rec_lcnetv3.py

# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
#
# 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 import ParamAttr
from paddle.nn.initializer import Constant, KaimingNormal
from paddle.nn import AdaptiveAvgPool2D, BatchNorm2D, Conv2D, Dropout, Hardsigmoid, Hardswish, Identity, Linear, ReLU
from paddle.regularizer import L2Decay

NET_CONFIG_det = {
    "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]],
    "blocks6": [[5, 256, 512, 2, True], [5, 512, 512, 1, True],
                [5, 512, 512, 1, False], [5, 512, 512, 1, False]]
}

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


def make_divisible(v, divisor=16, 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 LearnableAffineBlock(nn.Layer):
    def __init__(self, scale_value=1.0, bias_value=0.0, lr_mult=1.0,
                 lab_lr=0.1):
        super().__init__()
        self.scale = self.create_parameter(
            shape=[1, ],
            default_initializer=Constant(value=scale_value),
            attr=ParamAttr(learning_rate=lr_mult * lab_lr))
        self.add_parameter("scale", self.scale)
        self.bias = self.create_parameter(
            shape=[1, ],
            default_initializer=Constant(value=bias_value),
            attr=ParamAttr(learning_rate=lr_mult * lab_lr))
        self.add_parameter("bias", self.bias)

    def forward(self, x):
        return self.scale * x + self.bias


class ConvBNLayer(nn.Layer):
    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 stride,
                 groups=1,
                 lr_mult=1.0):
        super().__init__()
        self.conv = Conv2D(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
            stride=stride,
            padding=(kernel_size - 1) // 2,
            groups=groups,
            weight_attr=ParamAttr(
                initializer=KaimingNormal(), learning_rate=lr_mult),
            bias_attr=False)

        self.bn = BatchNorm2D(
            out_channels,
            weight_attr=ParamAttr(
                regularizer=L2Decay(0.0), learning_rate=lr_mult),
            bias_attr=ParamAttr(
                regularizer=L2Decay(0.0), learning_rate=lr_mult))

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


class Act(nn.Layer):
    def __init__(self, act="hswish", lr_mult=1.0, lab_lr=0.1):
        super().__init__()
        if act == "hswish":
            self.act = Hardswish()
        else:
            assert act == "relu"
            self.act = ReLU()
        self.lab = LearnableAffineBlock(lr_mult=lr_mult, lab_lr=lab_lr)

    def forward(self, x):
        return self.lab(self.act(x))


class LearnableRepLayer(nn.Layer):
    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 stride=1,
                 groups=1,
                 num_conv_branches=1,
                 lr_mult=1.0,
                 lab_lr=0.1):
        super().__init__()
        self.is_repped = False
        self.groups = groups
        self.stride = stride
        self.kernel_size = kernel_size
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.num_conv_branches = num_conv_branches
        self.padding = (kernel_size - 1) // 2

        self.identity = BatchNorm2D(
            num_features=in_channels,
            weight_attr=ParamAttr(learning_rate=lr_mult),
            bias_attr=ParamAttr(learning_rate=lr_mult)
        ) if out_channels == in_channels and stride == 1 else None

        self.conv_kxk = nn.LayerList([
            ConvBNLayer(
                in_channels,
                out_channels,
                kernel_size,
                stride,
                groups=groups,
                lr_mult=lr_mult) for _ in range(self.num_conv_branches)
        ])

        self.conv_1x1 = ConvBNLayer(
            in_channels,
            out_channels,
            1,
            stride,
            groups=groups,
            lr_mult=lr_mult) if kernel_size > 1 else None

        self.lab = LearnableAffineBlock(lr_mult=lr_mult, lab_lr=lab_lr)
        self.act = Act(lr_mult=lr_mult, lab_lr=lab_lr)

    def forward(self, x):
        # for export
        if self.is_repped:
            out = self.lab(self.reparam_conv(x))
            if self.stride != 2:
                out = self.act(out)
            return out

        out = 0
        if self.identity is not None:
            out += self.identity(x)

        if self.conv_1x1 is not None:
            out += self.conv_1x1(x)

        for conv in self.conv_kxk:
            out += conv(x)

        out = self.lab(out)
        if self.stride != 2:
            out = self.act(out)
        return out

    def rep(self):
        if self.is_repped:
            return
        kernel, bias = self._get_kernel_bias()
        self.reparam_conv = Conv2D(
            in_channels=self.in_channels,
            out_channels=self.out_channels,
            kernel_size=self.kernel_size,
            stride=self.stride,
            padding=self.padding,
            groups=self.groups)
        self.reparam_conv.weight.set_value(kernel)
        self.reparam_conv.bias.set_value(bias)
        self.is_repped = True

    def _pad_kernel_1x1_to_kxk(self, kernel1x1, pad):
        if not isinstance(kernel1x1, paddle.Tensor):
            return 0
        else:
            return nn.functional.pad(kernel1x1, [pad, pad, pad, pad])

    def _get_kernel_bias(self):
        kernel_conv_1x1, bias_conv_1x1 = self._fuse_bn_tensor(self.conv_1x1)
        kernel_conv_1x1 = self._pad_kernel_1x1_to_kxk(kernel_conv_1x1,
                                                      self.kernel_size // 2)

        kernel_identity, bias_identity = self._fuse_bn_tensor(self.identity)

        kernel_conv_kxk = 0
        bias_conv_kxk = 0
        for conv in self.conv_kxk:
            kernel, bias = self._fuse_bn_tensor(conv)
            kernel_conv_kxk += kernel
            bias_conv_kxk += bias

        kernel_reparam = kernel_conv_kxk + kernel_conv_1x1 + kernel_identity
        bias_reparam = bias_conv_kxk + bias_conv_1x1 + bias_identity
        return kernel_reparam, bias_reparam

    def _fuse_bn_tensor(self, branch):
        if not branch:
            return 0, 0
        elif isinstance(branch, ConvBNLayer):
            kernel = branch.conv.weight
            running_mean = branch.bn._mean
            running_var = branch.bn._variance
            gamma = branch.bn.weight
            beta = branch.bn.bias
            eps = branch.bn._epsilon
        else:
            assert isinstance(branch, BatchNorm2D)
            if not hasattr(self, 'id_tensor'):
                input_dim = self.in_channels // self.groups
                kernel_value = paddle.zeros(
                    (self.in_channels, input_dim, self.kernel_size,
                     self.kernel_size),
                    dtype=branch.weight.dtype)
                for i in range(self.in_channels):
                    kernel_value[i, i % input_dim, self.kernel_size // 2,
                                 self.kernel_size // 2] = 1
                self.id_tensor = kernel_value
            kernel = self.id_tensor
            running_mean = branch._mean
            running_var = branch._variance
            gamma = branch.weight
            beta = branch.bias
            eps = branch._epsilon
        std = (running_var + eps).sqrt()
        t = (gamma / std).reshape((-1, 1, 1, 1))
        return kernel * t, beta - running_mean * gamma / std


class SELayer(nn.Layer):
    def __init__(self, channel, reduction=4, lr_mult=1.0):
        super().__init__()
        self.avg_pool = AdaptiveAvgPool2D(1)
        self.conv1 = Conv2D(
            in_channels=channel,
            out_channels=channel // reduction,
            kernel_size=1,
            stride=1,
            padding=0,
            weight_attr=ParamAttr(learning_rate=lr_mult),
            bias_attr=ParamAttr(learning_rate=lr_mult))
        self.relu = ReLU()
        self.conv2 = Conv2D(
            in_channels=channel // reduction,
            out_channels=channel,
            kernel_size=1,
            stride=1,
            padding=0,
            weight_attr=ParamAttr(learning_rate=lr_mult),
            bias_attr=ParamAttr(learning_rate=lr_mult))
        self.hardsigmoid = 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 LCNetV3Block(nn.Layer):
    def __init__(self,
                 in_channels,
                 out_channels,
                 stride,
                 dw_size,
                 use_se=False,
                 conv_kxk_num=4,
                 lr_mult=1.0,
                 lab_lr=0.1):
        super().__init__()
        self.use_se = use_se
        self.dw_conv = LearnableRepLayer(
            in_channels=in_channels,
            out_channels=in_channels,
            kernel_size=dw_size,
            stride=stride,
            groups=in_channels,
            num_conv_branches=conv_kxk_num,
            lr_mult=lr_mult,
            lab_lr=lab_lr)
        if use_se:
            self.se = SELayer(in_channels, lr_mult=lr_mult)
        self.pw_conv = LearnableRepLayer(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=1,
            stride=1,
            num_conv_branches=conv_kxk_num,
            lr_mult=lr_mult,
            lab_lr=lab_lr)

    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 PPLCNetV3(nn.Layer):
    def __init__(self,
                 scale=1.0,
                 conv_kxk_num=4,
                 lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
                 lab_lr=0.1,
                 det=False,
                 **kwargs):
        super().__init__()
        self.scale = scale
        self.lr_mult_list = lr_mult_list
        self.det = det

        self.net_config = NET_CONFIG_det if self.det else NET_CONFIG_rec

        assert isinstance(self.lr_mult_list, (
            list, tuple
        )), "lr_mult_list should be in (list, tuple) but got {}".format(
            type(self.lr_mult_list))
        assert len(self.lr_mult_list
                   ) == 6, "lr_mult_list length should be 6 but got {}".format(
                       len(self.lr_mult_list))

        self.conv1 = ConvBNLayer(
            in_channels=3,
            out_channels=make_divisible(16 * scale),
            kernel_size=3,
            stride=2,
            lr_mult=self.lr_mult_list[0])

        self.blocks2 = nn.Sequential(*[
            LCNetV3Block(
                in_channels=make_divisible(in_c * scale),
                out_channels=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se,
                conv_kxk_num=conv_kxk_num,
                lr_mult=self.lr_mult_list[1],
                lab_lr=lab_lr)
            for i, (k, in_c, out_c, s, se
                    ) in enumerate(self.net_config["blocks2"])
        ])

        self.blocks3 = nn.Sequential(*[
            LCNetV3Block(
                in_channels=make_divisible(in_c * scale),
                out_channels=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se,
                conv_kxk_num=conv_kxk_num,
                lr_mult=self.lr_mult_list[2],
                lab_lr=lab_lr)
            for i, (k, in_c, out_c, s, se
                    ) in enumerate(self.net_config["blocks3"])
        ])

        self.blocks4 = nn.Sequential(*[
            LCNetV3Block(
                in_channels=make_divisible(in_c * scale),
                out_channels=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se,
                conv_kxk_num=conv_kxk_num,
                lr_mult=self.lr_mult_list[3],
                lab_lr=lab_lr)
            for i, (k, in_c, out_c, s, se
                    ) in enumerate(self.net_config["blocks4"])
        ])

        self.blocks5 = nn.Sequential(*[
            LCNetV3Block(
                in_channels=make_divisible(in_c * scale),
                out_channels=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se,
                conv_kxk_num=conv_kxk_num,
                lr_mult=self.lr_mult_list[4],
                lab_lr=lab_lr)
            for i, (k, in_c, out_c, s, se
                    ) in enumerate(self.net_config["blocks5"])
        ])

        self.blocks6 = nn.Sequential(*[
            LCNetV3Block(
                in_channels=make_divisible(in_c * scale),
                out_channels=make_divisible(out_c * scale),
                dw_size=k,
                stride=s,
                use_se=se,
                conv_kxk_num=conv_kxk_num,
                lr_mult=self.lr_mult_list[5],
                lab_lr=lab_lr)
            for i, (k, in_c, out_c, s, se
                    ) in enumerate(self.net_config["blocks6"])
        ])
        self.out_channels = make_divisible(512 * scale)

        if self.det:
            mv_c = [16, 24, 56, 480]
            self.out_channels = [
                make_divisible(self.net_config["blocks3"][-1][2] * scale),
                make_divisible(self.net_config["blocks4"][-1][2] * scale),
                make_divisible(self.net_config["blocks5"][-1][2] * scale),
                make_divisible(self.net_config["blocks6"][-1][2] * scale),
            ]

            self.layer_list = nn.LayerList([
                nn.Conv2D(self.out_channels[0], int(mv_c[0] * scale), 1, 1, 0),
                nn.Conv2D(self.out_channels[1], int(mv_c[1] * scale), 1, 1, 0),
                nn.Conv2D(self.out_channels[2], int(mv_c[2] * scale), 1, 1, 0),
                nn.Conv2D(self.out_channels[3], int(mv_c[3] * scale), 1, 1, 0)
            ])
            self.out_channels = [
                int(mv_c[0] * scale), int(mv_c[1] * scale),
                int(mv_c[2] * scale), int(mv_c[3] * scale)
            ]

    def forward(self, x):
        out_list = []
        x = self.conv1(x)

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

        if self.det:
            out_list[0] = self.layer_list[0](out_list[0])
            out_list[1] = self.layer_list[1](out_list[1])
            out_list[2] = self.layer_list[2](out_list[2])
            out_list[3] = self.layer_list[3](out_list[3])
            return out_list

        if self.training:
            x = F.adaptive_avg_pool2d(x, [1, 40])
        else:
            x = F.avg_pool2d(x, [3, 2])
        return x