PaddleRS/paddlers/models/ppdet/modeling/reid/pplcnet_embedding.py

# Copyright (c) 2022 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 __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.nn.initializer import Normal, Constant
from paddle import ParamAttr
from paddle.nn import AdaptiveAvgPool2D, BatchNorm2D, Conv2D, Linear
from paddle.regularizer import L2Decay
from paddle.nn.initializer import KaimingNormal, XavierNormal
from paddlers.models.ppdet.core.workspace import register

__all__ = ['PPLCNetEmbedding']

# Each element(list) represents a depthwise block, which is composed of k, in_c, out_c, s, use_se.
# k: kernel_size
# in_c: input channel number in depthwise block
# out_c: output channel number in depthwise block
# s: stride in depthwise block
# use_se: whether to use SE block

NET_CONFIG = {
    "blocks2":
    #k, in_c, out_c, s, use_se
    [[3, 16, 32, 1, False]],
    "blocks3": [[3, 32, 64, 2, False], [3, 64, 64, 1, False]],
    "blocks4": [[3, 64, 128, 2, False], [3, 128, 128, 1, False]],
    "blocks5":
    [[3, 128, 256, 2, False], [5, 256, 256, 1, False], [5, 256, 256, 1, False],
     [5, 256, 256, 1, False], [5, 256, 256, 1, False], [5, 256, 256, 1, False]],
    "blocks6": [[5, 256, 512, 2, True], [5, 512, 512, 1, True]]
}


def make_divisible(v, divisor=8, min_value=None):
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v


class ConvBNLayer(nn.Layer):
    def __init__(self,
                 num_channels,
                 filter_size,
                 num_filters,
                 stride,
                 num_groups=1):
        super().__init__()

        self.conv = Conv2D(
            in_channels=num_channels,
            out_channels=num_filters,
            kernel_size=filter_size,
            stride=stride,
            padding=(filter_size - 1) // 2,
            groups=num_groups,
            weight_attr=ParamAttr(initializer=KaimingNormal()),
            bias_attr=False)

        self.bn = BatchNorm2D(
            num_filters,
            weight_attr=ParamAttr(regularizer=L2Decay(0.0)),
            bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
        self.hardswish = nn.Hardswish()

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.hardswish(x)
        return x


class DepthwiseSeparable(nn.Layer):
    def __init__(self,
                 num_channels,
                 num_filters,
                 stride,
                 dw_size=3,
                 use_se=False):
        super().__init__()
        self.use_se = use_se
        self.dw_conv = ConvBNLayer(
            num_channels=num_channels,
            num_filters=num_channels,
            filter_size=dw_size,
            stride=stride,
            num_groups=num_channels)
        if use_se:
            self.se = SEModule(num_channels)
        self.pw_conv = ConvBNLayer(
            num_channels=num_channels,
            filter_size=1,
            num_filters=num_filters,
            stride=1)

    def forward(self, x):
        x = self.dw_conv(x)
        if self.use_se:
            x = self.se(x)
        x = self.pw_conv(x)
        return x


class SEModule(nn.Layer):
    def __init__(self, channel, reduction=4):
        super().__init__()
        self.avg_pool = AdaptiveAvgPool2D(1)
        self.conv1 = Conv2D(
            in_channels=channel,
            out_channels=channel // reduction,
            kernel_size=1,
            stride=1,
            padding=0)
        self.relu = nn.ReLU()
        self.conv2 = Conv2D(
            in_channels=channel // reduction,
            out_channels=channel,
            kernel_size=1,
            stride=1,
            padding=0)
        self.hardsigmoid = nn.Hardsigmoid()

    def forward(self, x):
        identity = x
        x = self.avg_pool(x)
        x = self.conv1(x)
        x = self.relu(x)
        x = self.conv2(x)
        x = self.hardsigmoid(x)
        x = paddle.multiply(x=identity, y=x)
        return x


class PPLCNet(nn.Layer):
    """
    PP-LCNet, see https://arxiv.org/abs/2109.15099.
    This code is different from PPLCNet in ppdet/modeling/backbones/lcnet.py
    or in PaddleClas, because the output is the flatten feature of last_conv.

    Args:
        scale (float): Scale ratio of channels.
        class_expand (int): Number of channels of conv feature.
    """

    def __init__(self, scale=1.0, class_expand=1280):
        super(PPLCNet, self).__init__()
        self.scale = scale
        self.class_expand = class_expand

        self.conv1 = ConvBNLayer(
            num_channels=3,
            filter_size=3,
            num_filters=make_divisible(16 * scale),
            stride=2)

        self.blocks2 = nn.Sequential(*[
            DepthwiseSeparable(
                num_channels=make_divisible(in_c * scale),
                num_filters=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se)
            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks2"])
        ])

        self.blocks3 = nn.Sequential(*[
            DepthwiseSeparable(
                num_channels=make_divisible(in_c * scale),
                num_filters=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se)
            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks3"])
        ])

        self.blocks4 = nn.Sequential(*[
            DepthwiseSeparable(
                num_channels=make_divisible(in_c * scale),
                num_filters=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se)
            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks4"])
        ])

        self.blocks5 = nn.Sequential(*[
            DepthwiseSeparable(
                num_channels=make_divisible(in_c * scale),
                num_filters=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se)
            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks5"])
        ])

        self.blocks6 = nn.Sequential(*[
            DepthwiseSeparable(
                num_channels=make_divisible(in_c * scale),
                num_filters=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se)
            for i, (k, in_c, out_c, s, se) in enumerate(NET_CONFIG["blocks6"])
        ])

        self.avg_pool = AdaptiveAvgPool2D(1)
        self.last_conv = Conv2D(
            in_channels=make_divisible(NET_CONFIG["blocks6"][-1][2] * scale),
            out_channels=self.class_expand,
            kernel_size=1,
            stride=1,
            padding=0,
            bias_attr=False)
        self.hardswish = nn.Hardswish()
        self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)

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

        x = self.blocks2(x)
        x = self.blocks3(x)
        x = self.blocks4(x)
        x = self.blocks5(x)
        x = self.blocks6(x)

        x = self.avg_pool(x)
        x = self.last_conv(x)
        x = self.hardswish(x)
        x = self.flatten(x)
        return x


class FC(nn.Layer):
    def __init__(self, input_ch, output_ch):
        super(FC, self).__init__()
        weight_attr = ParamAttr(initializer=XavierNormal())
        self.fc = paddle.nn.Linear(input_ch, output_ch, weight_attr=weight_attr)

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


@register
class PPLCNetEmbedding(nn.Layer):
    """
    PPLCNet Embedding

    Args:
        input_ch (int): Number of channels of input conv feature.
        output_ch (int): Number of channels of output conv feature.
    """

    def __init__(self, scale=2.5, input_ch=1280, output_ch=512):
        super(PPLCNetEmbedding, self).__init__()
        self.backbone = PPLCNet(scale=scale)
        self.neck = FC(input_ch, output_ch)

    def forward(self, x):
        feat = self.backbone(x)
        feat_out = self.neck(feat)
        return feat_out