ops.py 10.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
# 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.

from numbers import Integral

from paddle import fluid
Y
Yuan Gao 已提交
18 19
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.regularizer import L2Decay
20 21 22 23 24
from ppdet.core.workspace import register, serializable

__all__ = [
    'AnchorGenerator', 'RPNTargetAssign', 'GenerateProposals', 'MultiClassNMS',
    'BBoxAssigner', 'MaskAssigner', 'RoIAlign', 'RoIPool', 'MultiBoxHead',
25
    'SSDOutputDecoder', 'RetinaTargetAssign', 'RetinaOutputDecoder', 'ConvNorm'
26 27 28
]


Y
Yuan Gao 已提交
29 30 31 32 33 34 35
def ConvNorm(input,
             num_filters,
             filter_size,
             stride=1,
             groups=1,
             norm_decay=0.,
             norm_type='affine_channel',
Y
Yuan Gao 已提交
36 37
             norm_groups=32,
             dilation=1,
Y
Yuan Gao 已提交
38 39
             freeze_norm=False,
             act=None,
Y
Yuan Gao 已提交
40
             norm_name=None,
Y
Yuan Gao 已提交
41 42 43 44 45 46 47 48
             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
49 50
        padding=((filter_size - 1) // 2) * dilation,
        dilation=dilation,
Y
Yuan Gao 已提交
51 52 53 54 55 56 57 58 59
        groups=groups,
        act=None,
        param_attr=ParamAttr(
            name=name + "_weights", initializer=initializer),
        bias_attr=False,
        name=name + '.conv2d.output.1')

    norm_lr = 0. if freeze_norm else 1.
    pattr = ParamAttr(
Y
Yuan Gao 已提交
60
        name=norm_name + '_scale',
Y
Yuan Gao 已提交
61 62 63
        learning_rate=norm_lr,
        regularizer=L2Decay(norm_decay))
    battr = ParamAttr(
Y
Yuan Gao 已提交
64
        name=norm_name + '_offset',
Y
Yuan Gao 已提交
65 66 67 68 69 70 71 72
        learning_rate=norm_lr,
        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
73
            name=norm_name + '.output.1',
Y
Yuan Gao 已提交
74 75
            param_attr=pattr,
            bias_attr=battr,
Y
Yuan Gao 已提交
76 77
            moving_mean_name=norm_name + '_mean',
            moving_variance_name=norm_name + '_variance',
Y
Yuan Gao 已提交
78 79 80
            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
81 82 83 84 85 86 87 88 89 90
    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
91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
    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


110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173
@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
174
    __shared__ = ['num_classes']
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

    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


@register
class BBoxAssigner(object):
    __op__ = fluid.layers.generate_proposal_labels
    __append_doc__ = True
210
    __shared__ = ['num_classes']
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268

    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


@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,
269
                 base_size=300,
270 271 272 273
                 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.]],
274
                 steps=None,
275
                 offset=0.5,
276 277 278 279
                 flip=True,
                 min_max_aspect_ratios_order=False,
                 kernel_size=1,
                 pad=0):
280 281 282
        super(MultiBoxHead, self).__init__()
        self.min_ratio = min_ratio
        self.max_ratio = max_ratio
283
        self.base_size = base_size
284 285 286
        self.min_sizes = min_sizes
        self.max_sizes = max_sizes
        self.aspect_ratios = aspect_ratios
287
        self.steps = steps
288 289
        self.offset = offset
        self.flip = flip
290 291 292
        self.min_max_aspect_ratios_order = min_max_aspect_ratios_order
        self.kernel_size = kernel_size
        self.pad = pad
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346


@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