detection.py 63.1 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
23 24
from . import tensor
from . import nn
25
from . import ops
M
minqiyang 已提交
26
from ... import compat as cpt
C
chengduoZH 已提交
27
import math
M
minqiyang 已提交
28
import six
29
import numpy
30
from functools import reduce
31

C
chengduoZH 已提交
32
__all__ = [
33
    'prior_box',
C
chengduoZH 已提交
34
    'multi_box_head',
35 36 37 38
    'bipartite_match',
    'target_assign',
    'detection_output',
    'ssd_loss',
39
    'detection_map',
Y
Yuan Gao 已提交
40
    'rpn_target_assign',
41
    'anchor_generator',
W
whs 已提交
42
    'roi_perspective_transform',
43
    'generate_proposal_labels',
44
    'generate_proposals',
45 46
    'iou_similarity',
    'box_coder',
B
Bai Yifan 已提交
47
    'polygon_box_transform',
C
chengduoZH 已提交
48
]
49 50


51 52
def rpn_target_assign(bbox_pred,
                      cls_logits,
Y
Yuan Gao 已提交
53
                      anchor_box,
54
                      anchor_var,
55 56 57
                      gt_boxes,
                      is_crowd,
                      im_info,
Y
Yuan Gao 已提交
58
                      rpn_batch_size_per_im=256,
59 60
                      rpn_straddle_thresh=0.0,
                      rpn_fg_fraction=0.5,
Y
Yuan Gao 已提交
61
                      rpn_positive_overlap=0.7,
62 63
                      rpn_negative_overlap=0.3,
                      use_random=True):
Y
Yuan Gao 已提交
64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82
    """
    ** Target Assign Layer for region proposal network (RPN) in Faster-RCNN detection. **

    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:
83
        bbox_pred(Variable): A 3-D Tensor with shape [N, M, 4] represents the
Y
Yuan Gao 已提交
84 85 86
            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].
87 88 89
        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
90 91 92 93 94 95
        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.
96 97
        anchor_var(Variable): A 2-D Tensor with shape [M,4] holds expanded 
            variances of anchors.
98
        gt_boxes (Variable): The ground-truth boudding boxes (bboxes) are a 2D
Y
Yuan Gao 已提交
99 100
            LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth
            bboxes of mini-batch input.
101 102 103
        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
104
        rpn_batch_size_per_im(int): Total number of RPN examples per image.
105 106 107
        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
108 109 110 111 112 113 114 115 116
            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
117
        tuple:
Y
Yuan Gao 已提交
118 119 120 121 122 123 124 125 126 127
               A tuple(predicted_scores, predicted_location, target_label,
               target_bbox) is returned. The predicted_scores and
               predicted_location is the predicted result of the RPN.
               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
128
               anchors, the F and B is depends on the input of this operator.
Y
Yuan Gao 已提交
129 130 131 132

    Examples:
        .. code-block:: python

133
        bbox_pred = layers.data(name='bbox_pred', shape=[100, 4],
Y
Yuan Gao 已提交
134
                          append_batch_size=False, dtype='float32')
135
        cls_logits = layers.data(name='cls_logits', shape=[100, 1],
Y
Yuan Gao 已提交
136 137 138
                          append_batch_size=False, dtype='float32')
        anchor_box = layers.data(name='anchor_box', shape=[20, 4],
                          append_batch_size=False, dtype='float32')
139
        gt_boxes = layers.data(name='gt_boxes', shape=[10, 4],
Y
Yuan Gao 已提交
140 141
                         append_batch_size=False, dtype='float32')
        loc_pred, score_pred, loc_target, score_target =
142 143
            fluid.layers.rpn_target_assign(bbox_pred=bbox_pred,
                                          cls_logits=cls_logits,
Y
Yuan Gao 已提交
144
                                          anchor_box=anchor_box,
145
                                          gt_boxes=gt_boxes)
Y
Yuan Gao 已提交
146 147 148
    """

    helper = LayerHelper('rpn_target_assign', **locals())
149 150 151
    # Assign target label to anchors
    loc_index = helper.create_tmp_variable(dtype='int32')
    score_index = helper.create_tmp_variable(dtype='int32')
152
    target_label = helper.create_tmp_variable(dtype='int32')
153
    target_bbox = helper.create_tmp_variable(dtype=anchor_box.dtype)
Y
Yuan Gao 已提交
154 155
    helper.append_op(
        type="rpn_target_assign",
156 157 158 159 160 161
        inputs={
            'Anchor': anchor_box,
            'GtBoxes': gt_boxes,
            'IsCrowd': is_crowd,
            'ImInfo': im_info
        },
Y
Yuan Gao 已提交
162 163 164
        outputs={
            'LocationIndex': loc_index,
            'ScoreIndex': score_index,
165
            'TargetLabel': target_label,
166
            'TargetBBox': target_bbox
Y
Yuan Gao 已提交
167 168 169
        },
        attrs={
            'rpn_batch_size_per_im': rpn_batch_size_per_im,
170
            'rpn_straddle_thresh': rpn_straddle_thresh,
Y
Yuan Gao 已提交
171 172
            'rpn_positive_overlap': rpn_positive_overlap,
            'rpn_negative_overlap': rpn_negative_overlap,
173 174
            'rpn_fg_fraction': rpn_fg_fraction,
            'use_random': use_random
Y
Yuan Gao 已提交
175 176
        })

177 178 179 180
    loc_index.stop_gradient = True
    score_index.stop_gradient = True
    target_label.stop_gradient = True
    target_bbox.stop_gradient = True
Y
Yuan Gao 已提交
181

182 183 184 185
    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)
186

187
    return predicted_cls_logits, predicted_bbox_pred, target_label, target_bbox
Y
Yuan Gao 已提交
188 189


Y
Yuan Gao 已提交
190 191
def detection_output(loc,
                     scores,
192 193 194 195 196 197 198 199 200
                     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):
    """
201
    **Detection Output Layer for Single Shot Multibox Detector (SSD).**
202

203 204
    This operation is to get the detection results by performing following
    two steps:
C
caoying03 已提交
205

206 207 208 209 210 211
    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.
212 213 214 215 216 217

    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
218 219 220 221
        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.
222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
        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
244 245
        Variable:

246
            The detection outputs is a LoDTensor with shape [No, 6].
247 248 249 250 251 252 253 254
            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,
            all the elements in LoD are 0, and output tensor only contains one
            value, which is -1.
255 256 257 258

    Examples:
        .. code-block:: python

259
            pb = layers.data(name='prior_box', shape=[10, 4],
260
                         append_batch_size=False, dtype='float32')
261
            pbv = layers.data(name='prior_box_var', shape=[10, 4],
262
                          append_batch_size=False, dtype='float32')
263
            loc = layers.data(name='target_box', shape=[2, 21, 4],
264
                          append_batch_size=False, dtype='float32')
265
            scores = layers.data(name='scores', shape=[2, 21, 10],
266
                          append_batch_size=False, dtype='float32')
267
            nmsed_outs = fluid.layers.detection_output(scores=scores,
268 269 270 271 272
                                       loc=loc,
                                       prior_box=pb,
                                       prior_box_var=pbv)
    """
    helper = LayerHelper("detection_output", **locals())
273 274 275 276 277
    decoded_box = box_coder(
        prior_box=prior_box,
        prior_box_var=prior_box_var,
        target_box=loc,
        code_type='decode_center_size')
278
    compile_shape = scores.shape
G
merge  
gongweibao 已提交
279
    run_shape = nn.shape(scores)
280
    scores = nn.flatten(x=scores, axis=2)
281
    scores = nn.softmax(input=scores)
282
    scores = nn.reshape(x=scores, shape=compile_shape, actual_shape=run_shape)
Y
Yuan Gao 已提交
283
    scores = nn.transpose(scores, perm=[0, 2, 1])
284
    scores.stop_gradient = True
285
    nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype)
286 287 288 289 290 291 292 293 294 295 296 297 298
    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
        })
299
    nmsed_outs.stop_gradient = True
300
    return nmsed_outs
C
chengduoZH 已提交
301 302


X
Xin Pan 已提交
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
@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:
        out = helper.create_tmp_variable(dtype=x.dtype)
    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:
        output_box = helper.create_tmp_variable(dtype=prior_box.dtype)
    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:
        output = helper.create_tmp_variable(dtype=input.dtype)
    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


X
Xin Pan 已提交
398
@templatedoc()
399 400
def detection_map(detect_res,
                  label,
401 402
                  class_num,
                  background_label=0,
403 404
                  overlap_threshold=0.3,
                  evaluate_difficult=True,
405 406 407 408
                  has_state=None,
                  input_states=None,
                  out_states=None,
                  ap_version='integral'):
X
Xin Pan 已提交
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
    """
    ${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)
    """
450 451
    helper = LayerHelper("detection_map", **locals())

452 453 454 455 456 457 458 459 460 461 462 463 464 465
    def __create_var(type):
        return helper.create_tmp_variable(dtype=type)

    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

466 467 468 469 470
    helper.append_op(
        type="detection_map",
        inputs={
            'Label': label,
            'DetectRes': detect_res,
471
            'HasState': has_state,
472 473 474 475 476 477 478 479 480 481 482 483 484
            '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,
485 486
            'ap_type': ap_version,
            'class_num': class_num,
487
        })
488
    return map_out
489 490


491 492 493 494
def bipartite_match(dist_matrix,
                    match_type=None,
                    dist_threshold=None,
                    name=None):
495
    """
Y
yuyang18 已提交
496 497
    This operator implements a greedy bipartite matching algorithm, which is
    used to obtain the matching with the maximum distance based on the input
498
    distance matrix. For input 2D matrix, the bipartite matching algorithm can
Y
yuyang18 已提交
499 500 501 502 503 504 505 506
    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)
507 508 509
    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
510

Y
yuyang18 已提交
511
    NOTE: the input DistMat can be LoDTensor (with LoD) or Tensor.
512 513 514
    If LoDTensor with LoD, the height of ColToRowMatchIndices is batch size.
    If Tensor, the height of ColToRowMatchIndices is 1.

Y
yuyang18 已提交
515 516 517
    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.

518 519 520 521 522
    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
523 524 525 526 527 528
            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.
529
        match_type(string|None): The type of matching method, should be
Y
yuyang18 已提交
530
           'bipartite' or 'per_prediction'. [default 'bipartite'].
531 532
        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
533
            on the maximum distance, 0.5 by default.
534
    Returns:
Y
yuyang18 已提交
535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
        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)
558 559 560 561 562 563 564
    """
    helper = LayerHelper('bipartite_match', **locals())
    match_indices = helper.create_tmp_variable(dtype='int32')
    match_distance = helper.create_tmp_variable(dtype=dist_matrix.dtype)
    helper.append_op(
        type='bipartite_match',
        inputs={'DistMat': dist_matrix},
565 566 567 568
        attrs={
            'match_type': match_type,
            'dist_threshold': dist_threshold,
        },
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585
        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
586

587 588 589 590 591
    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
592

593
    1. Assigning all outpts based on `match_indices`:
C
chengduoZH 已提交
594

595 596 597
    .. code-block:: text

        If id = match_indices[i][j] > 0,
C
chengduoZH 已提交
598

599 600
            out[i][j][0 : K] = X[lod[i] + id][j % P][0 : K]
            out_weight[i][j] = 1.
C
chengduoZH 已提交
601

602
        Otherwise,
C
chengduoZH 已提交
603

604 605
            out[j][j][0 : K] = {mismatch_value, mismatch_value, ...}
            out_weight[i][j] = 0.
C
chengduoZH 已提交
606

607
    2. Assigning out_weight based on `neg_indices` if `neg_indices` is provided:
C
chengduoZH 已提交
608

609 610
    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
611

612
    .. code-block:: text
C
chengduoZH 已提交
613

614 615 616 617 618 619 620 621 622 623 624 625 626 627 628
        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
629 630 631 632 633
        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
634 635 636 637 638 639 640 641 642 643 644
               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)
645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675
    """
    helper = LayerHelper('target_assign', **locals())
    out = helper.create_tmp_variable(dtype=input.dtype)
    out_weight = helper.create_tmp_variable(dtype='float32')
    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',
676
             normalize=True,
677 678
             sample_size=None):
    """
Y
yuyang18 已提交
679
    **Multi-box loss layer for object detection algorithm of SSD**
680 681 682 683 684 685 686

    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
687
    1. Find matched bounding box by bipartite matching algorithm.
Y
yuyang18 已提交
688

689
      1.1 Compute IOU similarity between ground-truth boxes and prior boxes.
Y
yuyang18 已提交
690

691
      1.2 Compute matched boundding box by bipartite matching algorithm.
Y
yuyang18 已提交
692

693
    2. Compute confidence for mining hard examples
Y
yuyang18 已提交
694

695
      2.1. Get the target label based on matched indices.
Y
yuyang18 已提交
696

697
      2.2. Compute confidence loss.
Y
yuyang18 已提交
698

699 700
    3. Apply hard example mining to get the negative example indices and update
       the matched indices.
Y
yuyang18 已提交
701

702
    4. Assign classification and regression targets
Y
yuyang18 已提交
703

704
      4.1. Encoded bbox according to the prior boxes.
Y
yuyang18 已提交
705

706
      4.2. Assign regression targets.
Y
yuyang18 已提交
707

708
      4.3. Assign classification targets.
Y
yuyang18 已提交
709

710
    5. Compute the overall objective loss.
Y
yuyang18 已提交
711

712
      5.1 Compute confidence loss.
Y
yuyang18 已提交
713

714
      5.1 Compute localization loss.
Y
yuyang18 已提交
715

716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
      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
739
            boxes, used only when mining_type is 'max_negative', 3.0 by defalut.
740
        neg_overlap (float): The negative overlap upper bound for the unmatched
741
            predictions. Use only when mining_type is 'max_negative',
742 743 744 745
            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
746
            be 'bipartite' or 'per_prediction', 'per_prediction' by defalut.
747 748
        mining_type (str): The hard example mining type, should be 'hard_example'
            or 'max_negative', now only support `max_negative`.
749
        normalize (bool): Whether to normalize the SSD loss by the total number
Y
yuyang18 已提交
750
            of output locations, True by default.
751 752
        sample_size (int): The max sample size of negative box, used only when
            mining_type is 'hard_example'.
753 754

    Returns:
Y
yuyang18 已提交
755 756
        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`.
757 758

    Raises:
Y
yuyang18 已提交
759 760
        ValueError: If mining_type is 'hard_example', now only support mining \
        type of `max_negative`.
Y
yuyang18 已提交
761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779

    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)
780 781 782 783 784 785 786
    """

    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
787
    conf_shape = nn.shape(confidence)
788 789

    def __reshape_to_2d(var):
790
        return nn.flatten(x=var, axis=2)
791 792 793 794 795

    # 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.
796 797
    matched_indices, matched_dist = bipartite_match(iou, match_type,
                                                    overlap_threshold)
798 799 800

    # 2. Compute confidence for mining hard examples
    # 2.1. Get the target label based on matched indices
801 802
    gt_label = nn.reshape(
        x=gt_label, shape=(len(gt_label.shape) - 1) * (0, ) + (-1, 1))
803
    gt_label.stop_gradient = True
804 805 806 807 808 809 810
    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)
811
    target_label.stop_gradient = True
812 813
    conf_loss = nn.softmax_with_cross_entropy(confidence, target_label)
    # 3. Mining hard examples
G
merge  
gongweibao 已提交
814
    actual_shape = nn.slice(conf_shape, axes=[0], starts=[0], ends=[2])
815
    actual_shape.stop_gradient = True
816
    conf_loss = nn.reshape(
817
        x=conf_loss, shape=(num, num_prior), actual_shape=actual_shape)
818
    conf_loss.stop_gradient = True
819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835
    neg_indices = helper.create_tmp_variable(dtype='int32')
    dtype = matched_indices.dtype
    updated_matched_indices = helper.create_tmp_variable(dtype=dtype)
    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
836
            'neg_dist_threshold': neg_overlap,
837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861
            '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')
862

863 864 865 866
    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

867 868 869 870
    # the target_label and target_conf_weight do not have gradient.
    target_label.stop_gradient = True
    target_conf_weight.stop_gradient = True

871 872 873 874 875 876 877 878
    # 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

879 880 881 882
    # the target_bbox and target_loc_weight do not have gradient.
    target_bbox.stop_gradient = True
    target_loc_weight.stop_gradient = True

883 884
    # 5.3 Compute overall weighted loss.
    loss = conf_loss_weight * conf_loss + loc_loss_weight * loc_loss
885
    # reshape to [N, Np], N is the batch size and Np is the prior box number.
886
    loss = nn.reshape(x=loss, shape=(num, num_prior), actual_shape=actual_shape)
887 888 889 890 891
    loss = nn.reduce_sum(loss, dim=1, keep_dim=True)
    if normalize:
        normalizer = nn.reduce_sum(target_loc_weight)
        loss = loss / normalizer

892
    return loss
C
chengduoZH 已提交
893 894


895 896 897 898
def prior_box(input,
              image,
              min_sizes,
              max_sizes=None,
899
              aspect_ratios=[1.],
900 901 902 903 904
              variance=[0.1, 0.1, 0.2, 0.2],
              flip=False,
              clip=False,
              steps=[0.0, 0.0],
              offset=0.5,
905 906
              name=None,
              min_max_aspect_ratios_order=False):
907
    """
Q
update  
qiaolongfei 已提交
908
    **Prior Box Operator**
909 910 911 912 913 914 915 916 917 918 919

    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.
920
       min_sizes(list|tuple|float value): min sizes of generated prior boxes.
921 922
       max_sizes(list|tuple|None): max sizes of generated prior boxes.
            Default: None.
923 924
       aspect_ratios(list|tuple|float value): the aspect ratios of generated
            prior boxes. Default: [1.].
925 926 927 928
       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.
929
       step(list|turple): Prior boxes step across width and height, If
930
            step[0] == 0.0/step[1] == 0.0, the prior boxes step across
931 932
            height/weight of the input will be automatically calculated.
            Default: [0., 0.]
933 934
       offset(float): Prior boxes center offset. Default: 0.5
       name(str): Name of the prior box op. Default: None.
935
       min_max_aspect_ratios_order(bool): If set True, the output prior box is
M
minqiyang 已提交
936
            in order of [min, max, aspect_ratios], which is consistent with
937 938 939
            Caffe. Please note, this order affects the weights order of
            convolution layer followed by and does not affect the final
            detection results. Default: False.
940 941

    Returns:
Q
update  
qiaolongfei 已提交
942 943 944 945 946 947 948 949 950 951 952 953 954
        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
955 956 957 958


    Examples:
        .. code-block:: python
Q
update  
qiaolongfei 已提交
959 960 961 962 963 964 965

            box, var = fluid.layers.prior_box(
                input=conv1,
                image=images,
                min_sizes=[100.],
                flip=True,
                clip=True)
966 967 968 969
    """
    helper = LayerHelper("prior_box", **locals())
    dtype = helper.input_dtype()

970 971 972 973 974 975 976 977 978 979 980 981 982 983 984
    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))

985 986 987 988 989 990 991 992
    attrs = {
        'min_sizes': min_sizes,
        'aspect_ratios': aspect_ratios,
        'variances': variance,
        'flip': flip,
        'clip': clip,
        'step_w': steps[0],
        'step_h': steps[1],
993 994
        'offset': offset,
        'min_max_aspect_ratios_order': min_max_aspect_ratios_order
995 996
    }
    if max_sizes is not None and len(max_sizes) > 0 and max_sizes[0] > 0:
997 998
        if not _is_list_or_tuple_(max_sizes):
            max_sizes = [max_sizes]
999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014
        attrs['max_sizes'] = max_sizes

    box = helper.create_tmp_variable(dtype)
    var = helper.create_tmp_variable(dtype)
    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


C
chengduoZH 已提交
1015
def multi_box_head(inputs,
C
chengduoZH 已提交
1016 1017
                   image,
                   base_size,
C
chengduoZH 已提交
1018
                   num_classes,
C
chengduoZH 已提交
1019
                   aspect_ratios,
1020 1021
                   min_ratio=None,
                   max_ratio=None,
C
chengduoZH 已提交
1022 1023
                   min_sizes=None,
                   max_sizes=None,
C
chengduoZH 已提交
1024 1025 1026 1027
                   steps=None,
                   step_w=None,
                   step_h=None,
                   offset=0.5,
1028 1029
                   variance=[0.1, 0.1, 0.2, 0.2],
                   flip=True,
C
chengduoZH 已提交
1030
                   clip=False,
C
chengduoZH 已提交
1031
                   kernel_size=1,
C
chengduoZH 已提交
1032
                   pad=0,
C
chengduoZH 已提交
1033
                   stride=1,
1034 1035
                   name=None,
                   min_max_aspect_ratios_order=False):
C
chengduoZH 已提交
1036
    """
C
chengduoZH 已提交
1037 1038
    Generate prior boxes for SSD(Single Shot MultiBox Detector)
    algorithm. The details of this algorithm, please refer the
Q
update  
qiaolongfei 已提交
1039
    section 2.2 of SSD paper `SSD: Single Shot MultiBox Detector
C
chengduoZH 已提交
1040
    <https://arxiv.org/abs/1512.02325>`_ .
C
chengduoZH 已提交
1041 1042

    Args:
1043
       inputs(list|tuple): The list of input Variables, the format
C
chengduoZH 已提交
1044
            of all Variables is NCHW.
C
chengduoZH 已提交
1045 1046
       image(Variable): The input image data of PriorBoxOp,
            the layout is NCHW.
C
chengduoZH 已提交
1047 1048
       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
1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070
       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.
1071
            Default:[0.1, 0.1, 0.2, 0.2].
C
chengduoZH 已提交
1072 1073 1074 1075 1076 1077
       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.
1078
       min_max_aspect_ratios_order(bool): If set True, the output prior box is
M
minqiyang 已提交
1079
            in order of [min, max, aspect_ratios], which is consistent with
1080 1081 1082
            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
1083 1084

    Returns:
Q
update  
qiaolongfei 已提交
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099
        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
1100

C
chengduoZH 已提交
1101 1102 1103

    Examples:
        .. code-block:: python
Q
update  
qiaolongfei 已提交
1104 1105

          mbox_locs, mbox_confs, box, var = fluid.layers.multi_box_head(
C
chengduoZH 已提交
1106 1107 1108 1109 1110 1111 1112 1113 1114 1115
            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
1116 1117
    """

C
chengduoZH 已提交
1118
    def _reshape_with_axis_(input, axis=1):
1119
        out = nn.flatten(x=input, axis=axis)
C
chengduoZH 已提交
1120
        return out
1121

1122 1123
    def _is_list_or_tuple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))
1124

C
chengduoZH 已提交
1125 1126 1127 1128
    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)

1129 1130
    if not _is_list_or_tuple_(inputs):
        raise ValueError('inputs should be a list or tuple.')
C
chengduoZH 已提交
1131

C
chengduoZH 已提交
1132 1133 1134 1135 1136
    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
1137
    elif min_sizes is None and max_sizes is None:
C
chengduoZH 已提交
1138 1139 1140
        min_sizes = []
        max_sizes = []
        step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2)))
M
minqiyang 已提交
1141
        for ratio in six.moves.range(min_ratio, max_ratio + 1, step):
C
chengduoZH 已提交
1142 1143 1144 1145 1146
            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
1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169
    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
1170 1171
    mbox_locs = []
    mbox_confs = []
C
chengduoZH 已提交
1172 1173
    box_results = []
    var_results = []
C
chengduoZH 已提交
1174 1175
    for i, input in enumerate(inputs):
        min_size = min_sizes[i]
C
chengduoZH 已提交
1176 1177
        max_size = max_sizes[i]

1178
        if not _is_list_or_tuple_(min_size):
C
chengduoZH 已提交
1179
            min_size = [min_size]
C
chengduoZH 已提交
1180 1181
        if not _is_list_or_tuple_(max_size):
            max_size = [max_size]
C
chengduoZH 已提交
1182 1183 1184 1185

        aspect_ratio = []
        if aspect_ratios is not None:
            aspect_ratio = aspect_ratios[i]
1186
            if not _is_list_or_tuple_(aspect_ratio):
C
chengduoZH 已提交
1187
                aspect_ratio = [aspect_ratio]
1188
        step = [step_w[i] if step_w else 0.0, step_h[i] if step_w else 0.0]
C
chengduoZH 已提交
1189

1190
        box, var = prior_box(input, image, min_size, max_size, aspect_ratio,
1191 1192
                             variance, flip, clip, step, offset, None,
                             min_max_aspect_ratios_order)
C
chengduoZH 已提交
1193 1194 1195 1196 1197

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

        num_boxes = box.shape[2]
C
chengduoZH 已提交
1198

1199
        # get loc
Y
Yuan Gao 已提交
1200
        num_loc_output = num_boxes * 4
1201
        mbox_loc = nn.conv2d(
C
chengduoZH 已提交
1202
            input=input,
1203 1204 1205 1206 1207
            num_filters=num_loc_output,
            filter_size=kernel_size,
            padding=pad,
            stride=stride)

1208
        mbox_loc = nn.transpose(mbox_loc, perm=[0, 2, 3, 1])
1209
        compile_shape = [
1210
            mbox_loc.shape[0], cpt.floor_division(
M
minqiyang 已提交
1211
                mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3], 4), 4
Y
Yuan Gao 已提交
1212
        ]
1213 1214 1215
        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
1216
        mbox_locs.append(mbox_loc_flatten)
C
chengduoZH 已提交
1217

1218
        # get conf
C
chengduoZH 已提交
1219
        num_conf_output = num_boxes * num_classes
1220
        conf_loc = nn.conv2d(
C
chengduoZH 已提交
1221
            input=input,
1222 1223 1224 1225
            num_filters=num_conf_output,
            filter_size=kernel_size,
            padding=pad,
            stride=stride)
1226
        conf_loc = nn.transpose(conf_loc, perm=[0, 2, 3, 1])
1227 1228
        new_shape = [0, -1, num_classes]
        compile_shape = [
1229 1230 1231
            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
1232
        ]
1233 1234 1235 1236
        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
1237
        mbox_confs.append(conf_loc_flatten)
C
chengduoZH 已提交
1238

C
chengduoZH 已提交
1239 1240 1241
    if len(box_results) == 1:
        box = box_results[0]
        var = var_results[0]
Y
Yuan Gao 已提交
1242 1243
        mbox_locs_concat = mbox_locs[0]
        mbox_confs_concat = mbox_confs[0]
C
chengduoZH 已提交
1244 1245 1246 1247 1248 1249 1250 1251 1252
    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
1253 1254
        mbox_locs_concat = tensor.concat(mbox_locs, axis=1)
        mbox_confs_concat = tensor.concat(mbox_confs, axis=1)
C
chengduoZH 已提交
1255

1256 1257
    box.stop_gradient = True
    var.stop_gradient = True
Y
Yuan Gao 已提交
1258
    return mbox_locs_concat, mbox_confs_concat, box, var
1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350


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,
       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.
       aspect_ratios(list|tuple|float): The height / width ratios of generated
            anchors, e.g. [0.5, 1.0, 2.0].
       variance(list|tuple): The variances to be used in box regression deltas.
            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]
       offset(float): Prior boxes center offset. Default: 0.5
       name(str): Name of the prior box op. Default: None.

    Returns:
        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.


    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
    }

    anchor = helper.create_tmp_variable(dtype)
    var = helper.create_tmp_variable(dtype)
    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
1351 1352


W
whs 已提交
1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400
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()
    out = helper.create_tmp_variable(dtype)
    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


1401 1402
def generate_proposal_labels(rpn_rois,
                             gt_classes,
1403
                             is_crowd,
1404
                             gt_boxes,
1405
                             im_info,
1406 1407 1408 1409 1410 1411
                             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],
1412 1413
                             class_nums=None,
                             use_random=True):
1414 1415
    """
    ** Generate proposal labels Faster-RCNN **
B
buxingyuan 已提交
1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445
    This operator can be, for given the GenerateProposalOp output bounding boxes and groundtruth,
    to sample foregroud boxes and background boxes, and compute loss target.

    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,
    If an instance with a groundtruth overlap greater than fg_thresh, then it was considered as a foregroud sample.
    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.
    After all foregroud and background boxes are chosen (so called Rois),
    then we apply random sampling to make sure
    the number of foregroud boxes is no more than batch_size_per_im * fg_fraction.

    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.
1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460
    """

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

    rois = helper.create_tmp_variable(dtype=rpn_rois.dtype)
    labels_int32 = helper.create_tmp_variable(dtype=gt_classes.dtype)
    bbox_targets = helper.create_tmp_variable(dtype=rpn_rois.dtype)
    bbox_inside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype)
    bbox_outside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype)

    helper.append_op(
        type="generate_proposal_labels",
        inputs={
            'RpnRois': rpn_rois,
            'GtClasses': gt_classes,
1461
            'IsCrowd': is_crowd,
1462
            'GtBoxes': gt_boxes,
1463
            'ImInfo': im_info
1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478
        },
        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,
1479 1480
            'class_nums': class_nums,
            'use_random': use_random
1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491
        })

    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


1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503
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):
    """
B
buxingyuan 已提交
1504
    ** Generate proposal Faster-RCNN **
1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559
	
	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
	could be used to train detection net.

	For generating proposals, this operation performs following steps:

	1. Transposes and resizes scores and bbox_deltas in size of (H*W*A, 1) and (H*W*A, 4)
 	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:
		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
			between origin image size and the size of feature map.
		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.
		nms_thresh(float): Threshold in NMS, 0.5 by default.
		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.
    """
    helper = LayerHelper('generate_proposals', **locals())

    rpn_rois = helper.create_tmp_variable(dtype=bbox_deltas.dtype)
    rpn_roi_probs = helper.create_tmp_variable(dtype=scores.dtype)
    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