hrnet.py 23.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
# Copyright (c) 2020 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.

import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.regularizer import L2Decay
from paddle import ParamAttr
from paddle.nn.initializer import Normal
from numbers import Integral
import math

from ppdet.core.workspace import register, serializable
25
from ..shape_spec import ShapeSpec
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99

__all__ = ['HRNet']


class ConvNormLayer(nn.Layer):
    def __init__(self,
                 ch_in,
                 ch_out,
                 filter_size,
                 stride=1,
                 norm_type='bn',
                 norm_groups=32,
                 use_dcn=False,
                 norm_decay=0.,
                 freeze_norm=False,
                 act=None,
                 name=None):
        super(ConvNormLayer, self).__init__()
        assert norm_type in ['bn', 'sync_bn', 'gn']

        self.act = act
        self.conv = nn.Conv2D(
            in_channels=ch_in,
            out_channels=ch_out,
            kernel_size=filter_size,
            stride=stride,
            padding=(filter_size - 1) // 2,
            groups=1,
            weight_attr=ParamAttr(
                name=name + "_weights", initializer=Normal(
                    mean=0., std=0.01)),
            bias_attr=False)

        norm_lr = 0. if freeze_norm else 1.

        norm_name = name + '_bn'
        param_attr = ParamAttr(
            name=norm_name + "_scale",
            learning_rate=norm_lr,
            regularizer=L2Decay(norm_decay))
        bias_attr = ParamAttr(
            name=norm_name + "_offset",
            learning_rate=norm_lr,
            regularizer=L2Decay(norm_decay))
        global_stats = True if freeze_norm else False
        if norm_type in ['bn', 'sync_bn']:
            self.norm = nn.BatchNorm(
                ch_out,
                param_attr=param_attr,
                bias_attr=bias_attr,
                use_global_stats=global_stats,
                moving_mean_name=norm_name + '_mean',
                moving_variance_name=norm_name + '_variance')
        elif norm_type == 'gn':
            self.norm = nn.GroupNorm(
                num_groups=norm_groups,
                num_channels=ch_out,
                weight_attr=param_attr,
                bias_attr=bias_attr)
        norm_params = self.norm.parameters()
        if freeze_norm:
            for param in norm_params:
                param.stop_gradient = True

    def forward(self, inputs):
        out = self.conv(inputs)
        out = self.norm(out)

        if self.act == 'relu':
            out = F.relu(out)
        return out


class Layer1(nn.Layer):
F
Feng Ni 已提交
100 101 102 103 104 105
    def __init__(self,
                 num_channels,
                 has_se=False,
                 norm_decay=0.,
                 freeze_norm=True,
                 name=None):
106 107 108 109 110 111 112 113 114 115 116 117 118
        super(Layer1, self).__init__()

        self.bottleneck_block_list = []

        for i in range(4):
            bottleneck_block = self.add_sublayer(
                "block_{}_{}".format(name, i + 1),
                BottleneckBlock(
                    num_channels=num_channels if i == 0 else 256,
                    num_filters=64,
                    has_se=has_se,
                    stride=1,
                    downsample=True if i == 0 else False,
F
Feng Ni 已提交
119
                    norm_decay=norm_decay,
120 121 122 123 124 125 126 127 128 129 130 131
                    freeze_norm=freeze_norm,
                    name=name + '_' + str(i + 1)))
            self.bottleneck_block_list.append(bottleneck_block)

    def forward(self, input):
        conv = input
        for block_func in self.bottleneck_block_list:
            conv = block_func(conv)
        return conv


class TransitionLayer(nn.Layer):
F
Feng Ni 已提交
132 133 134 135 136 137
    def __init__(self,
                 in_channels,
                 out_channels,
                 norm_decay=0.,
                 freeze_norm=True,
                 name=None):
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
        super(TransitionLayer, self).__init__()

        num_in = len(in_channels)
        num_out = len(out_channels)
        out = []
        self.conv_bn_func_list = []
        for i in range(num_out):
            residual = None
            if i < num_in:
                if in_channels[i] != out_channels[i]:
                    residual = self.add_sublayer(
                        "transition_{}_layer_{}".format(name, i + 1),
                        ConvNormLayer(
                            ch_in=in_channels[i],
                            ch_out=out_channels[i],
                            filter_size=3,
F
Feng Ni 已提交
154
                            norm_decay=norm_decay,
155 156 157 158 159 160 161 162 163 164 165
                            freeze_norm=freeze_norm,
                            act='relu',
                            name=name + '_layer_' + str(i + 1)))
            else:
                residual = self.add_sublayer(
                    "transition_{}_layer_{}".format(name, i + 1),
                    ConvNormLayer(
                        ch_in=in_channels[-1],
                        ch_out=out_channels[i],
                        filter_size=3,
                        stride=2,
F
Feng Ni 已提交
166
                        norm_decay=norm_decay,
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
                        freeze_norm=freeze_norm,
                        act='relu',
                        name=name + '_layer_' + str(i + 1)))
            self.conv_bn_func_list.append(residual)

    def forward(self, input):
        outs = []
        for idx, conv_bn_func in enumerate(self.conv_bn_func_list):
            if conv_bn_func is None:
                outs.append(input[idx])
            else:
                if idx < len(input):
                    outs.append(conv_bn_func(input[idx]))
                else:
                    outs.append(conv_bn_func(input[-1]))
        return outs


class Branches(nn.Layer):
    def __init__(self,
                 block_num,
                 in_channels,
                 out_channels,
                 has_se=False,
F
Feng Ni 已提交
191
                 norm_decay=0.,
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
                 freeze_norm=True,
                 name=None):
        super(Branches, self).__init__()

        self.basic_block_list = []
        for i in range(len(out_channels)):
            self.basic_block_list.append([])
            for j in range(block_num):
                in_ch = in_channels[i] if j == 0 else out_channels[i]
                basic_block_func = self.add_sublayer(
                    "bb_{}_branch_layer_{}_{}".format(name, i + 1, j + 1),
                    BasicBlock(
                        num_channels=in_ch,
                        num_filters=out_channels[i],
                        has_se=has_se,
F
Feng Ni 已提交
207
                        norm_decay=norm_decay,
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230
                        freeze_norm=freeze_norm,
                        name=name + '_branch_layer_' + str(i + 1) + '_' +
                        str(j + 1)))
                self.basic_block_list[i].append(basic_block_func)

    def forward(self, inputs):
        outs = []
        for idx, input in enumerate(inputs):
            conv = input
            basic_block_list = self.basic_block_list[idx]
            for basic_block_func in basic_block_list:
                conv = basic_block_func(conv)
            outs.append(conv)
        return outs


class BottleneckBlock(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_filters,
                 has_se,
                 stride=1,
                 downsample=False,
F
Feng Ni 已提交
231
                 norm_decay=0.,
232 233 234 235 236 237 238 239 240 241 242
                 freeze_norm=True,
                 name=None):
        super(BottleneckBlock, self).__init__()

        self.has_se = has_se
        self.downsample = downsample

        self.conv1 = ConvNormLayer(
            ch_in=num_channels,
            ch_out=num_filters,
            filter_size=1,
F
Feng Ni 已提交
243
            norm_decay=norm_decay,
244 245 246 247 248 249 250 251
            freeze_norm=freeze_norm,
            act="relu",
            name=name + "_conv1")
        self.conv2 = ConvNormLayer(
            ch_in=num_filters,
            ch_out=num_filters,
            filter_size=3,
            stride=stride,
F
Feng Ni 已提交
252
            norm_decay=norm_decay,
253 254 255 256 257 258 259
            freeze_norm=freeze_norm,
            act="relu",
            name=name + "_conv2")
        self.conv3 = ConvNormLayer(
            ch_in=num_filters,
            ch_out=num_filters * 4,
            filter_size=1,
F
Feng Ni 已提交
260
            norm_decay=norm_decay,
261 262 263 264 265 266 267 268 269
            freeze_norm=freeze_norm,
            act=None,
            name=name + "_conv3")

        if self.downsample:
            self.conv_down = ConvNormLayer(
                ch_in=num_channels,
                ch_out=num_filters * 4,
                filter_size=1,
F
Feng Ni 已提交
270
                norm_decay=norm_decay,
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
                freeze_norm=freeze_norm,
                act=None,
                name=name + "_downsample")

        if self.has_se:
            self.se = SELayer(
                num_channels=num_filters * 4,
                num_filters=num_filters * 4,
                reduction_ratio=16,
                name='fc' + name)

    def forward(self, input):
        residual = input
        conv1 = self.conv1(input)
        conv2 = self.conv2(conv1)
        conv3 = self.conv3(conv2)

        if self.downsample:
            residual = self.conv_down(input)

        if self.has_se:
            conv3 = self.se(conv3)

        y = paddle.add(x=residual, y=conv3)
        y = F.relu(y)
        return y


class BasicBlock(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_filters,
                 stride=1,
                 has_se=False,
                 downsample=False,
F
Feng Ni 已提交
306
                 norm_decay=0.,
307 308 309 310 311 312 313 314 315 316
                 freeze_norm=True,
                 name=None):
        super(BasicBlock, self).__init__()

        self.has_se = has_se
        self.downsample = downsample
        self.conv1 = ConvNormLayer(
            ch_in=num_channels,
            ch_out=num_filters,
            filter_size=3,
F
Feng Ni 已提交
317
            norm_decay=norm_decay,
318 319 320 321 322 323 324 325
            freeze_norm=freeze_norm,
            stride=stride,
            act="relu",
            name=name + "_conv1")
        self.conv2 = ConvNormLayer(
            ch_in=num_filters,
            ch_out=num_filters,
            filter_size=3,
F
Feng Ni 已提交
326
            norm_decay=norm_decay,
327 328 329 330 331 332 333 334 335 336
            freeze_norm=freeze_norm,
            stride=1,
            act=None,
            name=name + "_conv2")

        if self.downsample:
            self.conv_down = ConvNormLayer(
                ch_in=num_channels,
                ch_out=num_filters * 4,
                filter_size=1,
F
Feng Ni 已提交
337
                norm_decay=norm_decay,
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
                freeze_norm=freeze_norm,
                act=None,
                name=name + "_downsample")

        if self.has_se:
            self.se = SELayer(
                num_channels=num_filters,
                num_filters=num_filters,
                reduction_ratio=16,
                name='fc' + name)

    def forward(self, input):
        residual = input
        conv1 = self.conv1(input)
        conv2 = self.conv2(conv1)

        if self.downsample:
            residual = self.conv_down(input)

        if self.has_se:
            conv2 = self.se(conv2)

        y = paddle.add(x=residual, y=conv2)
        y = F.relu(y)
        return y


class SELayer(nn.Layer):
    def __init__(self, num_channels, num_filters, reduction_ratio, name=None):
        super(SELayer, self).__init__()

        self.pool2d_gap = AdaptiveAvgPool2D(1)

        self._num_channels = num_channels

        med_ch = int(num_channels / reduction_ratio)
        stdv = 1.0 / math.sqrt(num_channels * 1.0)
        self.squeeze = Linear(
            num_channels,
            med_ch,
            weight_attr=ParamAttr(
                initializer=Uniform(-stdv, stdv), name=name + "_sqz_weights"),
            bias_attr=ParamAttr(name=name + '_sqz_offset'))

        stdv = 1.0 / math.sqrt(med_ch * 1.0)
        self.excitation = Linear(
            med_ch,
            num_filters,
            weight_attr=ParamAttr(
                initializer=Uniform(-stdv, stdv), name=name + "_exc_weights"),
            bias_attr=ParamAttr(name=name + '_exc_offset'))

    def forward(self, input):
        pool = self.pool2d_gap(input)
        pool = paddle.squeeze(pool, axis=[2, 3])
        squeeze = self.squeeze(pool)
        squeeze = F.relu(squeeze)
        excitation = self.excitation(squeeze)
        excitation = F.sigmoid(excitation)
        excitation = paddle.unsqueeze(excitation, axis=[2, 3])
        out = input * excitation
        return out


class Stage(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_modules,
                 num_filters,
                 has_se=False,
F
Feng Ni 已提交
408
                 norm_decay=0.,
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423
                 freeze_norm=True,
                 multi_scale_output=True,
                 name=None):
        super(Stage, self).__init__()

        self._num_modules = num_modules
        self.stage_func_list = []
        for i in range(num_modules):
            if i == num_modules - 1 and not multi_scale_output:
                stage_func = self.add_sublayer(
                    "stage_{}_{}".format(name, i + 1),
                    HighResolutionModule(
                        num_channels=num_channels,
                        num_filters=num_filters,
                        has_se=has_se,
F
Feng Ni 已提交
424
                        norm_decay=norm_decay,
425 426 427 428 429 430 431 432 433 434
                        freeze_norm=freeze_norm,
                        multi_scale_output=False,
                        name=name + '_' + str(i + 1)))
            else:
                stage_func = self.add_sublayer(
                    "stage_{}_{}".format(name, i + 1),
                    HighResolutionModule(
                        num_channels=num_channels,
                        num_filters=num_filters,
                        has_se=has_se,
F
Feng Ni 已提交
435
                        norm_decay=norm_decay,
436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453
                        freeze_norm=freeze_norm,
                        name=name + '_' + str(i + 1)))

            self.stage_func_list.append(stage_func)

    def forward(self, input):
        out = input
        for idx in range(self._num_modules):
            out = self.stage_func_list[idx](out)
        return out


class HighResolutionModule(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_filters,
                 has_se=False,
                 multi_scale_output=True,
F
Feng Ni 已提交
454
                 norm_decay=0.,
455 456 457 458 459 460 461 462
                 freeze_norm=True,
                 name=None):
        super(HighResolutionModule, self).__init__()
        self.branches_func = Branches(
            block_num=4,
            in_channels=num_channels,
            out_channels=num_filters,
            has_se=has_se,
F
Feng Ni 已提交
463
            norm_decay=norm_decay,
464 465 466 467 468 469 470
            freeze_norm=freeze_norm,
            name=name)

        self.fuse_func = FuseLayers(
            in_channels=num_filters,
            out_channels=num_filters,
            multi_scale_output=multi_scale_output,
F
Feng Ni 已提交
471
            norm_decay=norm_decay,
472 473 474 475 476 477 478 479 480 481 482 483 484 485
            freeze_norm=freeze_norm,
            name=name)

    def forward(self, input):
        out = self.branches_func(input)
        out = self.fuse_func(out)
        return out


class FuseLayers(nn.Layer):
    def __init__(self,
                 in_channels,
                 out_channels,
                 multi_scale_output=True,
F
Feng Ni 已提交
486
                 norm_decay=0.,
487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506
                 freeze_norm=True,
                 name=None):
        super(FuseLayers, self).__init__()

        self._actual_ch = len(in_channels) if multi_scale_output else 1
        self._in_channels = in_channels

        self.residual_func_list = []
        for i in range(self._actual_ch):
            for j in range(len(in_channels)):
                residual_func = None
                if j > i:
                    residual_func = self.add_sublayer(
                        "residual_{}_layer_{}_{}".format(name, i + 1, j + 1),
                        ConvNormLayer(
                            ch_in=in_channels[j],
                            ch_out=out_channels[i],
                            filter_size=1,
                            stride=1,
                            act=None,
F
Feng Ni 已提交
507
                            norm_decay=norm_decay,
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
                            freeze_norm=freeze_norm,
                            name=name + '_layer_' + str(i + 1) + '_' +
                            str(j + 1)))
                    self.residual_func_list.append(residual_func)
                elif j < i:
                    pre_num_filters = in_channels[j]
                    for k in range(i - j):
                        if k == i - j - 1:
                            residual_func = self.add_sublayer(
                                "residual_{}_layer_{}_{}_{}".format(
                                    name, i + 1, j + 1, k + 1),
                                ConvNormLayer(
                                    ch_in=pre_num_filters,
                                    ch_out=out_channels[i],
                                    filter_size=3,
                                    stride=2,
F
Feng Ni 已提交
524
                                    norm_decay=norm_decay,
525 526 527 528 529 530 531 532 533 534 535 536 537 538
                                    freeze_norm=freeze_norm,
                                    act=None,
                                    name=name + '_layer_' + str(i + 1) + '_' +
                                    str(j + 1) + '_' + str(k + 1)))
                            pre_num_filters = out_channels[i]
                        else:
                            residual_func = self.add_sublayer(
                                "residual_{}_layer_{}_{}_{}".format(
                                    name, i + 1, j + 1, k + 1),
                                ConvNormLayer(
                                    ch_in=pre_num_filters,
                                    ch_out=out_channels[j],
                                    filter_size=3,
                                    stride=2,
F
Feng Ni 已提交
539
                                    norm_decay=norm_decay,
540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580
                                    freeze_norm=freeze_norm,
                                    act="relu",
                                    name=name + '_layer_' + str(i + 1) + '_' +
                                    str(j + 1) + '_' + str(k + 1)))
                            pre_num_filters = out_channels[j]
                        self.residual_func_list.append(residual_func)

    def forward(self, input):
        outs = []
        residual_func_idx = 0
        for i in range(self._actual_ch):
            residual = input[i]
            for j in range(len(self._in_channels)):
                if j > i:
                    y = self.residual_func_list[residual_func_idx](input[j])
                    residual_func_idx += 1
                    y = F.interpolate(y, scale_factor=2**(j - i))
                    residual = paddle.add(x=residual, y=y)
                elif j < i:
                    y = input[j]
                    for k in range(i - j):
                        y = self.residual_func_list[residual_func_idx](y)
                        residual_func_idx += 1

                    residual = paddle.add(x=residual, y=y)
            residual = F.relu(residual)
            outs.append(residual)

        return outs


@register
class HRNet(nn.Layer):
    """
    HRNet, see https://arxiv.org/abs/1908.07919

    Args:
        width (int): the width of HRNet
        has_se (bool): whether to add SE block for each stage
        freeze_at (int): the stage to freeze
        freeze_norm (bool): whether to freeze norm in HRNet
F
Feng Ni 已提交
581
        norm_decay (float): weight decay for normalization layer weights
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
        return_idx (List): the stage to return
    """

    def __init__(self,
                 width=18,
                 has_se=False,
                 freeze_at=0,
                 freeze_norm=True,
                 norm_decay=0.,
                 return_idx=[0, 1, 2, 3]):
        super(HRNet, self).__init__()

        self.width = width
        self.has_se = has_se
        if isinstance(return_idx, Integral):
            return_idx = [return_idx]

        assert len(return_idx) > 0, "need one or more return index"
        self.freeze_at = freeze_at
        self.return_idx = return_idx

        self.channels = {
            18: [[18, 36], [18, 36, 72], [18, 36, 72, 144]],
            30: [[30, 60], [30, 60, 120], [30, 60, 120, 240]],
            32: [[32, 64], [32, 64, 128], [32, 64, 128, 256]],
            40: [[40, 80], [40, 80, 160], [40, 80, 160, 320]],
            44: [[44, 88], [44, 88, 176], [44, 88, 176, 352]],
            48: [[48, 96], [48, 96, 192], [48, 96, 192, 384]],
            60: [[60, 120], [60, 120, 240], [60, 120, 240, 480]],
            64: [[64, 128], [64, 128, 256], [64, 128, 256, 512]]
        }

        channels_2, channels_3, channels_4 = self.channels[width]
        num_modules_2, num_modules_3, num_modules_4 = 1, 4, 3
616 617
        self._out_channels = channels_4
        self._out_strides = [4, 8, 16, 32]
618 619 620 621 622 623

        self.conv_layer1_1 = ConvNormLayer(
            ch_in=3,
            ch_out=64,
            filter_size=3,
            stride=2,
F
Feng Ni 已提交
624
            norm_decay=norm_decay,
625 626 627 628 629 630 631 632 633
            freeze_norm=freeze_norm,
            act='relu',
            name="layer1_1")

        self.conv_layer1_2 = ConvNormLayer(
            ch_in=64,
            ch_out=64,
            filter_size=3,
            stride=2,
F
Feng Ni 已提交
634
            norm_decay=norm_decay,
635 636 637 638 639 640 641
            freeze_norm=freeze_norm,
            act='relu',
            name="layer1_2")

        self.la1 = Layer1(
            num_channels=64,
            has_se=has_se,
F
Feng Ni 已提交
642
            norm_decay=norm_decay,
643 644 645 646 647 648
            freeze_norm=freeze_norm,
            name="layer2")

        self.tr1 = TransitionLayer(
            in_channels=[256],
            out_channels=channels_2,
F
Feng Ni 已提交
649
            norm_decay=norm_decay,
650 651 652 653 654 655 656 657
            freeze_norm=freeze_norm,
            name="tr1")

        self.st2 = Stage(
            num_channels=channels_2,
            num_modules=num_modules_2,
            num_filters=channels_2,
            has_se=self.has_se,
F
Feng Ni 已提交
658
            norm_decay=norm_decay,
659 660 661 662 663 664
            freeze_norm=freeze_norm,
            name="st2")

        self.tr2 = TransitionLayer(
            in_channels=channels_2,
            out_channels=channels_3,
F
Feng Ni 已提交
665
            norm_decay=norm_decay,
666 667 668 669 670 671 672 673
            freeze_norm=freeze_norm,
            name="tr2")

        self.st3 = Stage(
            num_channels=channels_3,
            num_modules=num_modules_3,
            num_filters=channels_3,
            has_se=self.has_se,
F
Feng Ni 已提交
674
            norm_decay=norm_decay,
675 676 677 678 679 680
            freeze_norm=freeze_norm,
            name="st3")

        self.tr3 = TransitionLayer(
            in_channels=channels_3,
            out_channels=channels_4,
F
Feng Ni 已提交
681
            norm_decay=norm_decay,
682 683 684 685 686 687 688
            freeze_norm=freeze_norm,
            name="tr3")
        self.st4 = Stage(
            num_channels=channels_4,
            num_modules=num_modules_4,
            num_filters=channels_4,
            has_se=self.has_se,
F
Feng Ni 已提交
689
            norm_decay=norm_decay,
690
            freeze_norm=freeze_norm,
691
            multi_scale_output=len(return_idx) > 1,
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
            name="st4")

    def forward(self, inputs):
        x = inputs['image']
        conv1 = self.conv_layer1_1(x)
        conv2 = self.conv_layer1_2(conv1)

        la1 = self.la1(conv2)
        tr1 = self.tr1([la1])
        st2 = self.st2(tr1)
        tr2 = self.tr2(st2)

        st3 = self.st3(tr2)
        tr3 = self.tr3(st3)

        st4 = self.st4(tr3)

        res = []
        for i, layer in enumerate(st4):
            if i == self.freeze_at:
                layer.stop_gradient = True
            if i in self.return_idx:
                res.append(layer)

        return res
717 718 719 720 721 722 723 724

    @property
    def out_shape(self):
        return [
            ShapeSpec(
                channels=self._out_channels[i], stride=self._out_strides[i])
            for i in self.return_idx
        ]