detection.py 85.2 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
    'prior_box',
R
ruri 已提交
35
    'density_prior_box',
C
chengduoZH 已提交
36
    'multi_box_head',
37 38 39 40
    'bipartite_match',
    'target_assign',
    'detection_output',
    'ssd_loss',
41
    'detection_map',
Y
Yuan Gao 已提交
42
    'rpn_target_assign',
43
    'anchor_generator',
W
whs 已提交
44
    'roi_perspective_transform',
45
    'generate_proposal_labels',
46
    'generate_proposals',
47
    'generate_mask_labels',
48 49
    'iou_similarity',
    'box_coder',
B
Bai Yifan 已提交
50
    'polygon_box_transform',
D
dengkaipeng 已提交
51
    'yolov3_loss',
J
jerrywgz 已提交
52
    'multiclass_nms',
C
chengduoZH 已提交
53
]
54 55


56 57
def rpn_target_assign(bbox_pred,
                      cls_logits,
Y
Yuan Gao 已提交
58
                      anchor_box,
59
                      anchor_var,
60 61 62
                      gt_boxes,
                      is_crowd,
                      im_info,
Y
Yuan Gao 已提交
63
                      rpn_batch_size_per_im=256,
64 65
                      rpn_straddle_thresh=0.0,
                      rpn_fg_fraction=0.5,
Y
Yuan Gao 已提交
66
                      rpn_positive_overlap=0.7,
67 68
                      rpn_negative_overlap=0.3,
                      use_random=True):
Y
Yuan Gao 已提交
69
    """
H
haowang101779990 已提交
70
    **Target Assign Layer for region proposal network (RPN) in Faster-RCNN detection.**
Y
Yuan Gao 已提交
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87

    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:
88
        bbox_pred(Variable): A 3-D Tensor with shape [N, M, 4] represents the
Y
Yuan Gao 已提交
89 90 91
            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].
92 93 94
        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
95 96 97 98 99 100
        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.
101 102
        anchor_var(Variable): A 2-D Tensor with shape [M,4] holds expanded 
            variances of anchors.
103
        gt_boxes (Variable): The ground-truth boudding boxes (bboxes) are a 2D
Y
Yuan Gao 已提交
104 105
            LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth
            bboxes of mini-batch input.
106 107 108
        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
109
        rpn_batch_size_per_im(int): Total number of RPN examples per image.
110 111 112
        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
113 114 115 116 117 118 119 120 121
            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
122
        tuple:
Y
Yuan Gao 已提交
123
               A tuple(predicted_scores, predicted_location, target_label,
J
jerrywgz 已提交
124 125
               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
126 127 128 129 130 131 132
               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
133
               anchors, the F and B is depends on the input of this operator.
J
jerrywgz 已提交
134 135
               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
136 137 138 139

    Examples:
        .. code-block:: python

H
haowang101779990 已提交
140 141 142 143 144 145 146 147 148 149 150 151 152 153
            bbox_pred = layers.data(name='bbox_pred', shape=[100, 4],
                              append_batch_size=False, dtype='float32')
            cls_logits = layers.data(name='cls_logits', shape=[100, 1],
                              append_batch_size=False, dtype='float32')
            anchor_box = layers.data(name='anchor_box', shape=[20, 4],
                              append_batch_size=False, dtype='float32')
            gt_boxes = layers.data(name='gt_boxes', shape=[10, 4],
                             append_batch_size=False, dtype='float32')
            loc_pred, score_pred, loc_target, score_target, bbox_inside_weight =
                fluid.layers.rpn_target_assign(bbox_pred=bbox_pred,
                                              cls_logits=cls_logits,
                                              anchor_box=anchor_box,
                                              gt_boxes=gt_boxes)

Y
Yuan Gao 已提交
154 155 156
    """

    helper = LayerHelper('rpn_target_assign', **locals())
157
    # Assign target label to anchors
J
jerrywgz 已提交
158 159 160 161 162 163 164
    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
165 166
    helper.append_op(
        type="rpn_target_assign",
167 168 169 170 171 172
        inputs={
            'Anchor': anchor_box,
            'GtBoxes': gt_boxes,
            'IsCrowd': is_crowd,
            'ImInfo': im_info
        },
Y
Yuan Gao 已提交
173 174 175
        outputs={
            'LocationIndex': loc_index,
            'ScoreIndex': score_index,
176
            'TargetLabel': target_label,
J
jerrywgz 已提交
177
            'TargetBBox': target_bbox,
J
jerrywgz 已提交
178
            'BBoxInsideWeight': bbox_inside_weight
Y
Yuan Gao 已提交
179 180 181
        },
        attrs={
            'rpn_batch_size_per_im': rpn_batch_size_per_im,
182
            'rpn_straddle_thresh': rpn_straddle_thresh,
Y
Yuan Gao 已提交
183 184
            'rpn_positive_overlap': rpn_positive_overlap,
            'rpn_negative_overlap': rpn_negative_overlap,
185 186
            'rpn_fg_fraction': rpn_fg_fraction,
            'use_random': use_random
Y
Yuan Gao 已提交
187 188
        })

189 190 191 192
    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
193
    bbox_inside_weight.stop_gradient = True
Y
Yuan Gao 已提交
194

195 196 197 198
    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)
199

J
jerrywgz 已提交
200
    return predicted_cls_logits, predicted_bbox_pred, target_label, target_bbox, bbox_inside_weight
Y
Yuan Gao 已提交
201 202


Y
Yuan Gao 已提交
203 204
def detection_output(loc,
                     scores,
205 206 207 208 209 210 211 212 213
                     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):
    """
214
    **Detection Output Layer for Single Shot Multibox Detector (SSD).**
215

216 217
    This operation is to get the detection results by performing following
    two steps:
C
caoying03 已提交
218

219 220 221 222 223 224
    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.
225 226 227 228 229 230

    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
231 232 233 234
        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.
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
        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
257 258
        Variable:

259
            The detection outputs is a LoDTensor with shape [No, 6].
260 261 262 263 264 265
            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
266
            LoD will be set to {1}, and output tensor only contains one
267
            value, which is -1.
J
jerrywgz 已提交
268 269
            (After version 1.3, when no boxes detected, the lod is changed
             from {0} to {1}.)
270 271 272 273

    Examples:
        .. code-block:: python

274
            pb = layers.data(name='prior_box', shape=[10, 4],
275
                         append_batch_size=False, dtype='float32')
276
            pbv = layers.data(name='prior_box_var', shape=[10, 4],
277
                          append_batch_size=False, dtype='float32')
278
            loc = layers.data(name='target_box', shape=[2, 21, 4],
279
                          append_batch_size=False, dtype='float32')
280
            scores = layers.data(name='scores', shape=[2, 21, 10],
281
                          append_batch_size=False, dtype='float32')
282
            nmsed_outs = fluid.layers.detection_output(scores=scores,
283 284 285 286 287
                                       loc=loc,
                                       prior_box=pb,
                                       prior_box_var=pbv)
    """
    helper = LayerHelper("detection_output", **locals())
288 289 290 291 292
    decoded_box = box_coder(
        prior_box=prior_box,
        prior_box_var=prior_box_var,
        target_box=loc,
        code_type='decode_center_size')
293
    scores = nn.softmax(input=scores)
Y
Yuan Gao 已提交
294
    scores = nn.transpose(scores, perm=[0, 2, 1])
295
    scores.stop_gradient = True
X
Xin Pan 已提交
296 297
    nmsed_outs = helper.create_variable_for_type_inference(
        dtype=decoded_box.dtype)
298 299 300 301 302 303 304 305 306 307 308 309 310
    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
        })
311
    nmsed_outs.stop_gradient = True
312
    return nmsed_outs
C
chengduoZH 已提交
313 314


X
Xin Pan 已提交
315 316 317 318 319 320 321 322 323 324 325 326 327 328
@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}
    """
    helper = LayerHelper("iou_similarity", **locals())
    if name is None:
X
Xin Pan 已提交
329
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
X
Xin Pan 已提交
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365
    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,
              name=None):
    """
    ${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}): ${target_box_comment}
        code_type(${code_type_type}): ${code_type_comment}
        box_normalized(${box_normalized_type}): ${box_normalized_comment}

    Returns:
        output_box(${output_box_type}): ${output_box_comment}
    """
    helper = LayerHelper("box_coder", **locals())

    if name is None:
X
Xin Pan 已提交
366 367
        output_box = helper.create_variable_for_type_inference(
            dtype=prior_box.dtype)
X
Xin Pan 已提交
368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
    else:
        output_box = helper.create_variable(
            name=name, dtype=prior_box.dtype, persistable=False)

    helper.append_op(
        type="box_coder",
        inputs={
            "PriorBox": prior_box,
            "PriorBoxVar": prior_box_var,
            "TargetBox": target_box
        },
        attrs={"code_type": code_type,
               "box_normalized": box_normalized},
        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}
    """
    helper = LayerHelper("polygon_box_transform", **locals())
    if name is None:
X
Xin Pan 已提交
398
        output = helper.create_variable_for_type_inference(dtype=input.dtype)
X
Xin Pan 已提交
399 400 401 402 403 404 405 406 407 408 409 410
    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
411 412 413
@templatedoc(op_type="yolov3_loss")
def yolov3_loss(x,
                gtbox,
D
dengkaipeng 已提交
414
                gtlabel,
D
dengkaipeng 已提交
415
                anchors,
416
                anchor_mask,
D
dengkaipeng 已提交
417 418
                class_num,
                ignore_thresh,
419
                downsample_ratio,
D
dengkaipeng 已提交
420 421 422 423 424 425
                name=None):
    """
    ${comment}

    Args:
        x (Variable): ${x_comment}
D
dengkaipeng 已提交
426 427 428 429 430
        gtbox (Variable): groud truth boxes, should be in shape of [N, B, 4],
                          in the third dimenstion, x, y, w, h should be stored 
                          and x, y, w, h should be relative value of input image.
                          N is the batch number and B is the max box number in 
                          an image.
D
dengkaipeng 已提交
431
        gtlabel (Variable): class id of ground truth boxes, shoud be in shape
D
dengkaipeng 已提交
432
                            of [N, B].
D
dengkaipeng 已提交
433
        anchors (list|tuple): ${anchors_comment}
434
        anchor_mask (list|tuple): ${anchor_mask_comment}
D
dengkaipeng 已提交
435 436
        class_num (int): ${class_num_comment}
        ignore_thresh (float): ${ignore_thresh_comment}
437
        downsample_ratio (int): ${downsample_ratio_comment}
D
dengkaipeng 已提交
438 439 440 441 442 443 444 445
        name (string): the name of yolov3 loss

    Returns:
        Variable: A 1-D tensor with shape [1], the value of yolov3 loss

    Raises:
        TypeError: Input x of yolov3_loss must be Variable
        TypeError: Input gtbox of yolov3_loss must be Variable"
D
dengkaipeng 已提交
446
        TypeError: Input gtlabel of yolov3_loss must be Variable"
D
dengkaipeng 已提交
447 448 449 450 451 452 453
        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

    Examples:
    .. code-block:: python

D
dengkaipeng 已提交
454 455 456
        x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32')
        gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32')
        gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32')
457 458
        anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]
        anchors = [0, 1, 2]
459 460
        loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80, anchors=anchors, 
                                        ignore_thresh=0.5, downsample_ratio=32)
D
dengkaipeng 已提交
461 462 463 464 465 466 467
    """
    helper = LayerHelper('yolov3_loss', **locals())

    if not isinstance(x, Variable):
        raise TypeError("Input x of yolov3_loss must be Variable")
    if not isinstance(gtbox, Variable):
        raise TypeError("Input gtbox of yolov3_loss must be Variable")
D
dengkaipeng 已提交
468 469
    if not isinstance(gtlabel, Variable):
        raise TypeError("Input gtlabel of yolov3_loss must be Variable")
D
dengkaipeng 已提交
470 471
    if not isinstance(anchors, list) and not isinstance(anchors, tuple):
        raise TypeError("Attr anchors of yolov3_loss must be list or tuple")
472 473
    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
474 475 476 477 478 479 480 481 482 483 484 485
    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")

    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)

486 487 488
    objectness_mask = helper.create_variable_for_type_inference(dtype='int32')
    gt_match_mask = helper.create_variable_for_type_inference(dtype='int32')

D
dengkaipeng 已提交
489 490
    attrs = {
        "anchors": anchors,
491
        "anchor_mask": anchor_mask,
D
dengkaipeng 已提交
492 493
        "class_num": class_num,
        "ignore_thresh": ignore_thresh,
494
        "downsample_ratio": downsample_ratio,
D
dengkaipeng 已提交
495 496 497 498
    }

    helper.append_op(
        type='yolov3_loss',
D
dengkaipeng 已提交
499 500 501 502 503
        inputs={
            "X": x,
            "GTBox": gtbox,
            "GTLabel": gtlabel,
        },
504 505 506 507 508
        outputs={
            'Loss': loss,
            'ObjectnessMask': objectness_mask,
            'GTMatchMask': gt_match_mask
        },
D
dengkaipeng 已提交
509 510 511 512
        attrs=attrs)
    return loss


X
Xin Pan 已提交
513
@templatedoc()
514 515
def detection_map(detect_res,
                  label,
516 517
                  class_num,
                  background_label=0,
518 519
                  overlap_threshold=0.3,
                  evaluate_difficult=True,
520 521 522 523
                  has_state=None,
                  input_states=None,
                  out_states=None,
                  ap_version='integral'):
X
Xin Pan 已提交
524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564
    """
    ${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

            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')

            map_out = fluid.layers.detection_map(detect_res, label, 21)
    """
565 566
    helper = LayerHelper("detection_map", **locals())

567
    def __create_var(type):
X
Xin Pan 已提交
568
        return helper.create_variable_for_type_inference(dtype=type)
569 570 571 572 573 574 575 576 577 578 579 580

    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

581 582 583 584 585
    helper.append_op(
        type="detection_map",
        inputs={
            'Label': label,
            'DetectRes': detect_res,
586
            'HasState': has_state,
587 588 589 590 591 592 593 594 595 596 597 598 599
            '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,
600 601
            'ap_type': ap_version,
            'class_num': class_num,
602
        })
603
    return map_out
604 605


606 607 608 609
def bipartite_match(dist_matrix,
                    match_type=None,
                    dist_threshold=None,
                    name=None):
610
    """
Y
yuyang18 已提交
611 612
    This operator implements a greedy bipartite matching algorithm, which is
    used to obtain the matching with the maximum distance based on the input
613
    distance matrix. For input 2D matrix, the bipartite matching algorithm can
Y
yuyang18 已提交
614 615 616 617 618 619 620 621
    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)
622 623 624
    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
625

Y
yuyang18 已提交
626
    NOTE: the input DistMat can be LoDTensor (with LoD) or Tensor.
627 628 629
    If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size.
    If Tensor, the height of ColToRowMatchIndices is 1.

Y
yuyang18 已提交
630 631 632
    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.

633 634 635 636 637
    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
638 639 640 641 642 643
            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.
644
        match_type(string|None): The type of matching method, should be
Y
yuyang18 已提交
645
           'bipartite' or 'per_prediction'. [default 'bipartite'].
646 647
        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
648
            on the maximum distance, 0.5 by default.
649
    Returns:
Y
yuyang18 已提交
650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
        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)
673 674
    """
    helper = LayerHelper('bipartite_match', **locals())
X
Xin Pan 已提交
675 676 677
    match_indices = helper.create_variable_for_type_inference(dtype='int32')
    match_distance = helper.create_variable_for_type_inference(
        dtype=dist_matrix.dtype)
678 679 680
    helper.append_op(
        type='bipartite_match',
        inputs={'DistMat': dist_matrix},
681 682 683 684
        attrs={
            'match_type': match_type,
            'dist_threshold': dist_threshold,
        },
685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701
        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
702

703 704 705 706 707
    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
708

709
    1. Assigning all outpts based on `match_indices`:
C
chengduoZH 已提交
710

711 712 713
    .. code-block:: text

        If id = match_indices[i][j] > 0,
C
chengduoZH 已提交
714

715 716
            out[i][j][0 : K] = X[lod[i] + id][j % P][0 : K]
            out_weight[i][j] = 1.
C
chengduoZH 已提交
717

718
        Otherwise,
C
chengduoZH 已提交
719

720 721
            out[j][j][0 : K] = {mismatch_value, mismatch_value, ...}
            out_weight[i][j] = 0.
C
chengduoZH 已提交
722

723
    2. Assigning out_weight based on `neg_indices` if `neg_indices` is provided:
C
chengduoZH 已提交
724

725 726
    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
727

728
    .. code-block:: text
C
chengduoZH 已提交
729

730 731 732 733 734 735 736 737 738 739 740 741 742 743 744
        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
745 746 747 748 749
        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
750 751 752 753 754 755 756 757 758 759 760
               the shape of [N, P, 1].

    Examples:

        .. code-block:: python

            matched_indices, matched_dist = fluid.layers.bipartite_match(iou)
            gt = layers.data(
                        name='gt', shape=[1, 1], dtype='int32', lod_level=1)
            trg, trg_weight = layers.target_assign(
                            gt, matched_indices, mismatch_value=0)
761 762
    """
    helper = LayerHelper('target_assign', **locals())
X
Xin Pan 已提交
763 764
    out = helper.create_variable_for_type_inference(dtype=input.dtype)
    out_weight = helper.create_variable_for_type_inference(dtype='float32')
765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791
    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',
792
             normalize=True,
793 794
             sample_size=None):
    """
Y
yuyang18 已提交
795
    **Multi-box loss layer for object detection algorithm of SSD**
796 797 798 799 800 801 802

    This layer is to compute dection loss for SSD given the location offset
    predictions, confidence predictions, prior boxes and ground-truth boudding
    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
803
    1. Find matched bounding box by bipartite matching algorithm.
Y
yuyang18 已提交
804

805
      1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
Y
yuyang18 已提交
806

807
      1.2 Compute matched boundding box by bipartite matching algorithm.
Y
yuyang18 已提交
808

809
    2. Compute confidence for mining hard examples
Y
yuyang18 已提交
810

811
      2.1. Get the target label based on matched indices.
Y
yuyang18 已提交
812

813
      2.2. Compute confidence loss.
Y
yuyang18 已提交
814

815 816
    3. Apply hard example mining to get the negative example indices and update
       the matched indices.
Y
yuyang18 已提交
817

818
    4. Assign classification and regression targets
Y
yuyang18 已提交
819

820
      4.1. Encoded bbox according to the prior boxes.
Y
yuyang18 已提交
821

822
      4.2. Assign regression targets.
Y
yuyang18 已提交
823

824
      4.3. Assign classification targets.
Y
yuyang18 已提交
825

826
    5. Compute the overall objective loss.
Y
yuyang18 已提交
827

828
      5.1 Compute confidence loss.
Y
yuyang18 已提交
829

830
      5.1 Compute localization loss.
Y
yuyang18 已提交
831

832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854
      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.
        gt_box (Variable): The ground-truth boudding boxes (bboxes) are a 2D
            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
855
            boxes, used only when mining_type is 'max_negative', 3.0 by defalut.
856
        neg_overlap (float): The negative overlap upper bound for the unmatched
857
            predictions. Use only when mining_type is 'max_negative',
858 859 860 861
            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
862
            be 'bipartite' or 'per_prediction', 'per_prediction' by defalut.
863 864
        mining_type (str): The hard example mining type, should be 'hard_example'
            or 'max_negative', now only support `max_negative`.
865
        normalize (bool): Whether to normalize the SSD loss by the total number
Y
yuyang18 已提交
866
            of output locations, True by default.
867 868
        sample_size (int): The max sample size of negative box, used only when
            mining_type is 'hard_example'.
869 870

    Returns:
Y
yuyang18 已提交
871 872
        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`.
873 874

    Raises:
Y
yuyang18 已提交
875 876
        ValueError: If mining_type is 'hard_example', now only support mining \
        type of `max_negative`.
Y
yuyang18 已提交
877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895

    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)
896 897 898 899 900 901 902
    """

    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
903
    conf_shape = nn.shape(confidence)
904 905

    def __reshape_to_2d(var):
906
        return nn.flatten(x=var, axis=2)
907 908 909 910 911

    # 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.
912 913
    matched_indices, matched_dist = bipartite_match(iou, match_type,
                                                    overlap_threshold)
914 915 916

    # 2. Compute confidence for mining hard examples
    # 2.1. Get the target label based on matched indices
917 918
    gt_label = nn.reshape(
        x=gt_label, shape=(len(gt_label.shape) - 1) * (0, ) + (-1, 1))
919
    gt_label.stop_gradient = True
920 921 922 923 924 925 926
    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)
927
    target_label.stop_gradient = True
928 929
    conf_loss = nn.softmax_with_cross_entropy(confidence, target_label)
    # 3. Mining hard examples
G
merge  
gongweibao 已提交
930
    actual_shape = nn.slice(conf_shape, axes=[0], starts=[0], ends=[2])
931
    actual_shape.stop_gradient = True
932
    conf_loss = nn.reshape(
933
        x=conf_loss, shape=(num, num_prior), actual_shape=actual_shape)
934
    conf_loss.stop_gradient = True
X
Xin Pan 已提交
935
    neg_indices = helper.create_variable_for_type_inference(dtype='int32')
936
    dtype = matched_indices.dtype
X
Xin Pan 已提交
937 938
    updated_matched_indices = helper.create_variable_for_type_inference(
        dtype=dtype)
939 940 941 942 943 944 945 946 947 948 949 950 951 952
    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
953
            'neg_dist_threshold': neg_overlap,
954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978
            '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')
979

980 981 982 983
    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

984 985 986 987
    # the target_label and target_conf_weight do not have gradient.
    target_label.stop_gradient = True
    target_conf_weight.stop_gradient = True

988 989 990 991 992 993 994 995
    # 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

996 997 998 999
    # the target_bbox and target_loc_weight do not have gradient.
    target_bbox.stop_gradient = True
    target_loc_weight.stop_gradient = True

1000 1001
    # 5.3 Compute overall weighted loss.
    loss = conf_loss_weight * conf_loss + loc_loss_weight * loc_loss
1002
    # reshape to [N, Np], N is the batch size and Np is the prior box number.
1003
    loss = nn.reshape(x=loss, shape=(num, num_prior), actual_shape=actual_shape)
1004 1005 1006 1007 1008
    loss = nn.reduce_sum(loss, dim=1, keep_dim=True)
    if normalize:
        normalizer = nn.reduce_sum(target_loc_weight)
        loss = loss / normalizer

1009
    return loss
C
chengduoZH 已提交
1010 1011


1012 1013 1014 1015
def prior_box(input,
              image,
              min_sizes,
              max_sizes=None,
1016
              aspect_ratios=[1.],
1017 1018 1019 1020 1021
              variance=[0.1, 0.1, 0.2, 0.2],
              flip=False,
              clip=False,
              steps=[0.0, 0.0],
              offset=0.5,
1022 1023
              name=None,
              min_max_aspect_ratios_order=False):
1024
    """
Q
update  
qiaolongfei 已提交
1025
    **Prior Box Operator**
1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036

    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.
1037
       min_sizes(list|tuple|float value): min sizes of generated prior boxes.
1038 1039
       max_sizes(list|tuple|None): max sizes of generated prior boxes.
            Default: None.
1040 1041
       aspect_ratios(list|tuple|float value): the aspect ratios of generated
            prior boxes. Default: [1.].
1042 1043 1044 1045
       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.
1046
       step(list|turple): Prior boxes step across width and height, If
1047
            step[0] == 0.0/step[1] == 0.0, the prior boxes step across
1048 1049
            height/weight of the input will be automatically calculated.
            Default: [0., 0.]
1050 1051
       offset(float): Prior boxes center offset. Default: 0.5
       name(str): Name of the prior box op. Default: None.
1052
       min_max_aspect_ratios_order(bool): If set True, the output prior box is
M
minqiyang 已提交
1053
            in order of [min, max, aspect_ratios], which is consistent with
1054 1055 1056
            Caffe. Please note, this order affects the weights order of
            convolution layer followed by and does not affect the final
            detection results. Default: False.
1057 1058

    Returns:
Q
update  
qiaolongfei 已提交
1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
        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
1072 1073 1074 1075


    Examples:
        .. code-block:: python
Q
update  
qiaolongfei 已提交
1076 1077 1078 1079 1080 1081 1082

            box, var = fluid.layers.prior_box(
                input=conv1,
                image=images,
                min_sizes=[100.],
                flip=True,
                clip=True)
1083 1084 1085 1086
    """
    helper = LayerHelper("prior_box", **locals())
    dtype = helper.input_dtype()

1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101
    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))

1102 1103 1104 1105 1106 1107 1108 1109
    attrs = {
        'min_sizes': min_sizes,
        'aspect_ratios': aspect_ratios,
        'variances': variance,
        'flip': flip,
        'clip': clip,
        'step_w': steps[0],
        'step_h': steps[1],
1110 1111
        'offset': offset,
        'min_max_aspect_ratios_order': min_max_aspect_ratios_order
1112 1113
    }
    if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0:
1114 1115
        if not _is_list_or_tuple_(max_sizes):
            max_sizes = [max_sizes]
1116 1117
        attrs['max_sizes'] = max_sizes

X
Xin Pan 已提交
1118 1119
    box = helper.create_variable_for_type_inference(dtype)
    var = helper.create_variable_for_type_inference(dtype)
1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131
    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
1132 1133 1134 1135 1136 1137 1138 1139 1140
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,
1141
                      flatten_to_2d=False,
R
ruri 已提交
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177
                      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.
       step(list|turple): Prior boxes step across width and height, If
            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
1178 1179
       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
1180 1181 1182 1183 1184 1185
       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.
1186 1187 1188 1189
            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
1190 1191

        variances: the expanded variances of PriorBox.
1192 1193 1194 1195
            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
1196 1197 1198 1199 1200 1201 1202 1203


    Examples:
        .. code-block:: python

            box, var = fluid.layers.density_prior_box(
                input=conv1,
                image=images,
1204 1205 1206 1207 1208
                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
1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238
    """
    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,
1239 1240 1241 1242
        'densities': densities,
        'fixed_sizes': fixed_sizes,
        'fixed_ratios': fixed_ratios,
        'flatten_to_2d': flatten_to_2d,
R
ruri 已提交
1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257
    }
    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
1258
def multi_box_head(inputs,
C
chengduoZH 已提交
1259 1260
                   image,
                   base_size,
C
chengduoZH 已提交
1261
                   num_classes,
C
chengduoZH 已提交
1262
                   aspect_ratios,
1263 1264
                   min_ratio=None,
                   max_ratio=None,
C
chengduoZH 已提交
1265 1266
                   min_sizes=None,
                   max_sizes=None,
C
chengduoZH 已提交
1267 1268 1269 1270
                   steps=None,
                   step_w=None,
                   step_h=None,
                   offset=0.5,
1271 1272
                   variance=[0.1, 0.1, 0.2, 0.2],
                   flip=True,
C
chengduoZH 已提交
1273
                   clip=False,
C
chengduoZH 已提交
1274
                   kernel_size=1,
C
chengduoZH 已提交
1275
                   pad=0,
C
chengduoZH 已提交
1276
                   stride=1,
1277 1278
                   name=None,
                   min_max_aspect_ratios_order=False):
C
chengduoZH 已提交
1279
    """
C
chengduoZH 已提交
1280 1281
    Generate prior boxes for SSD(Single Shot MultiBox Detector)
    algorithm. The details of this algorithm, please refer the
Q
update  
qiaolongfei 已提交
1282
    section 2.2 of SSD paper `SSD: Single Shot MultiBox Detector
C
chengduoZH 已提交
1283
    <https://arxiv.org/abs/1512.02325>`_ .
C
chengduoZH 已提交
1284 1285

    Args:
1286
       inputs(list|tuple): The list of input Variables, the format
C
chengduoZH 已提交
1287
            of all Variables is NCHW.
C
chengduoZH 已提交
1288 1289
       image(Variable): The input image data of PriorBoxOp,
            the layout is NCHW.
C
chengduoZH 已提交
1290 1291
       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
1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313
       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.
1314
            Default:[0.1, 0.1, 0.2, 0.2].
C
chengduoZH 已提交
1315 1316 1317 1318 1319 1320
       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.
1321
       min_max_aspect_ratios_order(bool): If set True, the output prior box is
M
minqiyang 已提交
1322
            in order of [min, max, aspect_ratios], which is consistent with
1323 1324 1325
            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
1326 1327

    Returns:
Q
update  
qiaolongfei 已提交
1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
        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
1343

C
chengduoZH 已提交
1344 1345 1346

    Examples:
        .. code-block:: python
Q
update  
qiaolongfei 已提交
1347 1348

          mbox_locs, mbox_confs, box, var = fluid.layers.multi_box_head(
C
chengduoZH 已提交
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358
            inputs=[conv1, conv2, conv3, conv4, conv5, conv5],
            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
1359 1360
    """

C
chengduoZH 已提交
1361
    def _reshape_with_axis_(input, axis=1):
1362
        out = nn.flatten(x=input, axis=axis)
C
chengduoZH 已提交
1363
        return out
1364

1365 1366
    def _is_list_or_tuple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))
1367

C
chengduoZH 已提交
1368 1369 1370 1371
    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)

1372 1373
    if not _is_list_or_tuple_(inputs):
        raise ValueError('inputs should be a list or tuple.')
C
chengduoZH 已提交
1374

C
chengduoZH 已提交
1375 1376 1377 1378 1379
    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
1380
    elif min_sizes is None and max_sizes is None:
C
chengduoZH 已提交
1381 1382 1383
        min_sizes = []
        max_sizes = []
        step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2)))
M
minqiyang 已提交
1384
        for ratio in six.moves.range(min_ratio, max_ratio + 1, step):
C
chengduoZH 已提交
1385 1386 1387 1388 1389
            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
1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412
    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
1413 1414
    mbox_locs = []
    mbox_confs = []
C
chengduoZH 已提交
1415 1416
    box_results = []
    var_results = []
C
chengduoZH 已提交
1417 1418
    for i, input in enumerate(inputs):
        min_size = min_sizes[i]
C
chengduoZH 已提交
1419 1420
        max_size = max_sizes[i]

1421
        if not _is_list_or_tuple_(min_size):
C
chengduoZH 已提交
1422
            min_size = [min_size]
C
chengduoZH 已提交
1423 1424
        if not _is_list_or_tuple_(max_size):
            max_size = [max_size]
C
chengduoZH 已提交
1425 1426 1427 1428

        aspect_ratio = []
        if aspect_ratios is not None:
            aspect_ratio = aspect_ratios[i]
1429
            if not _is_list_or_tuple_(aspect_ratio):
C
chengduoZH 已提交
1430
                aspect_ratio = [aspect_ratio]
1431
        step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0]
C
chengduoZH 已提交
1432

1433
        box, var = prior_box(input, image, min_size, max_size, aspect_ratio,
1434 1435
                             variance, flip, clip, step, offset, None,
                             min_max_aspect_ratios_order)
C
chengduoZH 已提交
1436 1437 1438 1439 1440

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

        num_boxes = box.shape[2]
C
chengduoZH 已提交
1441

1442
        # get loc
Y
Yuan Gao 已提交
1443
        num_loc_output = num_boxes * 4
1444
        mbox_loc = nn.conv2d(
C
chengduoZH 已提交
1445
            input=input,
1446 1447 1448 1449 1450
            num_filters=num_loc_output,
            filter_size=kernel_size,
            padding=pad,
            stride=stride)

1451
        mbox_loc = nn.transpose(mbox_loc, perm=[0, 2, 3, 1])
1452
        compile_shape = [
1453
            mbox_loc.shape[0], cpt.floor_division(
M
minqiyang 已提交
1454
                mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3], 4), 4
Y
Yuan Gao 已提交
1455
        ]
1456 1457 1458
        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
1459
        mbox_locs.append(mbox_loc_flatten)
C
chengduoZH 已提交
1460

1461
        # get conf
C
chengduoZH 已提交
1462
        num_conf_output = num_boxes * num_classes
1463
        conf_loc = nn.conv2d(
C
chengduoZH 已提交
1464
            input=input,
1465 1466 1467 1468
            num_filters=num_conf_output,
            filter_size=kernel_size,
            padding=pad,
            stride=stride)
1469
        conf_loc = nn.transpose(conf_loc, perm=[0, 2, 3, 1])
1470 1471
        new_shape = [0, -1, num_classes]
        compile_shape = [
1472 1473 1474
            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
1475
        ]
1476 1477 1478 1479
        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
1480
        mbox_confs.append(conf_loc_flatten)
C
chengduoZH 已提交
1481

C
chengduoZH 已提交
1482 1483 1484
    if len(box_results) == 1:
        box = box_results[0]
        var = var_results[0]
Y
Yuan Gao 已提交
1485 1486
        mbox_locs_concat = mbox_locs[0]
        mbox_confs_concat = mbox_confs[0]
C
chengduoZH 已提交
1487 1488 1489 1490 1491 1492 1493 1494 1495
    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
1496 1497
        mbox_locs_concat = tensor.concat(mbox_locs, axis=1)
        mbox_confs_concat = tensor.concat(mbox_confs, axis=1)
C
chengduoZH 已提交
1498

1499 1500
    box.stop_gradient = True
    var.stop_gradient = True
Y
Yuan Gao 已提交
1501
    return mbox_locs_concat, mbox_confs_concat, box, var
1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521


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
1522 1523
                                       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.
1524
       aspect_ratios(list|tuple|float): The height / width ratios of generated
H
haowang101779990 已提交
1525
                                        anchors, e.g. [0.5, 1.0, 2.0].
1526
       variance(list|tuple): The variances to be used in box regression deltas.
H
haowang101779990 已提交
1527 1528
                             Default:[0.1, 0.1, 0.2, 0.2].
       stride(list|turple): The anchors stride across width and height,e.g. [16.0, 16.0]
1529 1530 1531 1532
       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
1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545
        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.
1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585


    Examples:

        .. code-block:: python

            anchor, var = anchor_generator(
                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
1586 1587
    anchor = helper.create_variable_for_type_inference(dtype)
    var = helper.create_variable_for_type_inference(dtype)
1588 1589 1590 1591 1592 1593 1594 1595 1596
    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
1597 1598


W
whs 已提交
1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
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.
        transformed_height (integer): The width of transformed output.
        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

            out = fluid.layers.roi_perspective_transform(input, rois, 7, 7, 1.0)
    """
    helper = LayerHelper('roi_perspective_transform', **locals())
    dtype = helper.input_dtype()
X
Xin Pan 已提交
1633
    out = helper.create_variable_for_type_inference(dtype)
W
whs 已提交
1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646
    helper.append_op(
        type="roi_perspective_transform",
        inputs={"X": input,
                "ROIs": rois},
        outputs={"Out": out},
        attrs={
            "transformed_height": transformed_height,
            "transformed_width": transformed_width,
            "spatial_scale": spatial_scale
        })
    return out


1647 1648
def generate_proposal_labels(rpn_rois,
                             gt_classes,
1649
                             is_crowd,
1650
                             gt_boxes,
1651
                             im_info,
1652 1653 1654 1655 1656 1657
                             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],
1658 1659
                             class_nums=None,
                             use_random=True):
1660
    """
1661
    ** Generate Proposal Labels of Faster-RCNN **
B
buxingyuan 已提交
1662
    This operator can be, for given the GenerateProposalOp output bounding boxes and groundtruth,
B
buxingyuan 已提交
1663
    to sample foreground boxes and background boxes, and compute loss target.
B
buxingyuan 已提交
1664 1665 1666

    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
1667
    If an instance with a groundtruth overlap greater than fg_thresh, then it was considered as a foreground sample.
B
buxingyuan 已提交
1668 1669
    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
1670
    After all foreground and background boxes are chosen (so called Rois),
B
buxingyuan 已提交
1671
    then we apply random sampling to make sure
B
buxingyuan 已提交
1672
    the number of foreground boxes is no more than batch_size_per_im * fg_fraction.
B
buxingyuan 已提交
1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691

    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.
1692 1693 1694 1695
    """

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

X
Xin Pan 已提交
1696 1697 1698 1699 1700 1701 1702 1703 1704
    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)
1705 1706 1707 1708 1709 1710

    helper.append_op(
        type="generate_proposal_labels",
        inputs={
            'RpnRois': rpn_rois,
            'GtClasses': gt_classes,
1711
            'IsCrowd': is_crowd,
1712
            'GtBoxes': gt_boxes,
1713
            'ImInfo': im_info
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728
        },
        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,
1729 1730
            'class_nums': class_nums,
            'use_random': use_random
1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741
        })

    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


1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875
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

          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)
          # rois, labels_int32 can be the output of
          # fluid.layers.generate_proposal_labels.
          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,
              labels_int32=labels_int32,
              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


1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887
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
1888 1889
    **Generate proposal Faster-RCNN**

1890 1891 1892 1893
    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
1894 1895 1896 1897
    could be used to train detection net.

    For generating proposals, this operation performs following steps:

1898 1899
    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
1900 1901 1902 1903 1904 1905
    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:
1906 1907 1908 1909 1910 1911 1912 1913 1914
        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
1915
            between origin image size and the size of feature map.
1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926
        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
1927
        nms_thresh(float): Threshold in NMS, 0.5 by default.
1928 1929 1930 1931
        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.
1932 1933 1934
    """
    helper = LayerHelper('generate_proposals', **locals())

X
Xin Pan 已提交
1935 1936 1937 1938
    rpn_rois = helper.create_variable_for_type_inference(
        dtype=bbox_deltas.dtype)
    rpn_roi_probs = helper.create_variable_for_type_inference(
        dtype=scores.dtype)
1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960
    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
1961 1962 1963 1964 1965 1966 1967


def multiclass_nms(bboxes,
                   scores,
                   score_threshold,
                   nms_top_k,
                   keep_top_k,
J
jerrywgz 已提交
1968
                   nms_threshold=0.3,
J
jerrywgz 已提交
1969 1970
                   normalized=True,
                   nms_eta=1.,
1971 1972
                   background_label=0,
                   name=None):
J
jerrywgz 已提交
1973
    """
1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034
    **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
2035 2036 2037 2038
             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}) 
2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054

    Examples:
        .. code-block:: python

            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
2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076
    """
    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

    return output