hrnet.py

# 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 math

import paddle
import paddle.nn as nn
import paddle.nn.functional as F

from paddle3d.apis import manager
from paddle3d.models import layers
from paddle3d.models.layers import param_init, reset_parameters
from paddle3d.utils import checkpoint

__all__ = ["HRNet_W18"]


@manager.BACKBONES.add_component
class HRNet_W18(nn.Layer):
    """
    The HRNet implementation based on PaddlePaddle.
    The original article refers to
    Jingdong Wang, et, al. "HRNet：Deep High-Resolution Representation Learning for Visual Recognition"
    (https://arxiv.org/pdf/1908.07919.pdf).
    Args:
        pretrained (str, optional): The path of pretrained model.
        stage1_num_modules (int, optional): Number of modules for stage1. Default 1.
        stage1_num_blocks (list, optional): Number of blocks per module for stage1. Default (4).
        stage1_num_channels (list, optional): Number of channels per branch for stage1. Default (64).
        stage2_num_modules (int, optional): Number of modules for stage2. Default 1.
        stage2_num_blocks (list, optional): Number of blocks per module for stage2. Default (4, 4).
        stage2_num_channels (list, optional): Number of channels per branch for stage2. Default (18, 36).
        stage3_num_modules (int, optional): Number of modules for stage3. Default 4.
        stage3_num_blocks (list, optional): Number of blocks per module for stage3. Default (4, 4, 4).
        stage3_num_channels (list, optional): Number of channels per branch for stage3. Default [18, 36, 72).
        stage4_num_modules (int, optional): Number of modules for stage4. Default 3.
        stage4_num_blocks (list, optional): Number of blocks per module for stage4. Default (4, 4, 4, 4).
        stage4_num_channels (list, optional): Number of channels per branch for stage4. Default (18, 36, 72. 144).
        has_se (bool, optional): Whether to use Squeeze-and-Excitation module. Default False.
        align_corners (bool, optional): An argument of F.interpolate. It should be set to False when the feature size is even,
            e.g. 1024x512, otherwise it is True, e.g. 769x769. Default: False.
    """

    def __init__(self,
                 pretrained=None,
                 stage1_num_modules=1,
                 stage1_num_blocks=(4, ),
                 stage1_num_channels=(64, ),
                 stage2_num_modules=1,
                 stage2_num_blocks=(4, 4),
                 stage2_num_channels=(18, 36),
                 stage3_num_modules=4,
                 stage3_num_blocks=(4, 4, 4),
                 stage3_num_channels=(18, 36, 72),
                 stage4_num_modules=3,
                 stage4_num_blocks=(4, 4, 4, 4),
                 stage4_num_channels=(18, 36, 72, 144),
                 has_se=False,
                 align_corners=False,
                 padding_same=True):
        super(HRNet_W18, self).__init__()
        self.pretrained = pretrained
        self.stage1_num_modules = stage1_num_modules
        self.stage1_num_blocks = stage1_num_blocks
        self.stage1_num_channels = stage1_num_channels
        self.stage2_num_modules = stage2_num_modules
        self.stage2_num_blocks = stage2_num_blocks
        self.stage2_num_channels = stage2_num_channels
        self.stage3_num_modules = stage3_num_modules
        self.stage3_num_blocks = stage3_num_blocks
        self.stage3_num_channels = stage3_num_channels
        self.stage4_num_modules = stage4_num_modules
        self.stage4_num_blocks = stage4_num_blocks
        self.stage4_num_channels = stage4_num_channels
        self.has_se = has_se
        self.align_corners = align_corners
        self.feat_channels = [sum(stage4_num_channels)]
        self.norm_mean = paddle.to_tensor([0.485, 0.456, 0.406])
        self.norm_std = paddle.to_tensor([0.229, 0.224, 0.225])

        self.conv_layer1_1 = layers.ConvBNReLU(
            in_channels=3,
            out_channels=64,
            kernel_size=3,
            stride=2,
            padding=1 if not padding_same else 'same',
            bias_attr=False)

        self.conv_layer1_2 = layers.ConvBNReLU(
            in_channels=64,
            out_channels=64,
            kernel_size=3,
            stride=2,
            padding=1 if not padding_same else 'same',
            bias_attr=False)

        self.la1 = Layer1(
            num_channels=64,
            num_blocks=self.stage1_num_blocks[0],
            num_filters=self.stage1_num_channels[0],
            has_se=has_se,
            name="layer2",
            padding_same=padding_same)

        self.tr1 = TransitionLayer(
            in_channels=[self.stage1_num_channels[0] * 4],
            out_channels=self.stage2_num_channels,
            name="tr1",
            padding_same=padding_same)

        self.st2 = Stage(
            num_channels=self.stage2_num_channels,
            num_modules=self.stage2_num_modules,
            num_blocks=self.stage2_num_blocks,
            num_filters=self.stage2_num_channels,
            has_se=self.has_se,
            name="st2",
            align_corners=align_corners,
            padding_same=padding_same)

        self.tr2 = TransitionLayer(
            in_channels=self.stage2_num_channels,
            out_channels=self.stage3_num_channels,
            name="tr2",
            padding_same=padding_same)
        self.st3 = Stage(
            num_channels=self.stage3_num_channels,
            num_modules=self.stage3_num_modules,
            num_blocks=self.stage3_num_blocks,
            num_filters=self.stage3_num_channels,
            has_se=self.has_se,
            name="st3",
            align_corners=align_corners,
            padding_same=padding_same)

        self.tr3 = TransitionLayer(
            in_channels=self.stage3_num_channels,
            out_channels=self.stage4_num_channels,
            name="tr3",
            padding_same=padding_same)
        self.st4 = Stage(
            num_channels=self.stage4_num_channels,
            num_modules=self.stage4_num_modules,
            num_blocks=self.stage4_num_blocks,
            num_filters=self.stage4_num_channels,
            has_se=self.has_se,
            name="st4",
            align_corners=align_corners,
            padding_same=padding_same)

        self.init_weight()

    def forward(self, x):
        x = self.preprocess(x)
        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)

        size = paddle.shape(st4[1])[2:]
        x1 = F.interpolate(
            st4[0], size, mode='bilinear', align_corners=self.align_corners)
        x2 = F.interpolate(
            st4[2], size, mode='bilinear', align_corners=self.align_corners)
        x3 = F.interpolate(
            st4[3], size, mode='bilinear', align_corners=self.align_corners)
        x = paddle.concat([x1, st4[1], x2, x3], axis=1)

        return [la1, x]

    def init_weight(self):
        for sublayer in self.sublayers():
            if isinstance(sublayer, nn.Conv2D):
                reset_parameters(sublayer)
        if self.pretrained is not None:
            checkpoint.load_pretrained_model(self, self.pretrained)

    def preprocess(self, images):
        """
        Preprocess images
        Args:
            images [paddle.Tensor(N, 3, H, W)]: Input images
        Return
            x [paddle.Tensor(N, 3, H, W)]: Preprocessed images
        """
        x = images
        # Create a mask for padded pixels
        mask = paddle.isnan(x)

        # Match ResNet pretrained preprocessing
        x = self.normalize(x, mean=self.norm_mean, std=self.norm_std)

        # Make padded pixels = 0
        a = paddle.zeros_like(x)
        x = paddle.where(mask, a, x)

        return x

    def normalize(self, image, mean, std):
        shape = paddle.shape(image)
        if mean.shape:
            mean = mean[..., :, None]
        if std.shape:
            std = std[..., :, None]
        out = (image.reshape([shape[0], shape[1], shape[2] * shape[3]]) -
               mean) / std
        return out.reshape(shape)


class Layer1(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_filters,
                 num_blocks,
                 has_se=False,
                 name=None,
                 padding_same=True):
        super(Layer1, self).__init__()

        self.bottleneck_block_list = []

        for i in range(num_blocks):
            bottleneck_block = self.add_sublayer(
                "bb_{}_{}".format(name, i + 1),
                BottleneckBlock(
                    num_channels=num_channels if i == 0 else num_filters * 4,
                    num_filters=num_filters,
                    has_se=has_se,
                    stride=1,
                    downsample=True if i == 0 else False,
                    name=name + '_' + str(i + 1),
                    padding_same=padding_same))
            self.bottleneck_block_list.append(bottleneck_block)

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


class TransitionLayer(nn.Layer):
    def __init__(self, in_channels, out_channels, name=None, padding_same=True):
        super(TransitionLayer, self).__init__()

        num_in = len(in_channels)
        num_out = len(out_channels)
        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),
                        layers.ConvBNReLU(
                            in_channels=in_channels[i],
                            out_channels=out_channels[i],
                            kernel_size=3,
                            padding=1 if not padding_same else 'same',
                            bias_attr=False))
            else:
                residual = self.add_sublayer(
                    "transition_{}_layer_{}".format(name, i + 1),
                    layers.ConvBNReLU(
                        in_channels=in_channels[-1],
                        out_channels=out_channels[i],
                        kernel_size=3,
                        stride=2,
                        padding=1 if not padding_same else 'same',
                        bias_attr=False))
            self.conv_bn_func_list.append(residual)

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


class Branches(nn.Layer):
    def __init__(self,
                 num_blocks,
                 in_channels,
                 out_channels,
                 has_se=False,
                 name=None,
                 padding_same=True):
        super(Branches, self).__init__()

        self.basic_block_list = []

        for i in range(len(out_channels)):
            self.basic_block_list.append([])
            for j in range(num_blocks[i]):
                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,
                        name=name + '_branch_layer_' + str(i + 1) + '_' +
                        str(j + 1),
                        padding_same=padding_same))
                self.basic_block_list[i].append(basic_block_func)

    def forward(self, x):
        outs = []
        for idx, input in enumerate(x):
            conv = input
            for basic_block_func in self.basic_block_list[idx]:
                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,
                 name=None,
                 padding_same=True):
        super(BottleneckBlock, self).__init__()

        self.has_se = has_se
        self.downsample = downsample

        self.conv1 = layers.ConvBNReLU(
            in_channels=num_channels,
            out_channels=num_filters,
            kernel_size=1,
            bias_attr=False)

        self.conv2 = layers.ConvBNReLU(
            in_channels=num_filters,
            out_channels=num_filters,
            kernel_size=3,
            stride=stride,
            padding=1 if not padding_same else 'same',
            bias_attr=False)

        self.conv3 = layers.ConvBN(
            in_channels=num_filters,
            out_channels=num_filters * 4,
            kernel_size=1,
            bias_attr=False)

        if self.downsample:
            self.conv_down = layers.ConvBN(
                in_channels=num_channels,
                out_channels=num_filters * 4,
                kernel_size=1,
                bias_attr=False)

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

        self.relu = nn.ReLU()

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

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

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

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


class BasicBlock(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_filters,
                 stride=1,
                 has_se=False,
                 downsample=False,
                 name=None,
                 padding_same=True):
        super(BasicBlock, self).__init__()

        self.has_se = has_se
        self.downsample = downsample

        self.conv1 = layers.ConvBNReLU(
            in_channels=num_channels,
            out_channels=num_filters,
            kernel_size=3,
            stride=stride,
            padding=1 if not padding_same else 'same',
            bias_attr=False)
        self.conv2 = layers.ConvBN(
            in_channels=num_filters,
            out_channels=num_filters,
            kernel_size=3,
            padding=1 if not padding_same else 'same',
            bias_attr=False)

        if self.downsample:
            self.conv_down = layers.ConvBNReLU(
                in_channels=num_channels,
                out_channels=num_filters,
                kernel_size=1,
                bias_attr=False)

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

        self.relu = nn.ReLU()

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

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

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

        y = paddle.add(conv2, residual)
        y = self.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 = nn.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 = nn.Linear(
            num_channels,
            med_ch,
            weight_attr=paddle.ParamAttr(
                initializer=nn.initializer.Uniform(-stdv, stdv)))

        stdv = 1.0 / math.sqrt(med_ch * 1.0)
        self.excitation = nn.Linear(
            med_ch,
            num_filters,
            weight_attr=paddle.ParamAttr(
                initializer=nn.initializer.Uniform(-stdv, stdv)))

    def forward(self, x):
        pool = self.pool2d_gap(x)
        pool = paddle.reshape(pool, shape=[-1, self._num_channels])
        squeeze = self.squeeze(pool)
        squeeze = F.relu(squeeze)
        excitation = self.excitation(squeeze)
        excitation = F.sigmoid(excitation)
        excitation = paddle.reshape(
            excitation, shape=[-1, self._num_channels, 1, 1])
        out = x * excitation
        return out


class Stage(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_modules,
                 num_blocks,
                 num_filters,
                 has_se=False,
                 multi_scale_output=True,
                 name=None,
                 align_corners=False,
                 padding_same=True):
        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_blocks=num_blocks,
                        num_filters=num_filters,
                        has_se=has_se,
                        multi_scale_output=False,
                        name=name + '_' + str(i + 1),
                        align_corners=align_corners,
                        padding_same=padding_same))
            else:
                stage_func = self.add_sublayer(
                    "stage_{}_{}".format(name, i + 1),
                    HighResolutionModule(
                        num_channels=num_channels,
                        num_blocks=num_blocks,
                        num_filters=num_filters,
                        has_se=has_se,
                        name=name + '_' + str(i + 1),
                        align_corners=align_corners,
                        padding_same=padding_same))

            self.stage_func_list.append(stage_func)

    def forward(self, x):
        out = x
        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_blocks,
                 num_filters,
                 has_se=False,
                 multi_scale_output=True,
                 name=None,
                 align_corners=False,
                 padding_same=True):
        super(HighResolutionModule, self).__init__()

        self.branches_func = Branches(
            num_blocks=num_blocks,
            in_channels=num_channels,
            out_channels=num_filters,
            has_se=has_se,
            name=name,
            padding_same=padding_same)

        self.fuse_func = FuseLayers(
            in_channels=num_filters,
            out_channels=num_filters,
            multi_scale_output=multi_scale_output,
            name=name,
            align_corners=align_corners,
            padding_same=padding_same)

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


class FuseLayers(nn.Layer):
    def __init__(self,
                 in_channels,
                 out_channels,
                 multi_scale_output=True,
                 name=None,
                 align_corners=False,
                 padding_same=True):
        super(FuseLayers, self).__init__()

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

        self.residual_func_list = []
        for i in range(self._actual_ch):
            for j in range(len(in_channels)):
                if j > i:
                    residual_func = self.add_sublayer(
                        "residual_{}_layer_{}_{}".format(name, i + 1, j + 1),
                        layers.ConvBN(
                            in_channels=in_channels[j],
                            out_channels=out_channels[i],
                            kernel_size=1,
                            bias_attr=False))
                    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),
                                layers.ConvBN(
                                    in_channels=pre_num_filters,
                                    out_channels=out_channels[i],
                                    kernel_size=3,
                                    stride=2,
                                    padding=1 if not padding_same else 'same',
                                    bias_attr=False))
                            pre_num_filters = out_channels[i]
                        else:
                            residual_func = self.add_sublayer(
                                "residual_{}_layer_{}_{}_{}".format(
                                    name, i + 1, j + 1, k + 1),
                                layers.ConvBNReLU(
                                    in_channels=pre_num_filters,
                                    out_channels=out_channels[j],
                                    kernel_size=3,
                                    stride=2,
                                    padding=1 if not padding_same else 'same',
                                    bias_attr=False))
                            pre_num_filters = out_channels[j]
                        self.residual_func_list.append(residual_func)

    def forward(self, x):
        outs = []
        residual_func_idx = 0
        for i in range(self._actual_ch):
            residual = x[i]
            residual_shape = paddle.shape(residual)[-2:]
            for j in range(len(self._in_channels)):
                if j > i:
                    y = self.residual_func_list[residual_func_idx](x[j])
                    residual_func_idx += 1

                    y = F.interpolate(
                        y,
                        residual_shape,
                        mode='bilinear',
                        align_corners=self.align_corners)
                    residual = residual + y
                elif j < i:
                    y = x[j]
                    for k in range(i - j):
                        y = self.residual_func_list[residual_func_idx](y)
                        residual_func_idx += 1

                    residual = residual + y

            residual = F.relu(residual)
            outs.append(residual)

        return outs