], dtype='float32')
        >>> y = fluid.layers.data(name='y', shape=[4], dtype='float32')
        >>> iou = fluid.layers.iou_similarity(x=x, y=y)
        >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou)
1127 1128
    """
    helper = LayerHelper('bipartite_match', **locals())
X
Xin Pan 已提交
1129 1130 1131
    match_indices = helper.create_variable_for_type_inference(dtype='int32')
    match_distance = helper.create_variable_for_type_inference(
        dtype=dist_matrix.dtype)
1132 1133 1134
    helper.append_op(
        type='bipartite_match',
        inputs={'DistMat': dist_matrix},
1135 1136 1137 1138
        attrs={
            'match_type': match_type,
            'dist_threshold': dist_threshold,
        },
1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155
        outputs={
            'ColToRowMatchIndices': match_indices,
            'ColToRowMatchDist': match_distance
        })
    return match_indices, match_distance


def target_assign(input,
                  matched_indices,
                  negative_indices=None,
                  mismatch_value=None,
                  name=None):
    """
    This operator can be, for given the target bounding boxes or labels,
    to assign classification and regression targets to each prediction as well as
    weights to prediction. The weights is used to specify which prediction would
    not contribute to training loss.
C
chengduoZH 已提交
1156

1157 1158 1159 1160 1161
    For each instance, the output `out` and`out_weight` are assigned based on
    `match_indices` and `negative_indices`.
    Assumed that the row offset for each instance in `input` is called lod,
    this operator assigns classification/regression targets by performing the
    following steps:
C
chengduoZH 已提交
1162

1163
    1. Assigning all outputs based on `match_indices`:
C
chengduoZH 已提交
1164

1165 1166 1167
    .. code-block:: text

        If id = match_indices[i][j] > 0,
C
chengduoZH 已提交
1168

1169 1170
            out[i][j][0 : K] = X[lod[i] + id][j % P][0 : K]
            out_weight[i][j] = 1.
C
chengduoZH 已提交
1171

1172
        Otherwise,
C
chengduoZH 已提交
1173

1174 1175
            out[j][j][0 : K] = {mismatch_value, mismatch_value, ...}
            out_weight[i][j] = 0.
C
chengduoZH 已提交
1176

1177
    2. Assigning out_weight based on `neg_indices` if `neg_indices` is provided:
C
chengduoZH 已提交
1178

1179 1180
    Assumed that the row offset for each instance in `neg_indices` is called neg_lod,
    for i-th instance and each `id` of neg_indices in this instance:
M
minqiyang 已提交
1181

1182
    .. code-block:: text
C
chengduoZH 已提交
1183

1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198
        out[i][id][0 : K] = {mismatch_value, mismatch_value, ...}
        out_weight[i][id] = 1.0

    Args:
       inputs (Variable): This input is a 3D LoDTensor with shape [M, P, K].
       matched_indices (Variable): Tensor<int>), The input matched indices
           is 2D Tenosr<int32> with shape [N, P], If MatchIndices[i][j] is -1,
           the j-th entity of column is not matched to any entity of row in
           i-th instance.
       negative_indices (Variable): The input negative example indices are
           an optional input with shape [Neg, 1] and int32 type, where Neg is
           the total number of negative example indices.
       mismatch_value (float32): Fill this value to the mismatched location.

    Returns:
M
minqiyang 已提交
1199 1200 1201 1202 1203
        tuple:
               A tuple(out, out_weight) is returned. out is a 3D Tensor with
               shape [N, P, K], N and P is the same as they are in
               `neg_indices`, K is the same as it in input of X. If
               `match_indices[i][j]`. out_weight is the weight for output with
1204 1205 1206 1207 1208 1209
               the shape of [N, P, 1].

    Examples:

        .. code-block:: python

1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225
            import paddle.fluid as fluid
            x = fluid.layers.data(
                name='x',
                shape=[4, 20, 4],
                dtype='float',
                lod_level=1,
                append_batch_size=False)
            matched_id = fluid.layers.data(
                name='indices',
                shape=[8, 20],
                dtype='int32',
                append_batch_size=False)
            trg, trg_weight = fluid.layers.target_assign(
                x,
                matched_id,
                mismatch_value=0)
1226 1227
    """
    helper = LayerHelper('target_assign', **locals())
X
Xin Pan 已提交
1228 1229
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    out_weight = helper.create_variable_for_type_inference(dtype='float32')
1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256
    helper.append_op(
        type='target_assign',
        inputs={
            'X': input,
            'MatchIndices': matched_indices,
            'NegIndices': negative_indices
        },
        outputs={'Out': out,
                 'OutWeight': out_weight},
        attrs={'mismatch_value': mismatch_value})
    return out, out_weight


def ssd_loss(location,
             confidence,
             gt_box,
             gt_label,
             prior_box,
             prior_box_var=None,
             background_label=0,
             overlap_threshold=0.5,
             neg_pos_ratio=3.0,
             neg_overlap=0.5,
             loc_loss_weight=1.0,
             conf_loss_weight=1.0,
             match_type='per_prediction',
             mining_type='max_negative',
1257
             normalize=True,
1258 1259
             sample_size=None):
    """
Y
yuyang18 已提交
1260
    **Multi-box loss layer for object detection algorithm of SSD**
1261

翟飞跃 已提交
1262 1263
    This layer is to compute detection loss for SSD given the location offset
    predictions, confidence predictions, prior boxes and ground-truth bounding
1264 1265 1266 1267
    boxes and labels, and the type of hard example mining. The returned loss
    is a weighted sum of the localization loss (or regression loss) and
    confidence loss (or classification loss) by performing the following steps:

Y
yuyang18 已提交
1268
    1. Find matched bounding box by bipartite matching algorithm.
Y
yuyang18 已提交
1269

1270
      1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
Y
yuyang18 已提交
1271

1272
      1.2 Compute matched boundding box by bipartite matching algorithm.
Y
yuyang18 已提交
1273

1274
    2. Compute confidence for mining hard examples
Y
yuyang18 已提交
1275

1276
      2.1. Get the target label based on matched indices.
Y
yuyang18 已提交
1277

1278
      2.2. Compute confidence loss.
Y
yuyang18 已提交
1279

1280 1281
    3. Apply hard example mining to get the negative example indices and update
       the matched indices.
Y
yuyang18 已提交
1282

1283
    4. Assign classification and regression targets
Y
yuyang18 已提交
1284

1285
      4.1. Encoded bbox according to the prior boxes.
Y
yuyang18 已提交
1286

1287
      4.2. Assign regression targets.
Y
yuyang18 已提交
1288

1289
      4.3. Assign classification targets.
Y
yuyang18 已提交
1290

1291
    5. Compute the overall objective loss.
Y
yuyang18 已提交
1292

1293
      5.1 Compute confidence loss.
Y
yuyang18 已提交
1294

1295
      5.1 Compute localization loss.
Y
yuyang18 已提交
1296

1297 1298 1299 1300 1301 1302 1303 1304 1305 1306
      5.3 Compute the overall weighted loss.

    Args:
        location (Variable): The location predictions are a 3D Tensor with
            shape [N, Np, 4], N is the batch size, Np is total number of
            predictions for each instance. 4 is the number of coordinate values,
            the layout is [xmin, ymin, xmax, ymax].
        confidence (Variable): The confidence predictions are a 3D Tensor
            with shape [N, Np, C], N and Np are the same as they are in
            `location`, C is the class number.
翟飞跃 已提交
1307
        gt_box (Variable): The ground-truth bounding boxes (bboxes) are a 2D
1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319
            LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth
            bboxes of mini-batch input.
        gt_label (Variable): The ground-truth labels are a 2D LoDTensor
            with shape [Ng, 1].
        prior_box (Variable): The prior boxes are a 2D Tensor with shape [Np, 4].
        prior_box_var (Variable): The variance of prior boxes are a 2D Tensor
            with shape [Np, 4].
        background_label (int): The index of background label, 0 by default.
        overlap_threshold (float): If match_type is 'per_prediction', use
            `overlap_threshold` to determine the extra matching bboxes when
             finding matched boxes. 0.5 by default.
        neg_pos_ratio (float): The ratio of the negative boxes to the positive
翟飞跃 已提交
1320
            boxes, used only when mining_type is 'max_negative', 3.0 by default.
1321
        neg_overlap (float): The negative overlap upper bound for the unmatched
1322
            predictions. Use only when mining_type is 'max_negative',
1323 1324 1325 1326
            0.5 by default.
        loc_loss_weight (float): Weight for localization loss, 1.0 by default.
        conf_loss_weight (float): Weight for confidence loss, 1.0 by default.
        match_type (str): The type of matching method during training, should
翟飞跃 已提交
1327
            be 'bipartite' or 'per_prediction', 'per_prediction' by default.
1328 1329
        mining_type (str): The hard example mining type, should be 'hard_example'
            or 'max_negative', now only support `max_negative`.
1330
        normalize (bool): Whether to normalize the SSD loss by the total number
Y
yuyang18 已提交
1331
            of output locations, True by default.
1332 1333
        sample_size (int): The max sample size of negative box, used only when
            mining_type is 'hard_example'.
1334 1335

    Returns:
Y
yuyang18 已提交
1336 1337
        The weighted sum of the localization loss and confidence loss, with \
        shape [N * Np, 1], N and Np are the same as they are in `location`.
1338 1339

    Raises:
Y
yuyang18 已提交
1340 1341
        ValueError: If mining_type is 'hard_example', now only support mining \
        type of `max_negative`.
Y
yuyang18 已提交
1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360

    Examples:
        >>> pb = fluid.layers.data(
        >>>                   name='prior_box',
        >>>                   shape=[10, 4],
        >>>                   append_batch_size=False,
        >>>                   dtype='float32')
        >>> pbv = fluid.layers.data(
        >>>                   name='prior_box_var',
        >>>                   shape=[10, 4],
        >>>                   append_batch_size=False,
        >>>                   dtype='float32')
        >>> loc = fluid.layers.data(name='target_box', shape=[10, 4], dtype='float32')
        >>> scores = fluid.layers.data(name='scores', shape=[10, 21], dtype='float32')
        >>> gt_box = fluid.layers.data(
        >>>         name='gt_box', shape=[4], lod_level=1, dtype='float32')
        >>> gt_label = fluid.layers.data(
        >>>         name='gt_label', shape=[1], lod_level=1, dtype='float32')
        >>> loss = fluid.layers.ssd_loss(loc, scores, gt_box, gt_label, pb, pbv)
1361 1362 1363 1364 1365 1366 1367
    """

    helper = LayerHelper('ssd_loss', **locals())
    if mining_type != 'max_negative':
        raise ValueError("Only support mining_type == max_negative now.")

    num, num_prior, num_class = confidence.shape
G
merge  
gongweibao 已提交
1368
    conf_shape = nn.shape(confidence)
1369 1370

    def __reshape_to_2d(var):
1371
        return nn.flatten(x=var, axis=2)
1372 1373 1374 1375 1376

    # 1. Find matched boundding box by prior box.
    #   1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
    iou = iou_similarity(x=gt_box, y=prior_box)
    #   1.2 Compute matched boundding box by bipartite matching algorithm.
1377 1378
    matched_indices, matched_dist = bipartite_match(iou, match_type,
                                                    overlap_threshold)
1379 1380 1381

    # 2. Compute confidence for mining hard examples
    # 2.1. Get the target label based on matched indices
1382 1383
    gt_label = nn.reshape(
        x=gt_label, shape=(len(gt_label.shape) - 1) * (0, ) + (-1, 1))
1384
    gt_label.stop_gradient = True
1385 1386 1387 1388 1389 1390 1391
    target_label, _ = target_assign(
        gt_label, matched_indices, mismatch_value=background_label)
    # 2.2. Compute confidence loss.
    # Reshape confidence to 2D tensor.
    confidence = __reshape_to_2d(confidence)
    target_label = tensor.cast(x=target_label, dtype='int64')
    target_label = __reshape_to_2d(target_label)
1392
    target_label.stop_gradient = True
1393 1394
    conf_loss = nn.softmax_with_cross_entropy(confidence, target_label)
    # 3. Mining hard examples
G
merge  
gongweibao 已提交
1395
    actual_shape = nn.slice(conf_shape, axes=[0], starts=[0], ends=[2])
1396
    actual_shape.stop_gradient = True
1397
    conf_loss = nn.reshape(
1398
        x=conf_loss, shape=(num, num_prior), actual_shape=actual_shape)
1399
    conf_loss.stop_gradient = True
X
Xin Pan 已提交
1400
    neg_indices = helper.create_variable_for_type_inference(dtype='int32')
1401
    dtype = matched_indices.dtype
X
Xin Pan 已提交
1402 1403
    updated_matched_indices = helper.create_variable_for_type_inference(
        dtype=dtype)
1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417
    helper.append_op(
        type='mine_hard_examples',
        inputs={
            'ClsLoss': conf_loss,
            'LocLoss': None,
            'MatchIndices': matched_indices,
            'MatchDist': matched_dist,
        },
        outputs={
            'NegIndices': neg_indices,
            'UpdatedMatchIndices': updated_matched_indices
        },
        attrs={
            'neg_pos_ratio': neg_pos_ratio,
B
Bai Yifan 已提交
1418
            'neg_dist_threshold': neg_overlap,
1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443
            'mining_type': mining_type,
            'sample_size': sample_size,
        })

    # 4. Assign classification and regression targets
    # 4.1. Encoded bbox according to the prior boxes.
    encoded_bbox = box_coder(
        prior_box=prior_box,
        prior_box_var=prior_box_var,
        target_box=gt_box,
        code_type='encode_center_size')
    # 4.2. Assign regression targets
    target_bbox, target_loc_weight = target_assign(
        encoded_bbox, updated_matched_indices, mismatch_value=background_label)
    # 4.3. Assign classification targets
    target_label, target_conf_weight = target_assign(
        gt_label,
        updated_matched_indices,
        negative_indices=neg_indices,
        mismatch_value=background_label)

    # 5. Compute loss.
    # 5.1 Compute confidence loss.
    target_label = __reshape_to_2d(target_label)
    target_label = tensor.cast(x=target_label, dtype='int64')
1444

1445 1446 1447 1448
    conf_loss = nn.softmax_with_cross_entropy(confidence, target_label)
    target_conf_weight = __reshape_to_2d(target_conf_weight)
    conf_loss = conf_loss * target_conf_weight

1449 1450 1451 1452
    # the target_label and target_conf_weight do not have gradient.
    target_label.stop_gradient = True
    target_conf_weight.stop_gradient = True

1453 1454 1455 1456 1457 1458 1459 1460
    # 5.2 Compute regression loss.
    location = __reshape_to_2d(location)
    target_bbox = __reshape_to_2d(target_bbox)

    loc_loss = nn.smooth_l1(location, target_bbox)
    target_loc_weight = __reshape_to_2d(target_loc_weight)
    loc_loss = loc_loss * target_loc_weight

1461 1462 1463 1464
    # the target_bbox and target_loc_weight do not have gradient.
    target_bbox.stop_gradient = True
    target_loc_weight.stop_gradient = True

1465 1466
    # 5.3 Compute overall weighted loss.
    loss = conf_loss_weight * conf_loss + loc_loss_weight * loc_loss
1467
    # reshape to [N, Np], N is the batch size and Np is the prior box number.
1468
    loss = nn.reshape(x=loss, shape=(num, num_prior), actual_shape=actual_shape)
1469 1470 1471 1472 1473
    loss = nn.reduce_sum(loss, dim=1, keep_dim=True)
    if normalize:
        normalizer = nn.reduce_sum(target_loc_weight)
        loss = loss / normalizer

1474
    return loss
C
chengduoZH 已提交
1475 1476


1477 1478 1479 1480
def prior_box(input,
              image,
              min_sizes,
              max_sizes=None,
1481
              aspect_ratios=[1.],
1482 1483 1484 1485 1486
              variance=[0.1, 0.1, 0.2, 0.2],
              flip=False,
              clip=False,
              steps=[0.0, 0.0],
              offset=0.5,
1487 1488
              name=None,
              min_max_aspect_ratios_order=False):
1489
    """
Q
update  
qiaolongfei 已提交
1490
    **Prior Box Operator**
1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501

    Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm.
    Each position of the input produce N prior boxes, N is determined by
    the count of min_sizes, max_sizes and aspect_ratios, The size of the
    box is in range(min_size, max_size) interval, which is generated in
    sequence according to the aspect_ratios.

    Args:
       input(Variable): The Input Variables, the format is NCHW.
       image(Variable): The input image data of PriorBoxOp,
            the layout is NCHW.
1502
       min_sizes(list|tuple|float value): min sizes of generated prior boxes.
1503 1504
       max_sizes(list|tuple|None): max sizes of generated prior boxes.
            Default: None.
1505 1506
       aspect_ratios(list|tuple|float value): the aspect ratios of generated
            prior boxes. Default: [1.].
1507 1508 1509 1510
       variance(list|tuple): the variances to be encoded in prior boxes.
            Default:[0.1, 0.1, 0.2, 0.2].
       flip(bool): Whether to flip aspect ratios. Default:False.
       clip(bool): Whether to clip out-of-boundary boxes. Default: False.
翟飞跃 已提交
1511
       step(list|tuple): Prior boxes step across width and height, If
1512
            step[0] == 0.0/step[1] == 0.0, the prior boxes step across
1513 1514
            height/weight of the input will be automatically calculated.
            Default: [0., 0.]
1515 1516
       offset(float): Prior boxes center offset. Default: 0.5
       name(str): Name of the prior box op. Default: None.
1517
       min_max_aspect_ratios_order(bool): If set True, the output prior box is
M
minqiyang 已提交
1518
            in order of [min, max, aspect_ratios], which is consistent with
1519 1520 1521
            Caffe. Please note, this order affects the weights order of
            convolution layer followed by and does not affect the final
            detection results. Default: False.
1522 1523

    Returns:
Q
update  
qiaolongfei 已提交
1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
        tuple: A tuple with two Variable (boxes, variances)

        boxes: the output prior boxes of PriorBox.
        The layout is [H, W, num_priors, 4].
        H is the height of input, W is the width of input,
        num_priors is the total
        box count of each position of input.

        variances: the expanded variances of PriorBox.
        The layout is [H, W, num_priors, 4].
        H is the height of input, W is the width of input
        num_priors is the total
        box count of each position of input
1537 1538 1539 1540


    Examples:
        .. code-block:: python
Q
update  
qiaolongfei 已提交
1541

R
ruri 已提交
1542 1543
            input = fluid.layers.data(name="input", shape=[3,6,9])
            images = fluid.layers.data(name="images", shape=[3,9,12])
Q
update  
qiaolongfei 已提交
1544
            box, var = fluid.layers.prior_box(
R
ruri 已提交
1545
                input=input,
Q
update  
qiaolongfei 已提交
1546 1547 1548 1549
                image=images,
                min_sizes=[100.],
                flip=True,
                clip=True)
1550 1551 1552 1553
    """
    helper = LayerHelper("prior_box", **locals())
    dtype = helper.input_dtype()

1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568
    def _is_list_or_tuple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))

    if not _is_list_or_tuple_(min_sizes):
        min_sizes = [min_sizes]
    if not _is_list_or_tuple_(aspect_ratios):
        aspect_ratios = [aspect_ratios]
    if not (_is_list_or_tuple_(steps) and len(steps) == 2):
        raise ValueError('steps should be a list or tuple ',
                         'with length 2, (step_width, step_height).')

    min_sizes = list(map(float, min_sizes))
    aspect_ratios = list(map(float, aspect_ratios))
    steps = list(map(float, steps))

1569 1570 1571 1572 1573 1574 1575 1576
    attrs = {
        'min_sizes': min_sizes,
        'aspect_ratios': aspect_ratios,
        'variances': variance,
        'flip': flip,
        'clip': clip,
        'step_w': steps[0],
        'step_h': steps[1],
1577 1578
        'offset': offset,
        'min_max_aspect_ratios_order': min_max_aspect_ratios_order
1579 1580
    }
    if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0:
1581 1582
        if not _is_list_or_tuple_(max_sizes):
            max_sizes = [max_sizes]
1583 1584
        attrs['max_sizes'] = max_sizes

X
Xin Pan 已提交
1585 1586
    box = helper.create_variable_for_type_inference(dtype)
    var = helper.create_variable_for_type_inference(dtype)
1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598
    helper.append_op(
        type="prior_box",
        inputs={"Input": input,
                "Image": image},
        outputs={"Boxes": box,
                 "Variances": var},
        attrs=attrs, )
    box.stop_gradient = True
    var.stop_gradient = True
    return box, var


R
ruri 已提交
1599 1600 1601 1602 1603 1604 1605 1606 1607
def density_prior_box(input,
                      image,
                      densities=None,
                      fixed_sizes=None,
                      fixed_ratios=None,
                      variance=[0.1, 0.1, 0.2, 0.2],
                      clip=False,
                      steps=[0.0, 0.0],
                      offset=0.5,
1608
                      flatten_to_2d=False,
R
ruri 已提交
1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639
                      name=None):
    """
    **Density Prior Box Operator**

    Generate density prior boxes for SSD(Single Shot MultiBox Detector) 
    algorithm. Each position of the input produce N prior boxes, N is 
    determined by the count of densities, fixed_sizes and fixed_ratios. 
    Boxes center at grid points around each input position is generated by 
    this operator, and the grid points is determined by densities and 
    the count of density prior box is determined by fixed_sizes and fixed_ratios. 
    Obviously, the number of fixed_sizes is equal to the number of densities.
    For densities_i in densities:
    N_density_prior_box =sum(N_fixed_ratios * densities_i^2),

    Args:
       input(Variable): The Input Variables, the format is NCHW.
       image(Variable): The input image data of PriorBoxOp,
            the layout is NCHW.
       densities(list|tuple|None): the densities of generated density prior 
            boxes, this attribute should be a list or tuple of integers. 
            Default: None.
       fixed_sizes(list|tuple|None): the fixed sizes of generated density
            prior boxes, this attribute should a list or tuple of same 
            length with :attr:`densities`. Default: None.
       fixed_ratios(list|tuple|None): the fixed ratios of generated density
            prior boxes, if this attribute is not set and :attr:`densities`
            and :attr:`fix_sizes` is set, :attr:`aspect_ratios` will be used
            to generate density prior boxes.
       variance(list|tuple): the variances to be encoded in density prior boxes.
            Default:[0.1, 0.1, 0.2, 0.2].
       clip(bool): Whether to clip out-of-boundary boxes. Default: False.
翟飞跃 已提交
1640
       step(list|tuple): Prior boxes step across width and height, If
R
ruri 已提交
1641 1642 1643 1644
            step[0] == 0.0/step[1] == 0.0, the density prior boxes step across
            height/weight of the input will be automatically calculated.
            Default: [0., 0.]
       offset(float): Prior boxes center offset. Default: 0.5
1645 1646
       flatten_to_2d(bool): Whether to flatten output prior boxes and variance
           to 2D shape, the second dim is 4. Default: False.
R
ruri 已提交
1647 1648 1649 1650 1651 1652
       name(str): Name of the density prior box op. Default: None.

    Returns:
        tuple: A tuple with two Variable (boxes, variances)

        boxes: the output density prior boxes of PriorBox.
1653 1654 1655 1656
            The layout is [H, W, num_priors, 4] when flatten_to_2d is False.
            The layout is [H * W * num_priors, 4] when flatten_to_2d is True.
            H is the height of input, W is the width of input,
            num_priors is the total box count of each position of input.
R
ruri 已提交
1657 1658

        variances: the expanded variances of PriorBox.
1659 1660 1661 1662
            The layout is [H, W, num_priors, 4] when flatten_to_2d is False.
            The layout is [H * W * num_priors, 4] when flatten_to_2d is True.
            H is the height of input, W is the width of input
            num_priors is the total box count of each position of input.
R
ruri 已提交
1663 1664 1665 1666 1667


    Examples:
        .. code-block:: python

R
ruri 已提交
1668 1669
            input = fluid.layers.data(name="input", shape=[3,6,9])
            images = fluid.layers.data(name="images", shape=[3,9,12])
R
ruri 已提交
1670
            box, var = fluid.layers.density_prior_box(
R
ruri 已提交
1671
                input=input,
R
ruri 已提交
1672
                image=images,
1673 1674 1675 1676 1677
                densities=[4, 2, 1],
                fixed_sizes=[32.0, 64.0, 128.0],
                fixed_ratios=[1.],
                clip=True,
                flatten_to_2d=True)
R
ruri 已提交
1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707
    """
    helper = LayerHelper("density_prior_box", **locals())
    dtype = helper.input_dtype()

    def _is_list_or_tuple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))

    if not _is_list_or_tuple_(densities):
        raise TypeError('densities should be a list or a tuple or None.')
    if not _is_list_or_tuple_(fixed_sizes):
        raise TypeError('fixed_sizes should be a list or a tuple or None.')
    if not _is_list_or_tuple_(fixed_ratios):
        raise TypeError('fixed_ratios should be a list or a tuple or None.')
    if len(densities) != len(fixed_sizes):
        raise ValueError('densities and fixed_sizes length should be euqal.')
    if not (_is_list_or_tuple_(steps) and len(steps) == 2):
        raise ValueError('steps should be a list or tuple ',
                         'with length 2, (step_width, step_height).')

    densities = list(map(int, densities))
    fixed_sizes = list(map(float, fixed_sizes))
    fixed_ratios = list(map(float, fixed_ratios))
    steps = list(map(float, steps))

    attrs = {
        'variances': variance,
        'clip': clip,
        'step_w': steps[0],
        'step_h': steps[1],
        'offset': offset,
1708 1709 1710 1711
        'densities': densities,
        'fixed_sizes': fixed_sizes,
        'fixed_ratios': fixed_ratios,
        'flatten_to_2d': flatten_to_2d,
R
ruri 已提交
1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726
    }
    box = helper.create_variable_for_type_inference(dtype)
    var = helper.create_variable_for_type_inference(dtype)
    helper.append_op(
        type="density_prior_box",
        inputs={"Input": input,
                "Image": image},
        outputs={"Boxes": box,
                 "Variances": var},
        attrs=attrs, )
    box.stop_gradient = True
    var.stop_gradient = True
    return box, var


C
chengduoZH 已提交
1727
def multi_box_head(inputs,
C
chengduoZH 已提交
1728 1729
                   image,
                   base_size,
C
chengduoZH 已提交
1730
                   num_classes,
C
chengduoZH 已提交
1731
                   aspect_ratios,
1732 1733
                   min_ratio=None,
                   max_ratio=None,
C
chengduoZH 已提交
1734 1735
                   min_sizes=None,
                   max_sizes=None,
C
chengduoZH 已提交
1736 1737 1738 1739
                   steps=None,
                   step_w=None,
                   step_h=None,
                   offset=0.5,
1740 1741
                   variance=[0.1, 0.1, 0.2, 0.2],
                   flip=True,
C
chengduoZH 已提交
1742
                   clip=False,
C
chengduoZH 已提交
1743
                   kernel_size=1,
C
chengduoZH 已提交
1744
                   pad=0,
C
chengduoZH 已提交
1745
                   stride=1,
1746 1747
                   name=None,
                   min_max_aspect_ratios_order=False):
C
chengduoZH 已提交
1748
    """
C
chengduoZH 已提交
1749 1750
    Generate prior boxes for SSD(Single Shot MultiBox Detector)
    algorithm. The details of this algorithm, please refer the
Q
update  
qiaolongfei 已提交
1751
    section 2.2 of SSD paper `SSD: Single Shot MultiBox Detector
C
chengduoZH 已提交
1752
    <https://arxiv.org/abs/1512.02325>`_ .
C
chengduoZH 已提交
1753 1754

    Args:
1755
       inputs(list|tuple): The list of input Variables, the format
C
chengduoZH 已提交
1756
            of all Variables is NCHW.
C
chengduoZH 已提交
1757 1758
       image(Variable): The input image data of PriorBoxOp,
            the layout is NCHW.
C
chengduoZH 已提交
1759 1760
       base_size(int): the base_size is used to get min_size
            and max_size according to min_ratio and max_ratio.
C
chengduoZH 已提交
1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782
       num_classes(int): The number of classes.
       aspect_ratios(list|tuple): the aspect ratios of generated prior
            boxes. The length of input and aspect_ratios must be equal.
       min_ratio(int): the min ratio of generated prior boxes.
       max_ratio(int): the max ratio of generated prior boxes.
       min_sizes(list|tuple|None): If `len(inputs) <=2`,
            min_sizes must be set up, and the length of min_sizes
            should equal to the length of inputs. Default: None.
       max_sizes(list|tuple|None): If `len(inputs) <=2`,
            max_sizes must be set up, and the length of min_sizes
            should equal to the length of inputs. Default: None.
       steps(list|tuple): If step_w and step_h are the same,
            step_w and step_h can be replaced by steps.
       step_w(list|tuple): Prior boxes step
            across width. If step_w[i] == 0.0, the prior boxes step
            across width of the inputs[i] will be automatically
            calculated. Default: None.
       step_h(list|tuple): Prior boxes step across height, If
            step_h[i] == 0.0, the prior boxes step across height of
            the inputs[i] will be automatically calculated. Default: None.
       offset(float): Prior boxes center offset. Default: 0.5
       variance(list|tuple): the variances to be encoded in prior boxes.
1783
            Default:[0.1, 0.1, 0.2, 0.2].
C
chengduoZH 已提交
1784 1785 1786 1787 1788 1789
       flip(bool): Whether to flip aspect ratios. Default:False.
       clip(bool): Whether to clip out-of-boundary boxes. Default: False.
       kernel_size(int): The kernel size of conv2d. Default: 1.
       pad(int|list|tuple): The padding of conv2d. Default:0.
       stride(int|list|tuple): The stride of conv2d. Default:1,
       name(str): Name of the prior box layer. Default: None.
1790
       min_max_aspect_ratios_order(bool): If set True, the output prior box is
M
minqiyang 已提交
1791
            in order of [min, max, aspect_ratios], which is consistent with
1792 1793 1794
            Caffe. Please note, this order affects the weights order of
            convolution layer followed by and does not affect the fininal
            detection results. Default: False.
C
chengduoZH 已提交
1795 1796

    Returns:
Q
update  
qiaolongfei 已提交
1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811
        tuple: A tuple with four Variables. (mbox_loc, mbox_conf, boxes, variances)

        mbox_loc: The predicted boxes' location of the inputs. The layout
        is [N, H*W*Priors, 4]. where Priors is the number of predicted
        boxes each position of each input.

        mbox_conf: The predicted boxes' confidence of the inputs. The layout
        is [N, H*W*Priors, C]. where Priors is the number of predicted boxes
        each position of each input and C is the number of Classes.

        boxes: the output prior boxes of PriorBox. The layout is [num_priors, 4].
        num_priors is the total box count of each position of inputs.

        variances: the expanded variances of PriorBox. The layout is
        [num_priors, 4]. num_priors is the total box count of each position of inputs
C
chengduoZH 已提交
1812

C
chengduoZH 已提交
1813 1814 1815

    Examples:
        .. code-block:: python
Q
update  
qiaolongfei 已提交
1816

1817 1818 1819 1820 1821 1822 1823 1824 1825 1826
          import paddle.fluid as fluid

          images = fluid.layers.data(name='data', shape=[3, 300, 300], dtype='float32')
          conv1 = fluid.layers.data(name='conv1', shape=[512, 19, 19], dtype='float32')
          conv2 = fluid.layers.data(name='conv2', shape=[1024, 10, 10], dtype='float32')
          conv3 = fluid.layers.data(name='conv3', shape=[512, 5, 5], dtype='float32')
          conv4 = fluid.layers.data(name='conv4', shape=[256, 3, 3], dtype='float32')
          conv5 = fluid.layers.data(name='conv5', shape=[256, 2, 2], dtype='float32')
          conv6 = fluid.layers.data(name='conv6', shape=[128, 1, 1], dtype='float32')

Q
update  
qiaolongfei 已提交
1827
          mbox_locs, mbox_confs, box, var = fluid.layers.multi_box_head(
1828
            inputs=[conv1, conv2, conv3, conv4, conv5, conv6],
C
chengduoZH 已提交
1829 1830 1831 1832 1833 1834 1835 1836 1837
            image=images,
            num_classes=21,
            min_ratio=20,
            max_ratio=90,
            aspect_ratios=[[2.], [2., 3.], [2., 3.], [2., 3.], [2.], [2.]],
            base_size=300,
            offset=0.5,
            flip=True,
            clip=True)
C
chengduoZH 已提交
1838 1839
    """

C
chengduoZH 已提交
1840
    def _reshape_with_axis_(input, axis=1):
1841
        out = nn.flatten(x=input, axis=axis)
C
chengduoZH 已提交
1842
        return out
1843

1844 1845
    def _is_list_or_tuple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))
1846

C
chengduoZH 已提交
1847 1848 1849 1850
    def _is_list_or_tuple_and_equal(data, length, err_info):
        if not (_is_list_or_tuple_(data) and len(data) == length):
            raise ValueError(err_info)

1851 1852
    if not _is_list_or_tuple_(inputs):
        raise ValueError('inputs should be a list or tuple.')
C
chengduoZH 已提交
1853

C
chengduoZH 已提交
1854 1855 1856 1857 1858
    num_layer = len(inputs)

    if num_layer <= 2:
        assert min_sizes is not None and max_sizes is not None
        assert len(min_sizes) == num_layer and len(max_sizes) == num_layer
1859
    elif min_sizes is None and max_sizes is None:
C
chengduoZH 已提交
1860 1861 1862
        min_sizes = []
        max_sizes = []
        step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2)))
M
minqiyang 已提交
1863
        for ratio in six.moves.range(min_ratio, max_ratio + 1, step):
C
chengduoZH 已提交
1864 1865 1866 1867 1868
            min_sizes.append(base_size * ratio / 100.)
            max_sizes.append(base_size * (ratio + step) / 100.)
        min_sizes = [base_size * .10] + min_sizes
        max_sizes = [base_size * .20] + max_sizes

C
chengduoZH 已提交
1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891
    if aspect_ratios:
        _is_list_or_tuple_and_equal(
            aspect_ratios, num_layer,
            'aspect_ratios should be list or tuple, and the length of inputs '
            'and aspect_ratios should be the same.')
    if step_h:
        _is_list_or_tuple_and_equal(
            step_h, num_layer,
            'step_h should be list or tuple, and the length of inputs and '
            'step_h should be the same.')
    if step_w:
        _is_list_or_tuple_and_equal(
            step_w, num_layer,
            'step_w should be list or tuple, and the length of inputs and '
            'step_w should be the same.')
    if steps:
        _is_list_or_tuple_and_equal(
            steps, num_layer,
            'steps should be list or tuple, and the length of inputs and '
            'step_w should be the same.')
        step_w = steps
        step_h = steps

C
chengduoZH 已提交
1892 1893
    mbox_locs = []
    mbox_confs = []
C
chengduoZH 已提交
1894 1895
    box_results = []
    var_results = []
C
chengduoZH 已提交
1896 1897
    for i, input in enumerate(inputs):
        min_size = min_sizes[i]
C
chengduoZH 已提交
1898 1899
        max_size = max_sizes[i]

1900
        if not _is_list_or_tuple_(min_size):
C
chengduoZH 已提交
1901
            min_size = [min_size]
C
chengduoZH 已提交
1902 1903
        if not _is_list_or_tuple_(max_size):
            max_size = [max_size]
C
chengduoZH 已提交
1904 1905 1906 1907

        aspect_ratio = []
        if aspect_ratios is not None:
            aspect_ratio = aspect_ratios[i]
1908
            if not _is_list_or_tuple_(aspect_ratio):
C
chengduoZH 已提交
1909
                aspect_ratio = [aspect_ratio]
1910
        step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0]
C
chengduoZH 已提交
1911

1912
        box, var = prior_box(input, image, min_size, max_size, aspect_ratio,
1913 1914
                             variance, flip, clip, step, offset, None,
                             min_max_aspect_ratios_order)
C
chengduoZH 已提交
1915 1916 1917 1918 1919

        box_results.append(box)
        var_results.append(var)

        num_boxes = box.shape[2]
C
chengduoZH 已提交
1920

1921
        # get loc
Y
Yuan Gao 已提交
1922
        num_loc_output = num_boxes * 4
1923
        mbox_loc = nn.conv2d(
C
chengduoZH 已提交
1924
            input=input,
1925 1926 1927 1928 1929
            num_filters=num_loc_output,
            filter_size=kernel_size,
            padding=pad,
            stride=stride)

1930
        mbox_loc = nn.transpose(mbox_loc, perm=[0, 2, 3, 1])
1931
        compile_shape = [
1932
            mbox_loc.shape[0], cpt.floor_division(
M
minqiyang 已提交
1933
                mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3], 4), 4
Y
Yuan Gao 已提交
1934
        ]
1935 1936 1937
        run_shape = tensor.assign(numpy.array([0, -1, 4]).astype("int32"))
        mbox_loc_flatten = nn.reshape(
            mbox_loc, shape=compile_shape, actual_shape=run_shape)
Y
Yuan Gao 已提交
1938
        mbox_locs.append(mbox_loc_flatten)
C
chengduoZH 已提交
1939

1940
        # get conf
C
chengduoZH 已提交
1941
        num_conf_output = num_boxes * num_classes
1942
        conf_loc = nn.conv2d(
C
chengduoZH 已提交
1943
            input=input,
1944 1945 1946 1947
            num_filters=num_conf_output,
            filter_size=kernel_size,
            padding=pad,
            stride=stride)
1948
        conf_loc = nn.transpose(conf_loc, perm=[0, 2, 3, 1])
1949 1950
        new_shape = [0, -1, num_classes]
        compile_shape = [
1951 1952 1953
            conf_loc.shape[0],
            cpt.floor_division(conf_loc.shape[1] * conf_loc.shape[2] *
                               conf_loc.shape[3], num_classes), num_classes
Y
Yuan Gao 已提交
1954
        ]
1955 1956 1957 1958
        run_shape = tensor.assign(
            numpy.array([0, -1, num_classes]).astype("int32"))
        conf_loc_flatten = nn.reshape(
            conf_loc, shape=compile_shape, actual_shape=run_shape)
Y
Yuan Gao 已提交
1959
        mbox_confs.append(conf_loc_flatten)
C
chengduoZH 已提交
1960

C
chengduoZH 已提交
1961 1962 1963
    if len(box_results) == 1:
        box = box_results[0]
        var = var_results[0]
Y
Yuan Gao 已提交
1964 1965
        mbox_locs_concat = mbox_locs[0]
        mbox_confs_concat = mbox_confs[0]
C
chengduoZH 已提交
1966 1967 1968 1969 1970 1971 1972 1973 1974
    else:
        reshaped_boxes = []
        reshaped_vars = []
        for i in range(len(box_results)):
            reshaped_boxes.append(_reshape_with_axis_(box_results[i], axis=3))
            reshaped_vars.append(_reshape_with_axis_(var_results[i], axis=3))

        box = tensor.concat(reshaped_boxes)
        var = tensor.concat(reshaped_vars)
Y
Yuan Gao 已提交
1975 1976
        mbox_locs_concat = tensor.concat(mbox_locs, axis=1)
        mbox_confs_concat = tensor.concat(mbox_confs, axis=1)
C
chengduoZH 已提交
1977

1978 1979
    box.stop_gradient = True
    var.stop_gradient = True
Y
Yuan Gao 已提交
1980
    return mbox_locs_concat, mbox_confs_concat, box, var
1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000


def anchor_generator(input,
                     anchor_sizes=None,
                     aspect_ratios=None,
                     variance=[0.1, 0.1, 0.2, 0.2],
                     stride=None,
                     offset=0.5,
                     name=None):
    """
    **Anchor generator operator**

    Generate anchors for Faster RCNN algorithm.
    Each position of the input produce N anchors, N =
    size(anchor_sizes) * size(aspect_ratios). The order of generated anchors
    is firstly aspect_ratios loop then anchor_sizes loop.

    Args:
       input(Variable): The input feature map, the format is NCHW.
       anchor_sizes(list|tuple|float): The anchor sizes of generated anchors,
H
haowang101779990 已提交
2001 2002
                                       given in absolute pixels e.g. [64., 128., 256., 512.].
                                       For instance, the anchor size of 64 means the area of this anchor equals to 64**2.
2003
       aspect_ratios(list|tuple|float): The height / width ratios of generated
H
haowang101779990 已提交
2004
                                        anchors, e.g. [0.5, 1.0, 2.0].
2005
       variance(list|tuple): The variances to be used in box regression deltas.
H
haowang101779990 已提交
2006
                             Default:[0.1, 0.1, 0.2, 0.2].
翟飞跃 已提交
2007
       stride(list|tuple): The anchors stride across width and height,e.g. [16.0, 16.0]
2008 2009 2010 2011
       offset(float): Prior boxes center offset. Default: 0.5
       name(str): Name of the prior box op. Default: None.

    Returns:
H
haowang101779990 已提交
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024
        Anchors(Variable),Variances(Variable):  
        
              two variables:
        
              - Anchors(Variable): The output anchors with a layout of [H, W, num_anchors, 4]. \
                H is the height of input, W is the width of input, \
                num_anchors is the box count of each position.  \
                Each anchor is in (xmin, ymin, xmax, ymax) format an unnormalized. 
              - Variances(Variable): The expanded variances of anchors \
                with a layout of [H, W, num_priors, 4]. \
                H is the height of input, W is the width of input \
                num_anchors is the box count of each position. \
                Each variance is in (xcenter, ycenter, w, h) format.
2025 2026 2027 2028 2029 2030


    Examples:

        .. code-block:: python

J
jerrywgz 已提交
2031 2032
            conv1 = fluid.layers.data(name='conv1', shape=[48, 16, 16], dtype='float32')
            anchor, var = fluid.layers.anchor_generator(
2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065
                input=conv1,
                anchor_sizes=[64, 128, 256, 512],
                aspect_ratios=[0.5, 1.0, 2.0],
                variance=[0.1, 0.1, 0.2, 0.2],
                stride=[16.0, 16.0],
                offset=0.5)
    """
    helper = LayerHelper("anchor_generator", **locals())
    dtype = helper.input_dtype()

    def _is_list_or_tuple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))

    if not _is_list_or_tuple_(anchor_sizes):
        anchor_sizes = [anchor_sizes]
    if not _is_list_or_tuple_(aspect_ratios):
        aspect_ratios = [aspect_ratios]
    if not (_is_list_or_tuple_(stride) and len(stride) == 2):
        raise ValueError('stride should be a list or tuple ',
                         'with length 2, (stride_width, stride_height).')

    anchor_sizes = list(map(float, anchor_sizes))
    aspect_ratios = list(map(float, aspect_ratios))
    stride = list(map(float, stride))

    attrs = {
        'anchor_sizes': anchor_sizes,
        'aspect_ratios': aspect_ratios,
        'variances': variance,
        'stride': stride,
        'offset': offset
    }

X
Xin Pan 已提交
2066 2067
    anchor = helper.create_variable_for_type_inference(dtype)
    var = helper.create_variable_for_type_inference(dtype)
2068 2069 2070 2071 2072 2073 2074 2075 2076
    helper.append_op(
        type="anchor_generator",
        inputs={"Input": input},
        outputs={"Anchors": anchor,
                 "Variances": var},
        attrs=attrs, )
    anchor.stop_gradient = True
    var.stop_gradient = True
    return anchor, var
2077 2078


W
whs 已提交
2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098
def roi_perspective_transform(input,
                              rois,
                              transformed_height,
                              transformed_width,
                              spatial_scale=1.0):
    """
    ROI perspective transform op.

    Args:
        input (Variable): The input of ROIPerspectiveTransformOp. The format of 
                          input tensor is NCHW. Where N is batch size, C is the
                          number of input channels, H is the height of the feature,
                          and W is the width of the feature.
        rois (Variable):  ROIs (Regions of Interest) to be transformed. It should be
                          a 2-D LoDTensor of shape (num_rois, 8). Given as 
                          [[x1, y1, x2, y2, x3, y3, x4, y4], ...], (x1, y1) is the 
                          top left coordinates, and (x2, y2) is the top right 
                          coordinates, and (x3, y3) is the bottom right coordinates, 
                          and (x4, y4) is the bottom left coordinates.
        transformed_height (integer): The height of transformed output.
S
SunGaofeng 已提交
2099
        transformed_width (integer): The width of transformed output.
W
whs 已提交
2100 2101 2102 2103 2104 2105 2106 2107 2108
        spatial_scale (float): Spatial scale factor to scale ROI coords. Default: 1.0

    Returns:
        Variable: The output of ROIPerspectiveTransformOp which is a 4-D tensor with shape 
                  (num_rois, channels, transformed_h, transformed_w).

    Examples:
        .. code-block:: python

S
SunGaofeng 已提交
2109
            import paddle.fluid as fluid
2110

S
SunGaofeng 已提交
2111 2112 2113
            x = fluid.layers.data(name='x', shape=[256, 28, 28], dtype='float32')
            rois = fluid.layers.data(name='rois', shape=[8], lod_level=1, dtype='float32')
            out = fluid.layers.roi_perspective_transform(x, rois, 7, 7, 1.0)
W
whs 已提交
2114 2115 2116
    """
    helper = LayerHelper('roi_perspective_transform', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
2117
    out = helper.create_variable_for_type_inference(dtype)
2118 2119
    out2in_idx = helper.create_variable_for_type_inference(dtype="int32")
    out2in_w = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
2120 2121 2122 2123
    helper.append_op(
        type="roi_perspective_transform",
        inputs={"X": input,
                "ROIs": rois},
2124 2125 2126 2127 2128
        outputs={
            "Out": out,
            "Out2InIdx": out2in_idx,
            "Out2InWeights": out2in_w
        },
W
whs 已提交
2129 2130 2131 2132 2133 2134 2135 2136
        attrs={
            "transformed_height": transformed_height,
            "transformed_width": transformed_width,
            "spatial_scale": spatial_scale
        })
    return out


2137 2138
def generate_proposal_labels(rpn_rois,
                             gt_classes,
2139
                             is_crowd,
2140
                             gt_boxes,
2141
                             im_info,
2142 2143 2144 2145 2146 2147
                             batch_size_per_im=256,
                             fg_fraction=0.25,
                             fg_thresh=0.25,
                             bg_thresh_hi=0.5,
                             bg_thresh_lo=0.0,
                             bbox_reg_weights=[0.1, 0.1, 0.2, 0.2],
2148
                             class_nums=None,
2149 2150 2151
                             use_random=True,
                             is_cls_agnostic=False,
                             is_cascade_rcnn=False):
2152
    """
2153

2154
    ** Generate Proposal Labels of Faster-RCNN **
2155

B
buxingyuan 已提交
2156
    This operator can be, for given the GenerateProposalOp output bounding boxes and groundtruth,
B
buxingyuan 已提交
2157
    to sample foreground boxes and background boxes, and compute loss target.
B
buxingyuan 已提交
2158 2159 2160

    RpnRois is the output boxes of RPN and was processed by generate_proposal_op, these boxes
    were combined with groundtruth boxes and sampled according to batch_size_per_im and fg_fraction,
B
buxingyuan 已提交
2161
    If an instance with a groundtruth overlap greater than fg_thresh, then it was considered as a foreground sample.
B
buxingyuan 已提交
2162 2163
    If an instance with a groundtruth overlap greater than bg_thresh_lo and lower than bg_thresh_hi,
    then it was considered as a background sample.
B
buxingyuan 已提交
2164
    After all foreground and background boxes are chosen (so called Rois),
B
buxingyuan 已提交
2165
    then we apply random sampling to make sure
B
buxingyuan 已提交
2166
    the number of foreground boxes is no more than batch_size_per_im * fg_fraction.
B
buxingyuan 已提交
2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185

    For each box in Rois, we assign the classification (class label) and regression targets (box label) to it.
    Finally BboxInsideWeights and BboxOutsideWeights are used to specify whether it would contribute to training loss.

    Args:
        rpn_rois(Variable): A 2-D LoDTensor with shape [N, 4]. N is the number of the GenerateProposalOp's output, each element is a bounding box with [xmin, ymin, xmax, ymax] format.
        gt_classes(Variable): A 2-D LoDTensor with shape [M, 1]. M is the number of groundtruth, each element is a class label of groundtruth.
        is_crowd(Variable): A 2-D LoDTensor with shape [M, 1]. M is the number of groundtruth, each element is a flag indicates whether a groundtruth is crowd.
        gt_boxes(Variable): A 2-D LoDTensor with shape [M, 4]. M is the number of groundtruth, each element is a bounding box with [xmin, ymin, xmax, ymax] format.
        im_info(Variable): A 2-D LoDTensor with shape [B, 3]. B is the number of input images, each element consists of im_height, im_width, im_scale.

        batch_size_per_im(int): Batch size of rois per images.
        fg_fraction(float): Foreground fraction in total batch_size_per_im.
        fg_thresh(float): Overlap threshold which is used to chose foreground sample.
        bg_thresh_hi(float): Overlap threshold upper bound which is used to chose background sample.
        bg_thresh_lo(float): Overlap threshold lower bound which is used to chose background sample.
        bbox_reg_weights(list|tuple): Box regression weights.
        class_nums(int): Class number.
        use_random(bool): Use random sampling to choose foreground and background boxes.
2186 2187
        is_cls_agnostic(bool): bbox regression use class agnostic simply which only represent fg and bg boxes.
        is_cascade_rcnn(bool): it will filter some bbox crossing the image's boundary when setting True.
B
Bai Yifan 已提交
2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            rpn_rois = fluid.layers.data(name='rpn_rois', shape=[2, 4],
                           append_batch_size=False, dtype='float32')
            gt_classes = fluid.layers.data(name='gt_classes', shape=[8, 1],
                           append_batch_size=False, dtype='float32')
            is_crowd = fluid.layers.data(name='is_crowd', shape=[8, 1],
                           append_batch_size=False, dtype='float32')
            gt_boxes = fluid.layers.data(name='gt_boxes', shape=[8, 4],
                           append_batch_size=False, dtype='float32')
            im_info = fluid.layers.data(name='im_info', shape=[10, 3],
                           append_batch_size=False, dtype='float32')
            rois, labels_int32, bbox_targets, bbox_inside_weights,
            bbox_outside_weights = fluid.layers.generate_proposal_labels(
                           rpn_rois, gt_classes, is_crowd, gt_boxes, im_info,
                           class_nums=10)

2208 2209 2210 2211
    """

    helper = LayerHelper('generate_proposal_labels', **locals())

X
Xin Pan 已提交
2212 2213 2214 2215 2216 2217 2218 2219 2220
    rois = helper.create_variable_for_type_inference(dtype=rpn_rois.dtype)
    labels_int32 = helper.create_variable_for_type_inference(
        dtype=gt_classes.dtype)
    bbox_targets = helper.create_variable_for_type_inference(
        dtype=rpn_rois.dtype)
    bbox_inside_weights = helper.create_variable_for_type_inference(
        dtype=rpn_rois.dtype)
    bbox_outside_weights = helper.create_variable_for_type_inference(
        dtype=rpn_rois.dtype)
2221 2222 2223 2224 2225 2226

    helper.append_op(
        type="generate_proposal_labels",
        inputs={
            'RpnRois': rpn_rois,
            'GtClasses': gt_classes,
2227
            'IsCrowd': is_crowd,
2228
            'GtBoxes': gt_boxes,
2229
            'ImInfo': im_info
2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244
        },
        outputs={
            'Rois': rois,
            'LabelsInt32': labels_int32,
            'BboxTargets': bbox_targets,
            'BboxInsideWeights': bbox_inside_weights,
            'BboxOutsideWeights': bbox_outside_weights
        },
        attrs={
            'batch_size_per_im': batch_size_per_im,
            'fg_fraction': fg_fraction,
            'fg_thresh': fg_thresh,
            'bg_thresh_hi': bg_thresh_hi,
            'bg_thresh_lo': bg_thresh_lo,
            'bbox_reg_weights': bbox_reg_weights,
2245
            'class_nums': class_nums,
2246 2247 2248
            'use_random': use_random,
            'is_cls_agnostic': is_cls_agnostic,
            'is_cascade_rcnn': is_cascade_rcnn
2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259
        })

    rois.stop_gradient = True
    labels_int32.stop_gradient = True
    bbox_targets.stop_gradient = True
    bbox_inside_weights.stop_gradient = True
    bbox_outside_weights.stop_gradient = True

    return rois, labels_int32, bbox_targets, bbox_inside_weights, bbox_outside_weights


2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339
def generate_mask_labels(im_info, gt_classes, is_crowd, gt_segms, rois,
                         labels_int32, num_classes, resolution):
    """
    ** Generate Mask Labels for Mask-RCNN **

    This operator can be, for given the RoIs and corresponding labels,
    to sample foreground RoIs. This mask branch also has
    a :math: `K \\times M^{2}` dimensional output targets for each foreground
    RoI, which encodes K binary masks of resolution M x M, one for each of the
    K classes. This mask targets are used to compute loss of mask branch.

    Please note, the data format of groud-truth segmentation, assumed the
    segmentations are as follows. The first instance has two gt objects.
    The second instance has one gt object, this object has two gt segmentations.

        .. code-block:: python

            #[
            #  [[[229.14, 370.9, 229.14, 370.9, ...]],
            #   [[343.7, 139.85, 349.01, 138.46, ...]]], # 0-th instance
            #  [[[500.0, 390.62, ...],[115.48, 187.86, ...]]] # 1-th instance
            #]

            batch_masks = []
            for semgs in batch_semgs:
                gt_masks = []
                for semg in semgs:
                    gt_segm = []
                    for polys in semg:
                        gt_segm.append(np.array(polys).reshape(-1, 2))
                    gt_masks.append(gt_segm)
                batch_masks.append(gt_masks)
            
            
            place = fluid.CPUPlace()
            feeder = fluid.DataFeeder(place=place, feed_list=feeds)
            feeder.feed(batch_masks)

    Args:
        im_info(Variable): A 2-D Tensor with shape [N, 3]. N is the batch size,
            each element is [height, width, scale] of image. Image scale is
            target_size) / original_size.
        gt_classes(Variable): A 2-D LoDTensor with shape [M, 1]. M is the total
            number of ground-truth, each element is a class label.
        is_crowd(Variable): A 2-D LoDTensor with shape as gt_classes,
            each element is a flag indicating whether a groundtruth is crowd.
        gt_segms(Variable): This input is a 2D LoDTensor with shape [S, 2],
            it's LoD level is 3. Usually users do not needs to understand LoD,
            The users should return correct data format in reader.



            The LoD[0] represents the gt objects number of
            each instance. LoD[1] represents the segmentation counts of each
            objects. LoD[2] represents the polygons number of each segmentation.
            S the total number of polygons coordinate points. Each element is
            (x, y) coordinate points.
        rois(Variable): A 2-D LoDTensor with shape [R, 4]. R is the total
            number of RoIs, each element is a bounding box with
            (xmin, ymin, xmax, ymax) format in the range of original image.
        labels_int32(Variable): A 2-D LoDTensor in shape of [R, 1] with type
            of int32. R is the same as it in `rois`. Each element repersents
            a class label of a RoI.
        num_classes(int): Class number.
        resolution(int): Resolution of mask predictions.

    Returns:
        mask_rois (Variable):  A 2D LoDTensor with shape [P, 4]. P is the total
            number of sampled RoIs. Each element is a bounding box with
            [xmin, ymin, xmax, ymax] format in range of orignal image size.
        mask_rois_has_mask_int32 (Variable): A 2D LoDTensor with shape [P, 1],
            each element repersents the output mask RoI index with regard to
            to input RoIs.
        mask_int32 (Variable): A 2D LoDTensor with shape [P, K * M * M],
            K is the classes number and M is the resolution of mask predictions.
            Each element repersents the binary mask targets.

    Examples:
        .. code-block:: python

2340 2341
          import paddle.fluid as fluid

2342 2343 2344 2345 2346 2347 2348 2349
          im_info = fluid.layers.data(name="im_info", shape=[3],
              dtype="float32")
          gt_classes = fluid.layers.data(name="gt_classes", shape=[1],
              dtype="float32", lod_level=1)
          is_crowd = fluid.layers.data(name="is_crowd", shape=[1],
              dtype="float32", lod_level=1)
          gt_masks = fluid.layers.data(name="gt_masks", shape=[2],
              dtype="float32", lod_level=3)
2350
          # rois, roi_labels can be the output of
2351
          # fluid.layers.generate_proposal_labels.
2352 2353 2354 2355
          rois = fluid.layers.data(name="rois", shape=[4],
              dtype="float32", lod_level=1)
          roi_labels = fluid.layers.data(name="roi_labels", shape=[1],
              dtype="int32", lod_level=1)
2356 2357 2358 2359 2360 2361
          mask_rois, mask_index, mask_int32 = fluid.layers.generate_mask_labels(
              im_info=im_info,
              gt_classes=gt_classes,
              is_crowd=is_crowd,
              gt_segms=gt_masks,
              rois=rois,
2362
              labels_int32=roi_labels,
2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399
              num_classes=81,
              resolution=14)
    """

    helper = LayerHelper('generate_mask_labels', **locals())

    mask_rois = helper.create_variable_for_type_inference(dtype=rois.dtype)
    roi_has_mask_int32 = helper.create_variable_for_type_inference(
        dtype=gt_classes.dtype)
    mask_int32 = helper.create_variable_for_type_inference(
        dtype=gt_classes.dtype)

    helper.append_op(
        type="generate_mask_labels",
        inputs={
            'ImInfo': im_info,
            'GtClasses': gt_classes,
            'IsCrowd': is_crowd,
            'GtSegms': gt_segms,
            'Rois': rois,
            'LabelsInt32': labels_int32
        },
        outputs={
            'MaskRois': mask_rois,
            'RoiHasMaskInt32': roi_has_mask_int32,
            'MaskInt32': mask_int32
        },
        attrs={'num_classes': num_classes,
               'resolution': resolution})

    mask_rois.stop_gradient = True
    roi_has_mask_int32.stop_gradient = True
    mask_int32.stop_gradient = True

    return mask_rois, roi_has_mask_int32, mask_int32


2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411
def generate_proposals(scores,
                       bbox_deltas,
                       im_info,
                       anchors,
                       variances,
                       pre_nms_top_n=6000,
                       post_nms_top_n=1000,
                       nms_thresh=0.5,
                       min_size=0.1,
                       eta=1.0,
                       name=None):
    """
H
haowang101779990 已提交
2412 2413
    **Generate proposal Faster-RCNN**

2414 2415 2416 2417
    This operation proposes RoIs according to each box with their
    probability to be a foreground object and 
    the box can be calculated by anchors. Bbox_deltais and scores
    to be an object are the output of RPN. Final proposals
H
haowang101779990 已提交
2418 2419 2420 2421
    could be used to train detection net.

    For generating proposals, this operation performs following steps:

2422 2423
    1. Transposes and resizes scores and bbox_deltas in size of
       (H*W*A, 1) and (H*W*A, 4)
H
haowang101779990 已提交
2424 2425 2426 2427 2428 2429
    2. Calculate box locations as proposals candidates. 
    3. Clip boxes to image
    4. Remove predicted boxes with small area. 
    5. Apply NMS to get final proposals as output.

    Args:
2430 2431 2432 2433 2434 2435 2436 2437 2438
        scores(Variable): A 4-D Tensor with shape [N, A, H, W] represents
            the probability for each box to be an object.
            N is batch size, A is number of anchors, H and W are height and
            width of the feature map.
        bbox_deltas(Variable): A 4-D Tensor with shape [N, 4*A, H, W]
            represents the differece between predicted box locatoin and
            anchor location.
        im_info(Variable): A 2-D Tensor with shape [N, 3] represents origin
            image information for N batch. Info contains height, width and scale
H
haowang101779990 已提交
2439
            between origin image size and the size of feature map.
2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450
        anchors(Variable):   A 4-D Tensor represents the anchors with a layout
            of [H, W, A, 4]. H and W are height and width of the feature map,
            num_anchors is the box count of each position. Each anchor is
            in (xmin, ymin, xmax, ymax) format an unnormalized.
        variances(Variable): The expanded variances of anchors with a layout of
            [H, W, num_priors, 4]. Each variance is in
            (xcenter, ycenter, w, h) format.
        pre_nms_top_n(float): Number of total bboxes to be kept per
            image before NMS. 6000 by default.
        post_nms_top_n(float): Number of total bboxes to be kept per
            image after NMS. 1000 by default.
H
haowang101779990 已提交
2451
        nms_thresh(float): Threshold in NMS, 0.5 by default.
2452 2453 2454 2455
        min_size(float): Remove predicted boxes with either height or
            width < min_size. 0.1 by default.
        eta(float): Apply in adaptive NMS, if adaptive threshold > 0.5,
            adaptive_threshold = adaptive_threshold * eta in each iteration.
B
Bai Yifan 已提交
2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473

    Examples:
        .. code-block:: python
        
            import paddle.fluid as fluid
            scores = fluid.layers.data(name='scores', shape=[2, 4, 5, 5],
                         append_batch_size=False, dtype='float32')
            bbox_deltas = fluid.layers.data(name='bbox_deltas', shape=[2, 16, 5, 5],
                         append_batch_size=False, dtype='float32')
            im_info = fluid.layers.data(name='im_info', shape=[2, 3],
                         append_batch_size=False, dtype='float32')
            anchors = fluid.layers.data(name='anchors', shape=[5, 5, 4, 4],
                         append_batch_size=False, dtype='float32')
            variances = fluid.layers.data(name='variances', shape=[5, 5, 10, 4],
                         append_batch_size=False, dtype='float32')
            rois, roi_probs = fluid.layers.generate_proposals(scores, bbox_deltas,
                         im_info, anchors, variances)

2474 2475 2476
    """
    helper = LayerHelper('generate_proposals', **locals())

X
Xin Pan 已提交
2477 2478 2479 2480
    rpn_rois = helper.create_variable_for_type_inference(
        dtype=bbox_deltas.dtype)
    rpn_roi_probs = helper.create_variable_for_type_inference(
        dtype=scores.dtype)
2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502
    helper.append_op(
        type="generate_proposals",
        inputs={
            'Scores': scores,
            'BboxDeltas': bbox_deltas,
            'ImInfo': im_info,
            'Anchors': anchors,
            'Variances': variances
        },
        attrs={
            'pre_nms_topN': pre_nms_top_n,
            'post_nms_topN': post_nms_top_n,
            'nms_thresh': nms_thresh,
            'min_size': min_size,
            'eta': eta
        },
        outputs={'RpnRois': rpn_rois,
                 'RpnRoiProbs': rpn_roi_probs})
    rpn_rois.stop_gradient = True
    rpn_roi_probs.stop_gradient = True

    return rpn_rois, rpn_roi_probs
J
jerrywgz 已提交
2503 2504


J
jerrywgz 已提交
2505
def box_clip(input, im_info, name=None):
J
jerrywgz 已提交
2506 2507
    """
    Clip the box into the size given by im_info
J
jerrywgz 已提交
2508
    For each input box, The formula is given as follows:
2509 2510 2511
        
    .. code-block:: text

J
jerrywgz 已提交
2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522
        xmin = max(min(xmin, im_w - 1), 0)
        ymin = max(min(ymin, im_h - 1), 0) 
        xmax = max(min(xmax, im_w - 1), 0)
        ymax = max(min(ymax, im_h - 1), 0)
    
    where im_w and im_h are computed from im_info:
 
    .. code-block:: text

        im_h = round(height / scale)
        im_w = round(weight / scale)
J
jerrywgz 已提交
2523 2524

    Args:
J
jerrywgz 已提交
2525
        input(variable): The input box, the last dimension is 4.
2526 2527 2528 2529
        im_info(variable): The information of image with shape [N, 3] with 
                            layout (height, width, scale). height and width
                            is the input size and scale is the ratio of input
                            size and original size.
J
jerrywgz 已提交
2530 2531 2532 2533
        name (str): The name of this layer. It is optional.
    
    Returns:
        Variable: The cliped tensor variable.
2534
        
J
jerrywgz 已提交
2535 2536
    Examples:
        .. code-block:: python
2537
        
J
jerrywgz 已提交
2538
            boxes = fluid.layers.data(
J
jerrywgz 已提交
2539
                name='boxes', shape=[8, 4], dtype='float32', lod_level=1)
J
jerrywgz 已提交
2540 2541
            im_info = fluid.layers.data(name='im_info', shape=[3])
            out = fluid.layers.box_clip(
J
jerrywgz 已提交
2542
                input=boxes, im_info=im_info)
J
jerrywgz 已提交
2543 2544 2545
    """

    helper = LayerHelper("box_clip", **locals())
J
jerrywgz 已提交
2546
    output = helper.create_variable_for_type_inference(dtype=input.dtype)
2547
    inputs = {"Input": input, "ImInfo": im_info}
J
jerrywgz 已提交
2548
    helper.append_op(type="box_clip", inputs=inputs, outputs={"Output": output})
J
jerrywgz 已提交
2549

2550 2551
    return output

J
jerrywgz 已提交
2552

2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659
def retinanet_detection_output(bboxes,
                               scores,
                               anchors,
                               im_info,
                               score_threshold=0.05,
                               nms_top_k=1000,
                               keep_top_k=100,
                               nms_threshold=0.3,
                               nms_eta=1.):
    """
    **Detection Output Layer for Retinanet.**

    This operation is to get the detection results by performing following
    steps:

    1. Decode top-scoring bounding box predictions per FPN level according 
       to the anchor boxes.
    2. Merge top predictions from all levels and apply multi-class non 
       maximum suppression (NMS) on them to get the final detections.

    Args:
        bboxes(List): A list of tensors from multiple FPN levels. Each
            element is a 3-D Tensor with shape [N, Mi, 4] representing the
            predicted locations of Mi bounding boxes. N is the batch size,
            Mi is the number of bounding boxes from i-th FPN level and each 
            bounding box has four coordinate values and the layout is
            [xmin, ymin, xmax, ymax].
        scores(List): A list of tensors from multiple FPN levels. Each
            element is a 3-D Tensor with shape [N, Mi, C] representing the
            predicted confidence predictions. N is the batch size, C is the
            class number (excluding background), Mi is the number of bounding
            boxes from i-th FPN level. For each bounding box, there are total
            C scores.
        anchors(List): A 2-D Tensor with shape [Mi, 4] represents the locations
            of Mi anchor boxes from all FPN level. Each bounding box has four
            coordinate values and the layout is [xmin, ymin, xmax, ymax].
        im_info(Variable): A 2-D LoDTensor with shape [N, 3] represents the
            image information. N is the batch size, each image information
            includes height, width and scale.
        score_threshold(float): Threshold to filter out bounding boxes
            with a confidence score.
        nms_top_k(int): Maximum number of detections per FPN layer to be
            kept according to the confidences before NMS.
        keep_top_k(int): Number of total bounding boxes to be kept per image after
            NMS step. -1 means keeping all bounding boxes after NMS step.
        nms_threshold(float): The threshold to be used in NMS.
        nms_eta(float): The parameter for adaptive NMS.

    Returns:
        Variable:
            The detection output is a LoDTensor with shape [No, 6].
            Each row has six values: [label, confidence, xmin, ymin, xmax, ymax].
            `No` is the total number of detections in this mini-batch. For each
            instance, the offsets in first dimension are called LoD, the offset
            number is N + 1, N is the batch size. The i-th image has
            `LoD[i + 1] - LoD[i]` detected results, if it is 0, the i-th image
            has no detected results. If all images have no detected results,
            LoD will be set to 0, and the output tensor is empty (None).

    Examples:
        .. code-block:: python
        
            import paddle.fluid as fluid

            bboxes = layers.data(name='bboxes', shape=[1, 21, 4],
                append_batch_size=False, dtype='float32')
            scores = layers.data(name='scores', shape=[1, 21, 10],
                append_batch_size=False, dtype='float32')
            anchors = layers.data(name='anchors', shape=[21, 4],
                append_batch_size=False, dtype='float32')
            im_info = layers.data(name="im_info", shape=[1, 3],
                append_batch_size=False, dtype='float32')
            nmsed_outs = fluid.layers.retinanet_detection_output(
                                                    bboxes=[bboxes, bboxes],
                                                    scores=[scores, scores],
                                                    anchors=[anchors, anchors],
                                                    im_info=im_info,
                                                    score_threshold=0.05,
                                                    nms_top_k=1000,
                                                    keep_top_k=100,
                                                    nms_threshold=0.3,
                                                    nms_eta=1.)
    """

    helper = LayerHelper('retinanet_detection_output', **locals())
    output = helper.create_variable_for_type_inference(
        dtype=helper.input_dtype('scores'))
    helper.append_op(
        type="retinanet_detection_output",
        inputs={
            'BBoxes': bboxes,
            'Scores': scores,
            'Anchors': anchors,
            'ImInfo': im_info
        },
        attrs={
            'score_threshold': score_threshold,
            'nms_top_k': nms_top_k,
            'nms_threshold': nms_threshold,
            'keep_top_k': keep_top_k,
            'nms_eta': 1.,
        },
        outputs={'Out': output})
    output.stop_gradient = True
    return output


J
jerrywgz 已提交
2660 2661 2662 2663 2664
def multiclass_nms(bboxes,
                   scores,
                   score_threshold,
                   nms_top_k,
                   keep_top_k,
J
jerrywgz 已提交
2665
                   nms_threshold=0.3,
J
jerrywgz 已提交
2666 2667
                   normalized=True,
                   nms_eta=1.,
2668 2669
                   background_label=0,
                   name=None):
J
jerrywgz 已提交
2670
    """
2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731
    **Multiclass NMS**
    
    This operator is to do multi-class non maximum suppression (NMS) on
    boxes and scores.

    In the NMS step, this operator greedily selects a subset of detection bounding
    boxes that have high scores larger than score_threshold, if providing this
    threshold, then selects the largest nms_top_k confidences scores if nms_top_k
    is larger than -1. Then this operator pruns away boxes that have high IOU
    (intersection over union) overlap with already selected boxes by adaptive
    threshold NMS based on parameters of nms_threshold and nms_eta.

    Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
    per image if keep_top_k is larger than -1.

    Args:
        bboxes (Variable): Two types of bboxes are supported:
                           1. (Tensor) A 3-D Tensor with shape
                           [N, M, 4 or 8 16 24 32] represents the
                           predicted locations of M bounding bboxes,
                           N is the batch size. Each bounding box has four
                           coordinate values and the layout is 
                           [xmin, ymin, xmax, ymax], when box size equals to 4.
                           2. (LoDTensor) A 3-D Tensor with shape [M, C, 4]
                           M is the number of bounding boxes, C is the 
                           class number   
        scores (Variable): Two types of scores are supported:
                           1. (Tensor) A 3-D Tensor with shape [N, C, M]
                           represents the predicted confidence predictions.
                           N is the batch size, C is the class number, M is 
                           number of bounding boxes. For each category there 
                           are total M scores which corresponding M bounding
                           boxes. Please note, M is equal to the 2nd dimension
                           of BBoxes.
                           2. (LoDTensor) A 2-D LoDTensor with shape [M, C].
                           M is the number of bbox, C is the class number.
                           In this case, input BBoxes should be the second
                           case with shape [M, C, 4].
        background_label (int): The index of background label, the background 
                                label will be ignored. If set to -1, then all
                                categories will be considered. Default: 0
        score_threshold (float): Threshold to filter out bounding boxes with
                                 low confidence score. If not provided, 
                                 consider all boxes.
        nms_top_k (int): Maximum number of detections to be kept according to
                         the confidences aftern the filtering detections based
                         on score_threshold.
        nms_threshold (float): The threshold to be used in NMS. Default: 0.3
        nms_eta (float): The threshold to be used in NMS. Default: 1.0
        keep_top_k (int): Number of total bboxes to be kept per image after NMS
                          step. -1 means keeping all bboxes after NMS step.
        normalized (bool): Whether detections are normalized. Default: True
        name(str): Name of the multiclass nms op. Default: None.

    Returns:
        Out: A 2-D LoDTensor with shape [No, 6] represents the detections.
             Each row has 6 values: [label, confidence, xmin, ymin, xmax, ymax]
             or A 2-D LoDTensor with shape [No, 10] represents the detections.
             Each row has 10 values: 
             [label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the 
             total number of detections. If there is no detected boxes for all
J
jerrywgz 已提交
2732 2733 2734 2735
             images, lod will be set to {1} and Out only contains one value
             which is -1.
             (After version 1.3, when no boxes detected, the lod is changed 
             from {0} to {1}) 
2736

2737

2738 2739 2740
    Examples:
        .. code-block:: python

2741

2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753
            boxes = fluid.layers.data(name='bboxes', shape=[81, 4],
                                      dtype='float32', lod_level=1)
            scores = fluid.layers.data(name='scores', shape=[81],
                                      dtype='float32', lod_level=1)
            out = fluid.layers.multiclass_nms(bboxes=boxes,
                                              scores=scores,
                                              background_label=0,
                                              score_threshold=0.5,
                                              nms_top_k=400,
                                              nms_threshold=0.3,
                                              keep_top_k=200,
                                              normalized=False)
J
jerrywgz 已提交
2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773
    """
    helper = LayerHelper('multiclass_nms', **locals())

    output = helper.create_variable_for_type_inference(dtype=bboxes.dtype)
    helper.append_op(
        type="multiclass_nms",
        inputs={'BBoxes': bboxes,
                'Scores': scores},
        attrs={
            'background_label': background_label,
            'score_threshold': score_threshold,
            'nms_top_k': nms_top_k,
            'nms_threshold': nms_threshold,
            'nms_eta': nms_eta,
            'keep_top_k': keep_top_k,
            'nms_eta': nms_eta,
            'normalized': normalized
        },
        outputs={'Out': output})
    output.stop_gradient = True
J
jerrywgz 已提交
2774 2775

    return output
2776 2777 2778 2779 2780 2781 2782 2783 2784


def distribute_fpn_proposals(fpn_rois,
                             min_level,
                             max_level,
                             refer_level,
                             refer_scale,
                             name=None):
    """
J
jerrywgz 已提交
2785 2786 2787 2788 2789 2790
    In Feature Pyramid Networks (FPN) models, it is needed to distribute all 
    proposals into different FPN level, with respect to scale of the proposals,
    the referring scale and the referring level. Besides, to restore the order
    of proposals, we return an array which indicates the original index of rois
    in current proposals. To compute FPN level for each roi, the formula is 
    given as follows:
2791
    
J
jerrywgz 已提交
2792
    .. math::
2793

J
jerrywgz 已提交
2794
        roi\_scale &= \sqrt{BBoxArea(fpn\_roi)}
2795

J
jerrywgz 已提交
2796 2797 2798
        level = floor(&\log(\\frac{roi\_scale}{refer\_scale}) + refer\_level)

    where BBoxArea is a function to compute the area of each roi.
2799 2800

    Args:
J
jerrywgz 已提交
2801
        fpn_rois(variable): The input fpn_rois, the second dimension is 4.
2802 2803 2804 2805 2806 2807
        min_level(int): The lowest level of FPN layer where the proposals come 
                        from.
        max_level(int): The highest level of FPN layer where the proposals
                        come from.
        refer_level(int): The referring level of FPN layer with specified scale.
        refer_scale(int): The referring scale of FPN layer with specified level.
J
jerrywgz 已提交
2808 2809
        name(str|None): The name of this operator.        

2810
    Returns:
J
jerrywgz 已提交
2811 2812 2813 2814 2815
        tuple: 
               A tuple(multi_rois, restore_ind) is returned. The multi_rois is 
               a list of segmented tensor variables. The restore_ind is a 2D 
               Tensor with shape [N, 1], N is the number of total rois. It is
               used to restore the order of fpn_rois.
2816 2817 2818 2819 2820 2821 2822

    Examples:
        .. code-block:: python

            fpn_rois = fluid.layers.data(
                name='data', shape=[4], dtype='float32', lod_level=1)
            multi_rois, restore_ind = fluid.layers.distribute_fpn_proposals(
2823 2824 2825
                fpn_rois=fpn_rois,
                min_level=2,
                max_level=5,
2826 2827 2828 2829 2830
                refer_level=4,
                refer_scale=224)
    """

    helper = LayerHelper('distribute_fpn_proposals', **locals())
2831
    dtype = helper.input_dtype('fpn_rois')
2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848
    num_lvl = max_level - min_level + 1
    multi_rois = [
        helper.create_variable_for_type_inference(dtype) for i in range(num_lvl)
    ]
    restore_ind = helper.create_variable_for_type_inference(dtype='int32')
    helper.append_op(
        type='distribute_fpn_proposals',
        inputs={'FpnRois': fpn_rois},
        outputs={'MultiFpnRois': multi_rois,
                 'RestoreIndex': restore_ind},
        attrs={
            'min_level': min_level,
            'max_level': max_level,
            'refer_level': refer_level,
            'refer_scale': refer_scale
        })
    return multi_rois, restore_ind
2849 2850


2851
@templatedoc()
J
jerrywgz 已提交
2852 2853 2854 2855 2856 2857
def box_decoder_and_assign(prior_box,
                           prior_box_var,
                           target_box,
                           box_score,
                           box_clip,
                           name=None):
2858 2859 2860 2861 2862 2863 2864
    """
    ${comment}
    Args:
        prior_box(${prior_box_type}): ${prior_box_comment}
        prior_box_var(${prior_box_var_type}): ${prior_box_var_comment}
        target_box(${target_box_type}): ${target_box_comment}
        box_score(${box_score_type}): ${box_score_comment}
J
jerrywgz 已提交
2865
        box_clip(${box_clip_type}): ${box_clip_comment}
J
jerrywgz 已提交
2866
        name(str|None): The name of this operator
2867
    Returns:
J
jerrywgz 已提交
2868 2869 2870 2871 2872 2873 2874
        decode_box(Variable), output_assign_box(Variable):

            two variables:

            - decode_box(${decode_box_type}): ${decode_box_comment}
            - output_assign_box(${output_assign_box_type}): ${output_assign_box_comment}

2875 2876 2877
    Examples:
        .. code-block:: python

J
jerrywgz 已提交
2878
            pb = fluid.layers.data(
J
jerrywgz 已提交
2879
                name='prior_box', shape=[4], dtype='float32')
J
jerrywgz 已提交
2880
            pbv = fluid.layers.data(
J
jerrywgz 已提交
2881 2882
                name='prior_box_var', shape=[4], 
                dtype='float32', append_batch_size=False)
J
jerrywgz 已提交
2883
            loc = fluid.layers.data(
J
jerrywgz 已提交
2884
                name='target_box', shape=[4*81], dtype='float32')
J
jerrywgz 已提交
2885
            scores = fluid.layers.data(
J
jerrywgz 已提交
2886
                name='scores', shape=[81], dtype='float32')
J
jerrywgz 已提交
2887
            decoded_box, output_assign_box = fluid.layers.box_decoder_and_assign(
J
jerrywgz 已提交
2888
                pb, pbv, loc, scores, 4.135)
2889 2890 2891 2892

    """
    helper = LayerHelper("box_decoder_and_assign", **locals())

J
jerrywgz 已提交
2893
    decoded_box = helper.create_variable_for_type_inference(
2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907
        dtype=prior_box.dtype)
    output_assign_box = helper.create_variable_for_type_inference(
        dtype=prior_box.dtype)

    helper.append_op(
        type="box_decoder_and_assign",
        inputs={
            "PriorBox": prior_box,
            "PriorBoxVar": prior_box_var,
            "TargetBox": target_box,
            "BoxScore": box_score
        },
        attrs={"box_clip": box_clip},
        outputs={
J
jerrywgz 已提交
2908
            "DecodeBox": decoded_box,
2909 2910
            "OutputAssignBox": output_assign_box
        })
J
jerrywgz 已提交
2911
    return decoded_box, output_assign_box
2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976


def collect_fpn_proposals(multi_rois,
                          multi_scores,
                          min_level,
                          max_level,
                          post_nms_top_n,
                          name=None):
    """
    Concat multi-level RoIs (Region of Interest) and select N RoIs 
    with respect to multi_scores. This operation performs the following steps:

    1. Choose num_level RoIs and scores as input: num_level = max_level - min_level
    2. Concat multi-level RoIs and scores
    3. Sort scores and select post_nms_top_n scores
    4. Gather RoIs by selected indices from scores
    5. Re-sort RoIs by corresponding batch_id

    Args:
        multi_ros(list): List of RoIs to collect
        multi_scores(list): List of scores
        min_level(int): The lowest level of FPN layer to collect
        max_level(int): The highest level of FPN layer to collect
        post_nms_top_n(int): The number of selected RoIs
        name(str|None): A name for this layer(optional)
        
    Returns:
        Variable: Output variable of selected RoIs. 

    Examples:
        .. code-block:: python
           
            multi_rois = []
            multi_scores = []
            for i in range(4):
                multi_rois.append(fluid.layers.data(
                    name='roi_'+str(i), shape=[4], dtype='float32', lod_level=1))
            for i in range(4):
                multi_scores.append(fluid.layers.data(
                    name='score_'+str(i), shape=[1], dtype='float32', lod_level=1))

            fpn_rois = fluid.layers.collect_fpn_proposals(
                multi_rois=multi_rois, 
                multi_scores=multi_scores,
                min_level=2, 
                max_level=5, 
                post_nms_top_n=2000)
    """

    helper = LayerHelper('collect_fpn_proposals', **locals())
    dtype = helper.input_dtype('multi_rois')
    num_lvl = max_level - min_level + 1
    input_rois = multi_rois[:num_lvl]
    input_scores = multi_scores[:num_lvl]
    output_rois = helper.create_variable_for_type_inference(dtype)
    output_rois.stop_gradient = True
    helper.append_op(
        type='collect_fpn_proposals',
        inputs={
            'MultiLevelRois': input_rois,
            'MultiLevelScores': input_scores
        },
        outputs={'FpnRois': output_rois},
        attrs={'post_nms_topN': post_nms_top_n})
    return output_rois