ops.py 42.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

15
import numpy as np
16 17 18
from numbers import Integral

from paddle import fluid
Y
Yuan Gao 已提交
19 20
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.regularizer import L2Decay
21
from ppdet.core.workspace import register, serializable
littletomatodonkey's avatar
littletomatodonkey 已提交
22
from ppdet.utils.bbox_utils import bbox_overlaps, box_to_delta
23 24

__all__ = [
25 26 27
    'AnchorGenerator', 'DropBlock', 'RPNTargetAssign', 'GenerateProposals',
    'MultiClassNMS', 'BBoxAssigner', 'MaskAssigner', 'RoIAlign', 'RoIPool',
    'MultiBoxHead', 'SSDOutputDecoder', 'RetinaTargetAssign',
littletomatodonkey's avatar
littletomatodonkey 已提交
28
    'RetinaOutputDecoder', 'ConvNorm', 'MultiClassSoftNMS', 'LibraBBoxAssigner'
29 30 31
]


Y
Yuan Gao 已提交
32 33 34 35 36 37 38
def ConvNorm(input,
             num_filters,
             filter_size,
             stride=1,
             groups=1,
             norm_decay=0.,
             norm_type='affine_channel',
Y
Yuan Gao 已提交
39 40
             norm_groups=32,
             dilation=1,
41
             lr_scale=1,
Y
Yuan Gao 已提交
42 43
             freeze_norm=False,
             act=None,
Y
Yuan Gao 已提交
44
             norm_name=None,
Y
Yuan Gao 已提交
45 46 47 48 49 50 51 52
             initializer=None,
             name=None):
    fan = num_filters
    conv = fluid.layers.conv2d(
        input=input,
        num_filters=num_filters,
        filter_size=filter_size,
        stride=stride,
Y
Yuan Gao 已提交
53 54
        padding=((filter_size - 1) // 2) * dilation,
        dilation=dilation,
Y
Yuan Gao 已提交
55 56 57
        groups=groups,
        act=None,
        param_attr=ParamAttr(
58 59 60
            name=name + "_weights",
            initializer=initializer,
            learning_rate=lr_scale),
Y
Yuan Gao 已提交
61 62 63 64 65
        bias_attr=False,
        name=name + '.conv2d.output.1')

    norm_lr = 0. if freeze_norm else 1.
    pattr = ParamAttr(
Y
Yuan Gao 已提交
66
        name=norm_name + '_scale',
67
        learning_rate=norm_lr * lr_scale,
Y
Yuan Gao 已提交
68 69
        regularizer=L2Decay(norm_decay))
    battr = ParamAttr(
Y
Yuan Gao 已提交
70
        name=norm_name + '_offset',
71
        learning_rate=norm_lr * lr_scale,
Y
Yuan Gao 已提交
72 73 74 75 76 77 78
        regularizer=L2Decay(norm_decay))

    if norm_type in ['bn', 'sync_bn']:
        global_stats = True if freeze_norm else False
        out = fluid.layers.batch_norm(
            input=conv,
            act=act,
Y
Yuan Gao 已提交
79
            name=norm_name + '.output.1',
Y
Yuan Gao 已提交
80 81
            param_attr=pattr,
            bias_attr=battr,
Y
Yuan Gao 已提交
82 83
            moving_mean_name=norm_name + '_mean',
            moving_variance_name=norm_name + '_variance',
Y
Yuan Gao 已提交
84 85 86
            use_global_stats=global_stats)
        scale = fluid.framework._get_var(pattr.name)
        bias = fluid.framework._get_var(battr.name)
Y
Yuan Gao 已提交
87 88 89 90 91 92 93 94 95 96
    elif norm_type == 'gn':
        out = fluid.layers.group_norm(
            input=conv,
            act=act,
            name=norm_name + '.output.1',
            groups=norm_groups,
            param_attr=pattr,
            bias_attr=battr)
        scale = fluid.framework._get_var(pattr.name)
        bias = fluid.framework._get_var(battr.name)
Y
Yuan Gao 已提交
97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
    elif norm_type == 'affine_channel':
        scale = fluid.layers.create_parameter(
            shape=[conv.shape[1]],
            dtype=conv.dtype,
            attr=pattr,
            default_initializer=fluid.initializer.Constant(1.))
        bias = fluid.layers.create_parameter(
            shape=[conv.shape[1]],
            dtype=conv.dtype,
            attr=battr,
            default_initializer=fluid.initializer.Constant(0.))
        out = fluid.layers.affine_channel(
            x=conv, scale=scale, bias=bias, act=act)
    if freeze_norm:
        scale.stop_gradient = True
        bias.stop_gradient = True
    return out


116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
def DropBlock(input, block_size, keep_prob, is_test):
    if is_test:
        return input

    def CalculateGamma(input, block_size, keep_prob):
        input_shape = fluid.layers.shape(input)
        feat_shape_tmp = fluid.layers.slice(input_shape, [0], [3], [4])
        feat_shape_tmp = fluid.layers.cast(feat_shape_tmp, dtype="float32")
        feat_shape_t = fluid.layers.reshape(feat_shape_tmp, [1, 1, 1, 1])
        feat_area = fluid.layers.pow(feat_shape_t, factor=2)

        block_shape_t = fluid.layers.fill_constant(
            shape=[1, 1, 1, 1], value=block_size, dtype='float32')
        block_area = fluid.layers.pow(block_shape_t, factor=2)

        useful_shape_t = feat_shape_t - block_shape_t + 1
        useful_area = fluid.layers.pow(useful_shape_t, factor=2)

        upper_t = feat_area * (1 - keep_prob)
        bottom_t = block_area * useful_area
        output = upper_t / bottom_t
        return output

    gamma = CalculateGamma(input, block_size=block_size, keep_prob=keep_prob)
    input_shape = fluid.layers.shape(input)
    p = fluid.layers.expand_as(gamma, input)

    input_shape_tmp = fluid.layers.cast(input_shape, dtype="int64")
    random_matrix = fluid.layers.uniform_random(
        input_shape_tmp, dtype='float32', min=0.0, max=1.0)
    one_zero_m = fluid.layers.less_than(random_matrix, p)
    one_zero_m.stop_gradient = True
    one_zero_m = fluid.layers.cast(one_zero_m, dtype="float32")

    mask_flag = fluid.layers.pool2d(
        one_zero_m,
        pool_size=block_size,
        pool_type='max',
        pool_stride=1,
        pool_padding=block_size // 2)
    mask = 1.0 - mask_flag

    elem_numel = fluid.layers.reduce_prod(input_shape)
C
CodesFarmer 已提交
159 160
    elem_numel_m = fluid.layers.cast(elem_numel, dtype="float32")
    elem_numel_m.stop_gradient = True
161 162

    elem_sum = fluid.layers.reduce_sum(mask)
C
CodesFarmer 已提交
163 164
    elem_sum_m = fluid.layers.cast(elem_sum, dtype="float32")
    elem_sum_m.stop_gradient = True
165 166 167 168 169

    output = input * mask * elem_numel_m / elem_sum_m
    return output


170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
@register
@serializable
class AnchorGenerator(object):
    __op__ = fluid.layers.anchor_generator
    __append_doc__ = True

    def __init__(self,
                 stride=[16.0, 16.0],
                 anchor_sizes=[32, 64, 128, 256, 512],
                 aspect_ratios=[0.5, 1., 2.],
                 variance=[1., 1., 1., 1.]):
        super(AnchorGenerator, self).__init__()
        self.anchor_sizes = anchor_sizes
        self.aspect_ratios = aspect_ratios
        self.variance = variance
        self.stride = stride


@register
@serializable
class RPNTargetAssign(object):
    __op__ = fluid.layers.rpn_target_assign
    __append_doc__ = True

    def __init__(self,
                 rpn_batch_size_per_im=256,
                 rpn_straddle_thresh=0.,
                 rpn_fg_fraction=0.5,
                 rpn_positive_overlap=0.7,
                 rpn_negative_overlap=0.3,
                 use_random=True):
        super(RPNTargetAssign, self).__init__()
        self.rpn_batch_size_per_im = rpn_batch_size_per_im
        self.rpn_straddle_thresh = rpn_straddle_thresh
        self.rpn_fg_fraction = rpn_fg_fraction
        self.rpn_positive_overlap = rpn_positive_overlap
        self.rpn_negative_overlap = rpn_negative_overlap
        self.use_random = use_random


@register
@serializable
class GenerateProposals(object):
    __op__ = fluid.layers.generate_proposals
    __append_doc__ = True

    def __init__(self,
                 pre_nms_top_n=6000,
                 post_nms_top_n=1000,
                 nms_thresh=.5,
                 min_size=.1,
                 eta=1.):
        super(GenerateProposals, self).__init__()
        self.pre_nms_top_n = pre_nms_top_n
        self.post_nms_top_n = post_nms_top_n
        self.nms_thresh = nms_thresh
        self.min_size = min_size
        self.eta = eta


@register
class MaskAssigner(object):
    __op__ = fluid.layers.generate_mask_labels
    __append_doc__ = True
234
    __shared__ = ['num_classes']
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264

    def __init__(self, num_classes=81, resolution=14):
        super(MaskAssigner, self).__init__()
        self.num_classes = num_classes
        self.resolution = resolution


@register
@serializable
class MultiClassNMS(object):
    __op__ = fluid.layers.multiclass_nms
    __append_doc__ = True

    def __init__(self,
                 score_threshold=.05,
                 nms_top_k=-1,
                 keep_top_k=100,
                 nms_threshold=.5,
                 normalized=False,
                 nms_eta=1.0,
                 background_label=0):
        super(MultiClassNMS, self).__init__()
        self.score_threshold = score_threshold
        self.nms_top_k = nms_top_k
        self.keep_top_k = keep_top_k
        self.nms_threshold = nms_threshold
        self.normalized = normalized
        self.nms_eta = nms_eta
        self.background_label = background_label

265

266 267 268
@register
@serializable
class MultiClassSoftNMS(object):
269 270 271 272 273 274 275
    def __init__(
            self,
            score_threshold=0.01,
            keep_top_k=300,
            softnms_sigma=0.5,
            normalized=False,
            background_label=0, ):
276 277 278 279 280 281
        super(MultiClassSoftNMS, self).__init__()
        self.score_threshold = score_threshold
        self.keep_top_k = keep_top_k
        self.softnms_sigma = softnms_sigma
        self.normalized = normalized
        self.background_label = background_label
282 283

    def __call__(self, bboxes, scores):
littletomatodonkey's avatar
littletomatodonkey 已提交
284
        def create_tmp_var(program, name, dtype, shape, lod_level):
285
            return program.current_block().create_var(
littletomatodonkey's avatar
littletomatodonkey 已提交
286
                name=name, dtype=dtype, shape=shape, lod_level=lod_level)
287

288 289 290 291 292 293 294 295
        def _soft_nms_for_cls(dets, sigma, thres):
            """soft_nms_for_cls"""
            dets_final = []
            while len(dets) > 0:
                maxpos = np.argmax(dets[:, 0])
                dets_final.append(dets[maxpos].copy())
                ts, tx1, ty1, tx2, ty2 = dets[maxpos]
                scores = dets[:, 0]
296 297
                # force remove bbox at maxpos
                scores[maxpos] = -1
298 299 300 301 302 303 304 305 306 307 308 309 310
                x1 = dets[:, 1]
                y1 = dets[:, 2]
                x2 = dets[:, 3]
                y2 = dets[:, 4]
                eta = 0 if self.normalized else 1
                areas = (x2 - x1 + eta) * (y2 - y1 + eta)
                xx1 = np.maximum(tx1, x1)
                yy1 = np.maximum(ty1, y1)
                xx2 = np.minimum(tx2, x2)
                yy2 = np.minimum(ty2, y2)
                w = np.maximum(0.0, xx2 - xx1 + eta)
                h = np.maximum(0.0, yy2 - yy1 + eta)
                inter = w * h
311
                ovr = inter / (areas + areas[maxpos] - inter)
312 313 314 315 316 317
                weight = np.exp(-(ovr * ovr) / sigma)
                scores = scores * weight
                idx_keep = np.where(scores >= thres)
                dets[:, 0] = scores
                dets = dets[idx_keep]
            dets_final = np.array(dets_final).reshape(-1, 5)
318
            return dets_final
319 320 321 322 323

        def _soft_nms(bboxes, scores):
            bboxes = np.array(bboxes)
            scores = np.array(scores)
            class_nums = scores.shape[-1]
324

325 326 327
            softnms_thres = self.score_threshold
            softnms_sigma = self.softnms_sigma
            keep_top_k = self.keep_top_k
328

329 330
            cls_boxes = [[] for _ in range(class_nums)]
            cls_ids = [[] for _ in range(class_nums)]
331

332 333 334 335 336
            start_idx = 1 if self.background_label == 0 else 0
            for j in range(start_idx, class_nums):
                inds = np.where(scores[:, j] >= softnms_thres)[0]
                scores_j = scores[inds, j]
                rois_j = bboxes[inds, j, :]
337 338
                dets_j = np.hstack((scores_j[:, np.newaxis], rois_j)).astype(
                    np.float32, copy=False)
339 340 341
                cls_rank = np.argsort(-dets_j[:, 0])
                dets_j = dets_j[cls_rank]

342 343 344 345 346
                cls_boxes[j] = _soft_nms_for_cls(
                    dets_j, sigma=softnms_sigma, thres=softnms_thres)
                cls_ids[j] = np.array([j] * cls_boxes[j].shape[0]).reshape(-1,
                                                                           1)

347 348
            cls_boxes = np.vstack(cls_boxes[start_idx:])
            cls_ids = np.vstack(cls_ids[start_idx:])
349
            pred_result = np.hstack([cls_ids, cls_boxes])
350 351

            # Limit to max_per_image detections **over all classes**
352
            image_scores = cls_boxes[:, 0]
353 354 355 356
            if len(image_scores) > keep_top_k:
                image_thresh = np.sort(image_scores)[-keep_top_k]
                keep = np.where(cls_boxes[:, 0] >= image_thresh)[0]
                pred_result = pred_result[keep, :]
357

358 359 360
            res = fluid.LoDTensor()
            res.set_lod([[0, pred_result.shape[0]]])
            if pred_result.shape[0] == 0:
361
                pred_result = np.array([[1]], dtype=np.float32)
362
            res.set(pred_result, fluid.CPUPlace())
363

364
            return res
365 366 367 368 369 370

        pred_result = create_tmp_var(
            fluid.default_main_program(),
            name='softnms_pred_result',
            dtype='float32',
            shape=[6],
littletomatodonkey's avatar
littletomatodonkey 已提交
371
            lod_level=1)
372 373
        fluid.layers.py_func(
            func=_soft_nms, x=[bboxes, scores], out=pred_result)
374 375
        return pred_result

376

littletomatodonkey's avatar
littletomatodonkey 已提交
377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
@register
@serializable
class MultiClassDiouNMS(object):
    def __init__(
            self,
            score_threshold=0.05,
            keep_top_k=100,
            nms_threshold=0.5,
            normalized=False,
            background_label=0, ):
        super(MultiClassDiouNMS, self).__init__()
        self.score_threshold = score_threshold
        self.nms_threshold = nms_threshold
        self.keep_top_k = keep_top_k
        self.normalized = normalized
        self.background_label = background_label

    def __call__(self, bboxes, scores):
        def create_tmp_var(program, name, dtype, shape, lod_level):
            return program.current_block().create_var(
                name=name, dtype=dtype, shape=shape, lod_level=lod_level)

        def _calc_diou_term(dets1, dets2):
            eps = 1.e-10
            eta = 0 if self.normalized else 1

            x1, y1, x2, y2 = dets1[0], dets1[1], dets1[2], dets1[3]
            x1g, y1g, x2g, y2g = dets2[0], dets2[1], dets2[2], dets2[3]

            cx = (x1 + x2) / 2
            cy = (y1 + y2) / 2
            w = x2 - x1 + eta
            h = y2 - y1 + eta

            cxg = (x1g + x2g) / 2
            cyg = (y1g + y2g) / 2
            wg = x2g - x1g + eta
            hg = y2g - y1g + eta

            x2 = np.maximum(x1, x2)
            y2 = np.maximum(y1, y2)

            # A or B
            xc1 = np.minimum(x1, x1g)
            yc1 = np.minimum(y1, y1g)
            xc2 = np.maximum(x2, x2g)
            yc2 = np.maximum(y2, y2g)

            # DIOU term
            dist_intersection = (cx - cxg)**2 + (cy - cyg)**2
            dist_union = (xc2 - xc1)**2 + (yc2 - yc1)**2
            diou_term = (dist_intersection + eps) / (dist_union + eps)
            return diou_term

        def _diou_nms_for_cls(dets, thres):
            """_diou_nms_for_cls"""
            scores = dets[:, 0]
            x1 = dets[:, 1]
            y1 = dets[:, 2]
            x2 = dets[:, 3]
            y2 = dets[:, 4]
            eta = 0 if self.normalized else 1
            areas = (x2 - x1 + eta) * (y2 - y1 + eta)
            dt_num = dets.shape[0]
            order = np.array(range(dt_num))

            keep = []
            while order.size > 0:
                i = order[0]
                keep.append(i)
                xx1 = np.maximum(x1[i], x1[order[1:]])
                yy1 = np.maximum(y1[i], y1[order[1:]])
                xx2 = np.minimum(x2[i], x2[order[1:]])
                yy2 = np.minimum(y2[i], y2[order[1:]])

                w = np.maximum(0.0, xx2 - xx1 + eta)
                h = np.maximum(0.0, yy2 - yy1 + eta)
                inter = w * h
                ovr = inter / (areas[i] + areas[order[1:]] - inter)

                diou_term = _calc_diou_term([x1[i], y1[i], x2[i], y2[i]], [
                    x1[order[1:]], y1[order[1:]], x2[order[1:]], y2[order[1:]]
                ])

                inds = np.where(ovr - diou_term <= thres)[0]

                order = order[inds + 1]

            dets_final = dets[keep]
            return dets_final

        def _diou_nms(bboxes, scores):
            bboxes = np.array(bboxes)
            scores = np.array(scores)
            class_nums = scores.shape[-1]

            score_threshold = self.score_threshold
            nms_threshold = self.nms_threshold
            keep_top_k = self.keep_top_k

            cls_boxes = [[] for _ in range(class_nums)]
            cls_ids = [[] for _ in range(class_nums)]

            start_idx = 1 if self.background_label == 0 else 0
            for j in range(start_idx, class_nums):
                inds = np.where(scores[:, j] >= score_threshold)[0]
                scores_j = scores[inds, j]
                rois_j = bboxes[inds, j, :]
                dets_j = np.hstack((scores_j[:, np.newaxis], rois_j)).astype(
                    np.float32, copy=False)
                cls_rank = np.argsort(-dets_j[:, 0])
                dets_j = dets_j[cls_rank]

                cls_boxes[j] = _diou_nms_for_cls(dets_j, thres=nms_threshold)
                cls_ids[j] = np.array([j] * cls_boxes[j].shape[0]).reshape(-1,
                                                                           1)

            cls_boxes = np.vstack(cls_boxes[start_idx:])
            cls_ids = np.vstack(cls_ids[start_idx:])
            pred_result = np.hstack([cls_ids, cls_boxes])

            # Limit to max_per_image detections **over all classes**
            image_scores = cls_boxes[:, 0]
            if len(image_scores) > keep_top_k:
                image_thresh = np.sort(image_scores)[-keep_top_k]
                keep = np.where(cls_boxes[:, 0] >= image_thresh)[0]
                pred_result = pred_result[keep, :]

            res = fluid.LoDTensor()
            res.set_lod([[0, pred_result.shape[0]]])
            if pred_result.shape[0] == 0:
                pred_result = np.array([[1]], dtype=np.float32)
            res.set(pred_result, fluid.CPUPlace())

            return res

        pred_result = create_tmp_var(
            fluid.default_main_program(),
            name='diou_nms_pred_result',
            dtype='float32',
            shape=[6],
            lod_level=1)
        fluid.layers.py_func(
            func=_diou_nms, x=[bboxes, scores], out=pred_result)
        return pred_result


524 525 526 527
@register
class BBoxAssigner(object):
    __op__ = fluid.layers.generate_proposal_labels
    __append_doc__ = True
528
    __shared__ = ['num_classes']
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549

    def __init__(self,
                 batch_size_per_im=512,
                 fg_fraction=.25,
                 fg_thresh=.5,
                 bg_thresh_hi=.5,
                 bg_thresh_lo=0.,
                 bbox_reg_weights=[0.1, 0.1, 0.2, 0.2],
                 num_classes=81,
                 shuffle_before_sample=True):
        super(BBoxAssigner, self).__init__()
        self.batch_size_per_im = batch_size_per_im
        self.fg_fraction = fg_fraction
        self.fg_thresh = fg_thresh
        self.bg_thresh_hi = bg_thresh_hi
        self.bg_thresh_lo = bg_thresh_lo
        self.bbox_reg_weights = bbox_reg_weights
        self.class_nums = num_classes
        self.use_random = shuffle_before_sample


littletomatodonkey's avatar
littletomatodonkey 已提交
550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003
@register
class LibraBBoxAssigner(object):
    def __init__(self,
                 batch_size_per_im=512,
                 fg_fraction=.25,
                 fg_thresh=.5,
                 bg_thresh_hi=.5,
                 bg_thresh_lo=0.,
                 bbox_reg_weights=[0.1, 0.1, 0.2, 0.2],
                 num_classes=81,
                 shuffle_before_sample=True,
                 is_cls_agnostic=False,
                 num_bins=3):
        super(LibraBBoxAssigner, self).__init__()
        self.batch_size_per_im = batch_size_per_im
        self.fg_fraction = fg_fraction
        self.fg_thresh = fg_thresh
        self.bg_thresh_hi = bg_thresh_hi
        self.bg_thresh_lo = bg_thresh_lo
        self.bbox_reg_weights = bbox_reg_weights
        self.class_nums = num_classes
        self.use_random = shuffle_before_sample
        self.is_cls_agnostic = is_cls_agnostic
        self.num_bins = num_bins

    def __call__(
            self,
            rpn_rois,
            gt_classes,
            is_crowd,
            gt_boxes,
            im_info, ):
        return self.generate_proposal_label_libra(
            rpn_rois=rpn_rois,
            gt_classes=gt_classes,
            is_crowd=is_crowd,
            gt_boxes=gt_boxes,
            im_info=im_info,
            batch_size_per_im=self.batch_size_per_im,
            fg_fraction=self.fg_fraction,
            fg_thresh=self.fg_thresh,
            bg_thresh_hi=self.bg_thresh_hi,
            bg_thresh_lo=self.bg_thresh_lo,
            bbox_reg_weights=self.bbox_reg_weights,
            class_nums=self.class_nums,
            use_random=self.use_random,
            is_cls_agnostic=self.is_cls_agnostic,
            is_cascade_rcnn=False)

    def generate_proposal_label_libra(
            self, rpn_rois, gt_classes, is_crowd, gt_boxes, im_info,
            batch_size_per_im, fg_fraction, fg_thresh, bg_thresh_hi,
            bg_thresh_lo, bbox_reg_weights, class_nums, use_random,
            is_cls_agnostic, is_cascade_rcnn):
        num_bins = self.num_bins

        def create_tmp_var(program, name, dtype, shape, lod_level=None):
            return program.current_block().create_var(
                name=name, dtype=dtype, shape=shape, lod_level=lod_level)

        def _sample_pos(max_overlaps, max_classes, pos_inds, num_expected):
            if len(pos_inds) <= num_expected:
                return pos_inds
            else:
                unique_gt_inds = np.unique(max_classes[pos_inds])
                num_gts = len(unique_gt_inds)
                num_per_gt = int(round(num_expected / float(num_gts)) + 1)

                sampled_inds = []
                for i in unique_gt_inds:
                    inds = np.nonzero(max_classes == i)[0]
                    before_len = len(inds)
                    inds = list(set(inds) & set(pos_inds))
                    after_len = len(inds)
                    if len(inds) > num_per_gt:
                        inds = np.random.choice(
                            inds, size=num_per_gt, replace=False)
                    sampled_inds.extend(list(inds))  # combine as a new sampler
                if len(sampled_inds) < num_expected:
                    num_extra = num_expected - len(sampled_inds)
                    extra_inds = np.array(
                        list(set(pos_inds) - set(sampled_inds)))
                    assert len(sampled_inds)+len(extra_inds) == len(pos_inds), \
                        "sum of sampled_inds({}) and extra_inds({}) length must be equal with pos_inds({})!".format(
                            len(sampled_inds), len(extra_inds), len(pos_inds))
                    if len(extra_inds) > num_extra:
                        extra_inds = np.random.choice(
                            extra_inds, size=num_extra, replace=False)
                    sampled_inds.extend(extra_inds.tolist())
                elif len(sampled_inds) > num_expected:
                    sampled_inds = np.random.choice(
                        sampled_inds, size=num_expected, replace=False)
                return sampled_inds

        def sample_via_interval(max_overlaps, full_set, num_expected, floor_thr,
                                num_bins, bg_thresh_hi):
            max_iou = max_overlaps.max()
            iou_interval = (max_iou - floor_thr) / num_bins
            per_num_expected = int(num_expected / num_bins)

            sampled_inds = []
            for i in range(num_bins):
                start_iou = floor_thr + i * iou_interval
                end_iou = floor_thr + (i + 1) * iou_interval

                tmp_set = set(
                    np.where(
                        np.logical_and(max_overlaps >= start_iou, max_overlaps <
                                       end_iou))[0])
                tmp_inds = list(tmp_set & full_set)

                if len(tmp_inds) > per_num_expected:
                    tmp_sampled_set = np.random.choice(
                        tmp_inds, size=per_num_expected, replace=False)
                else:
                    tmp_sampled_set = np.array(tmp_inds, dtype=np.int)
                sampled_inds.append(tmp_sampled_set)

            sampled_inds = np.concatenate(sampled_inds)
            if len(sampled_inds) < num_expected:
                num_extra = num_expected - len(sampled_inds)
                extra_inds = np.array(list(full_set - set(sampled_inds)))
                assert len(sampled_inds)+len(extra_inds) == len(full_set), \
                    "sum of sampled_inds({}) and extra_inds({}) length must be equal with full_set({})!".format(
                            len(sampled_inds), len(extra_inds), len(full_set))

                if len(extra_inds) > num_extra:
                    extra_inds = np.random.choice(
                        extra_inds, num_extra, replace=False)
                sampled_inds = np.concatenate([sampled_inds, extra_inds])

            return sampled_inds

        def _sample_neg(max_overlaps,
                        max_classes,
                        neg_inds,
                        num_expected,
                        floor_thr=-1,
                        floor_fraction=0,
                        num_bins=3,
                        bg_thresh_hi=0.5):
            if len(neg_inds) <= num_expected:
                return neg_inds
            else:
                # balance sampling for negative samples
                neg_set = set(neg_inds)
                if floor_thr > 0:
                    floor_set = set(
                        np.where(
                            np.logical_and(max_overlaps >= 0, max_overlaps <
                                           floor_thr))[0])
                    iou_sampling_set = set(
                        np.where(max_overlaps >= floor_thr)[0])
                elif floor_thr == 0:
                    floor_set = set(np.where(max_overlaps == 0)[0])
                    iou_sampling_set = set(
                        np.where(max_overlaps > floor_thr)[0])
                else:
                    floor_set = set()
                    iou_sampling_set = set(
                        np.where(max_overlaps > floor_thr)[0])
                    floor_thr = 0

                floor_neg_inds = list(floor_set & neg_set)
                iou_sampling_neg_inds = list(iou_sampling_set & neg_set)

                num_expected_iou_sampling = int(num_expected *
                                                (1 - floor_fraction))
                if len(iou_sampling_neg_inds) > num_expected_iou_sampling:
                    if num_bins >= 2:
                        iou_sampled_inds = sample_via_interval(
                            max_overlaps,
                            set(iou_sampling_neg_inds),
                            num_expected_iou_sampling, floor_thr, num_bins,
                            bg_thresh_hi)
                    else:
                        iou_sampled_inds = np.random.choice(
                            iou_sampling_neg_inds,
                            size=num_expected_iou_sampling,
                            replace=False)
                else:
                    iou_sampled_inds = np.array(
                        iou_sampling_neg_inds, dtype=np.int)
                num_expected_floor = num_expected - len(iou_sampled_inds)
                if len(floor_neg_inds) > num_expected_floor:
                    sampled_floor_inds = np.random.choice(
                        floor_neg_inds, size=num_expected_floor, replace=False)
                else:
                    sampled_floor_inds = np.array(floor_neg_inds, dtype=np.int)
                sampled_inds = np.concatenate(
                    (sampled_floor_inds, iou_sampled_inds))
                if len(sampled_inds) < num_expected:
                    num_extra = num_expected - len(sampled_inds)
                    extra_inds = np.array(list(neg_set - set(sampled_inds)))
                    if len(extra_inds) > num_extra:
                        extra_inds = np.random.choice(
                            extra_inds, size=num_extra, replace=False)
                    sampled_inds = np.concatenate((sampled_inds, extra_inds))
                return sampled_inds

        def _sample_rois(rpn_rois, gt_classes, is_crowd, gt_boxes, im_info,
                         batch_size_per_im, fg_fraction, fg_thresh,
                         bg_thresh_hi, bg_thresh_lo, bbox_reg_weights,
                         class_nums, use_random, is_cls_agnostic,
                         is_cascade_rcnn):
            rois_per_image = int(batch_size_per_im)
            fg_rois_per_im = int(np.round(fg_fraction * rois_per_image))

            # Roidb
            im_scale = im_info[2]
            inv_im_scale = 1. / im_scale
            rpn_rois = rpn_rois * inv_im_scale
            if is_cascade_rcnn:
                rpn_rois = rpn_rois[gt_boxes.shape[0]:, :]
            boxes = np.vstack([gt_boxes, rpn_rois])
            gt_overlaps = np.zeros((boxes.shape[0], class_nums))
            box_to_gt_ind_map = np.zeros((boxes.shape[0]), dtype=np.int32)
            if len(gt_boxes) > 0:
                proposal_to_gt_overlaps = bbox_overlaps(boxes, gt_boxes)

                overlaps_argmax = proposal_to_gt_overlaps.argmax(axis=1)
                overlaps_max = proposal_to_gt_overlaps.max(axis=1)
                # Boxes which with non-zero overlap with gt boxes
                overlapped_boxes_ind = np.where(overlaps_max > 0)[0]

                overlapped_boxes_gt_classes = gt_classes[overlaps_argmax[
                    overlapped_boxes_ind]]

                for idx in range(len(overlapped_boxes_ind)):
                    gt_overlaps[overlapped_boxes_ind[
                        idx], overlapped_boxes_gt_classes[idx]] = overlaps_max[
                            overlapped_boxes_ind[idx]]
                    box_to_gt_ind_map[overlapped_boxes_ind[
                        idx]] = overlaps_argmax[overlapped_boxes_ind[idx]]

            crowd_ind = np.where(is_crowd)[0]
            gt_overlaps[crowd_ind] = -1

            max_overlaps = gt_overlaps.max(axis=1)
            max_classes = gt_overlaps.argmax(axis=1)

            # Cascade RCNN Decode Filter
            if is_cascade_rcnn:
                ws = boxes[:, 2] - boxes[:, 0] + 1
                hs = boxes[:, 3] - boxes[:, 1] + 1
                keep = np.where((ws > 0) & (hs > 0))[0]
                boxes = boxes[keep]
                fg_inds = np.where(max_overlaps >= fg_thresh)[0]
                bg_inds = np.where((max_overlaps < bg_thresh_hi) & (
                    max_overlaps >= bg_thresh_lo))[0]
                fg_rois_per_this_image = fg_inds.shape[0]
                bg_rois_per_this_image = bg_inds.shape[0]
            else:
                # Foreground
                fg_inds = np.where(max_overlaps >= fg_thresh)[0]
                fg_rois_per_this_image = np.minimum(fg_rois_per_im,
                                                    fg_inds.shape[0])
                # Sample foreground if there are too many
                if fg_inds.shape[0] > fg_rois_per_this_image:
                    if use_random:
                        fg_inds = _sample_pos(max_overlaps, max_classes,
                                              fg_inds, fg_rois_per_this_image)
                fg_inds = fg_inds[:fg_rois_per_this_image]

                # Background
                bg_inds = np.where((max_overlaps < bg_thresh_hi) & (
                    max_overlaps >= bg_thresh_lo))[0]
                bg_rois_per_this_image = rois_per_image - fg_rois_per_this_image
                bg_rois_per_this_image = np.minimum(bg_rois_per_this_image,
                                                    bg_inds.shape[0])
                assert bg_rois_per_this_image >= 0, "bg_rois_per_this_image must be >= 0 but got {}".format(
                    bg_rois_per_this_image)

                # Sample background if there are too many
                if bg_inds.shape[0] > bg_rois_per_this_image:
                    if use_random:
                        # libra neg sample
                        bg_inds = _sample_neg(
                            max_overlaps,
                            max_classes,
                            bg_inds,
                            bg_rois_per_this_image,
                            num_bins=num_bins,
                            bg_thresh_hi=bg_thresh_hi)
                bg_inds = bg_inds[:bg_rois_per_this_image]

            keep_inds = np.append(fg_inds, bg_inds)
            sampled_labels = max_classes[keep_inds]  # N x 1
            sampled_labels[fg_rois_per_this_image:] = 0
            sampled_boxes = boxes[keep_inds]  # N x 324
            sampled_gts = gt_boxes[box_to_gt_ind_map[keep_inds]]
            sampled_gts[fg_rois_per_this_image:, :] = gt_boxes[0]
            bbox_label_targets = _compute_targets(
                sampled_boxes, sampled_gts, sampled_labels, bbox_reg_weights)
            bbox_targets, bbox_inside_weights = _expand_bbox_targets(
                bbox_label_targets, class_nums, is_cls_agnostic)
            bbox_outside_weights = np.array(
                bbox_inside_weights > 0, dtype=bbox_inside_weights.dtype)
            # Scale rois
            sampled_rois = sampled_boxes * im_scale

            # Faster RCNN blobs
            frcn_blobs = dict(
                rois=sampled_rois,
                labels_int32=sampled_labels,
                bbox_targets=bbox_targets,
                bbox_inside_weights=bbox_inside_weights,
                bbox_outside_weights=bbox_outside_weights)
            return frcn_blobs

        def _compute_targets(roi_boxes, gt_boxes, labels, bbox_reg_weights):
            assert roi_boxes.shape[0] == gt_boxes.shape[0]
            assert roi_boxes.shape[1] == 4
            assert gt_boxes.shape[1] == 4

            targets = np.zeros(roi_boxes.shape)
            bbox_reg_weights = np.asarray(bbox_reg_weights)
            targets = box_to_delta(
                ex_boxes=roi_boxes, gt_boxes=gt_boxes, weights=bbox_reg_weights)

            return np.hstack([labels[:, np.newaxis], targets]).astype(
                np.float32, copy=False)

        def _expand_bbox_targets(bbox_targets_input, class_nums,
                                 is_cls_agnostic):
            class_labels = bbox_targets_input[:, 0]
            fg_inds = np.where(class_labels > 0)[0]
            bbox_targets = np.zeros((class_labels.shape[0], 4 * class_nums
                                     if not is_cls_agnostic else 4 * 2))
            bbox_inside_weights = np.zeros(bbox_targets.shape)
            for ind in fg_inds:
                class_label = int(class_labels[
                    ind]) if not is_cls_agnostic else 1
                start_ind = class_label * 4
                end_ind = class_label * 4 + 4
                bbox_targets[ind, start_ind:end_ind] = bbox_targets_input[ind,
                                                                          1:]
                bbox_inside_weights[ind, start_ind:end_ind] = (1.0, 1.0, 1.0,
                                                               1.0)
            return bbox_targets, bbox_inside_weights

        def generate_func(
                rpn_rois,
                gt_classes,
                is_crowd,
                gt_boxes,
                im_info, ):
            rpn_rois_lod = rpn_rois.lod()[0]
            gt_classes_lod = gt_classes.lod()[0]

            # convert
            rpn_rois = np.array(rpn_rois)
            gt_classes = np.array(gt_classes)
            is_crowd = np.array(is_crowd)
            gt_boxes = np.array(gt_boxes)
            im_info = np.array(im_info)

            rois = []
            labels_int32 = []
            bbox_targets = []
            bbox_inside_weights = []
            bbox_outside_weights = []
            lod = [0]

            for idx in range(len(rpn_rois_lod) - 1):
                rois_si = rpn_rois_lod[idx]
                rois_ei = rpn_rois_lod[idx + 1]

                gt_si = gt_classes_lod[idx]
                gt_ei = gt_classes_lod[idx + 1]
                frcn_blobs = _sample_rois(
                    rpn_rois[rois_si:rois_ei], gt_classes[gt_si:gt_ei],
                    is_crowd[gt_si:gt_ei], gt_boxes[gt_si:gt_ei], im_info[idx],
                    batch_size_per_im, fg_fraction, fg_thresh, bg_thresh_hi,
                    bg_thresh_lo, bbox_reg_weights, class_nums, use_random,
                    is_cls_agnostic, is_cascade_rcnn)
                lod.append(frcn_blobs['rois'].shape[0] + lod[-1])
                rois.append(frcn_blobs['rois'])
                labels_int32.append(frcn_blobs['labels_int32'].reshape(-1, 1))
                bbox_targets.append(frcn_blobs['bbox_targets'])
                bbox_inside_weights.append(frcn_blobs['bbox_inside_weights'])
                bbox_outside_weights.append(frcn_blobs['bbox_outside_weights'])

            rois = np.vstack(rois)
            labels_int32 = np.vstack(labels_int32)
            bbox_targets = np.vstack(bbox_targets)
            bbox_inside_weights = np.vstack(bbox_inside_weights)
            bbox_outside_weights = np.vstack(bbox_outside_weights)

            # create lod-tensor for return
            # notice that the func create_lod_tensor does not work well here
            ret_rois = fluid.LoDTensor()
            ret_rois.set_lod([lod])
            ret_rois.set(rois.astype("float32"), fluid.CPUPlace())

            ret_labels_int32 = fluid.LoDTensor()
            ret_labels_int32.set_lod([lod])
            ret_labels_int32.set(labels_int32.astype("int32"), fluid.CPUPlace())

            ret_bbox_targets = fluid.LoDTensor()
            ret_bbox_targets.set_lod([lod])
            ret_bbox_targets.set(
                bbox_targets.astype("float32"), fluid.CPUPlace())

            ret_bbox_inside_weights = fluid.LoDTensor()
            ret_bbox_inside_weights.set_lod([lod])
            ret_bbox_inside_weights.set(
                bbox_inside_weights.astype("float32"), fluid.CPUPlace())

            ret_bbox_outside_weights = fluid.LoDTensor()
            ret_bbox_outside_weights.set_lod([lod])
            ret_bbox_outside_weights.set(
                bbox_outside_weights.astype("float32"), fluid.CPUPlace())

            return ret_rois, ret_labels_int32, ret_bbox_targets, ret_bbox_inside_weights, ret_bbox_outside_weights

        rois = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'rois', 
            dtype='float32',
            shape=[-1, 4], )
        bbox_inside_weights = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'bbox_inside_weights', 
            dtype='float32',
            shape=[-1, 8 if self.is_cls_agnostic else self.class_nums * 4], )
        bbox_outside_weights = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'bbox_outside_weights', 
            dtype='float32',
            shape=[-1, 8 if self.is_cls_agnostic else self.class_nums * 4], )
        bbox_targets = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'bbox_targets', 
            dtype='float32',
            shape=[-1, 8 if self.is_cls_agnostic else self.class_nums * 4], )
        labels_int32 = create_tmp_var(
            fluid.default_main_program(),
            name=None,  #'labels_int32', 
            dtype='int32',
            shape=[-1, 1], )

        outs = [
            rois, labels_int32, bbox_targets, bbox_inside_weights,
            bbox_outside_weights
        ]

        fluid.layers.py_func(
            func=generate_func,
            x=[rpn_rois, gt_classes, is_crowd, gt_boxes, im_info],
            out=outs)
        return outs


1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
@register
class RoIAlign(object):
    __op__ = fluid.layers.roi_align
    __append_doc__ = True

    def __init__(self, resolution=7, spatial_scale=1. / 16, sampling_ratio=0):
        super(RoIAlign, self).__init__()
        if isinstance(resolution, Integral):
            resolution = [resolution, resolution]
        self.pooled_height = resolution[0]
        self.pooled_width = resolution[1]
        self.spatial_scale = spatial_scale
        self.sampling_ratio = sampling_ratio


@register
class RoIPool(object):
    __op__ = fluid.layers.roi_pool
    __append_doc__ = True

    def __init__(self, resolution=7, spatial_scale=1. / 16):
        super(RoIPool, self).__init__()
        if isinstance(resolution, Integral):
            resolution = [resolution, resolution]
        self.pooled_height = resolution[0]
        self.pooled_width = resolution[1]
        self.spatial_scale = spatial_scale


@register
class MultiBoxHead(object):
    __op__ = fluid.layers.multi_box_head
    __append_doc__ = True

    def __init__(self,
                 min_ratio=20,
                 max_ratio=90,
1041
                 base_size=300,
1042 1043 1044 1045
                 min_sizes=[60.0, 105.0, 150.0, 195.0, 240.0, 285.0],
                 max_sizes=[[], 150.0, 195.0, 240.0, 285.0, 300.0],
                 aspect_ratios=[[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.],
                                [2., 3.]],
1046
                 steps=None,
1047
                 offset=0.5,
1048 1049 1050 1051
                 flip=True,
                 min_max_aspect_ratios_order=False,
                 kernel_size=1,
                 pad=0):
1052 1053 1054
        super(MultiBoxHead, self).__init__()
        self.min_ratio = min_ratio
        self.max_ratio = max_ratio
1055
        self.base_size = base_size
1056 1057 1058
        self.min_sizes = min_sizes
        self.max_sizes = max_sizes
        self.aspect_ratios = aspect_ratios
1059
        self.steps = steps
1060 1061
        self.offset = offset
        self.flip = flip
1062 1063 1064
        self.min_max_aspect_ratios_order = min_max_aspect_ratios_order
        self.kernel_size = kernel_size
        self.pad = pad
1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118


@register
@serializable
class SSDOutputDecoder(object):
    __op__ = fluid.layers.detection_output
    __append_doc__ = True

    def __init__(self,
                 nms_threshold=0.45,
                 nms_top_k=400,
                 keep_top_k=200,
                 score_threshold=0.01,
                 nms_eta=1.0,
                 background_label=0):
        super(SSDOutputDecoder, self).__init__()
        self.nms_threshold = nms_threshold
        self.background_label = background_label
        self.nms_top_k = nms_top_k
        self.keep_top_k = keep_top_k
        self.score_threshold = score_threshold
        self.nms_eta = nms_eta


@register
@serializable
class RetinaTargetAssign(object):
    __op__ = fluid.layers.retinanet_target_assign
    __append_doc__ = True

    def __init__(self, positive_overlap=0.5, negative_overlap=0.4):
        super(RetinaTargetAssign, self).__init__()
        self.positive_overlap = positive_overlap
        self.negative_overlap = negative_overlap


@register
@serializable
class RetinaOutputDecoder(object):
    __op__ = fluid.layers.retinanet_detection_output
    __append_doc__ = True

    def __init__(self,
                 score_thresh=0.05,
                 nms_thresh=0.3,
                 pre_nms_top_n=1000,
                 detections_per_im=100,
                 nms_eta=1.0):
        super(RetinaOutputDecoder, self).__init__()
        self.score_threshold = score_thresh
        self.nms_threshold = nms_thresh
        self.nms_top_k = pre_nms_top_n
        self.keep_top_k = detections_per_im
        self.nms_eta = nms_eta