rec_repvit.py

# copyright (c) 2024 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.

"""
This code is refer from:
https://github.com/THU-MIG/RepViT
"""

import paddle.nn as nn
import paddle
from paddle.nn.initializer import TruncatedNormal, Constant, Normal

trunc_normal_ = TruncatedNormal(std=0.02)
normal_ = Normal
zeros_ = Constant(value=0.0)
ones_ = Constant(value=1.0)


def _make_divisible(v, divisor, min_value=None):
    """
    This function is taken from the original tf repo.
    It ensures that all layers have a channel number that is divisible by 8
    It can be seen here:
    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
    :param v:
    :param divisor:
    :param min_value:
    :return:
    """
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v


# from timm.models.layers import SqueezeExcite


def make_divisible(v, divisor=8, min_value=None, round_limit=0.9):
    min_value = min_value or divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_v < round_limit * v:
        new_v += divisor
    return new_v


class SEModule(nn.Layer):
    """SE Module as defined in original SE-Nets with a few additions
    Additions include:
        * divisor can be specified to keep channels % div == 0 (default: 8)
        * reduction channels can be specified directly by arg (if rd_channels is set)
        * reduction channels can be specified by float rd_ratio (default: 1/16)
        * global max pooling can be added to the squeeze aggregation
        * customizable activation, normalization, and gate layer
    """

    def __init__(
        self,
        channels,
        rd_ratio=1.0 / 16,
        rd_channels=None,
        rd_divisor=8,
        act_layer=nn.ReLU,
    ):
        super(SEModule, self).__init__()
        if not rd_channels:
            rd_channels = make_divisible(
                channels * rd_ratio, rd_divisor, round_limit=0.0
            )
        self.fc1 = nn.Conv2D(channels, rd_channels, kernel_size=1, bias_attr=True)
        self.act = act_layer()
        self.fc2 = nn.Conv2D(rd_channels, channels, kernel_size=1, bias_attr=True)

    def forward(self, x):
        x_se = x.mean((2, 3), keepdim=True)
        x_se = self.fc1(x_se)
        x_se = self.act(x_se)
        x_se = self.fc2(x_se)
        return x * nn.functional.sigmoid(x_se)


class Conv2D_BN(nn.Sequential):
    def __init__(
        self,
        a,
        b,
        ks=1,
        stride=1,
        pad=0,
        dilation=1,
        groups=1,
        bn_weight_init=1,
        resolution=-10000,
    ):
        super().__init__()
        self.add_sublayer(
            "c", nn.Conv2D(a, b, ks, stride, pad, dilation, groups, bias_attr=False)
        )
        self.add_sublayer("bn", nn.BatchNorm2D(b))
        if bn_weight_init == 1:
            ones_(self.bn.weight)
        else:
            zeros_(self.bn.weight)
        zeros_(self.bn.bias)

    @paddle.no_grad()
    def fuse(self):
        c, bn = self.c, self.bn
        w = bn.weight / (bn._variance + bn._epsilon) ** 0.5
        w = c.weight * w[:, None, None, None]
        b = bn.bias - bn._mean * bn.weight / (bn._variance + bn._epsilon) ** 0.5
        m = nn.Conv2D(
            w.shape[1] * self.c._groups,
            w.shape[0],
            w.shape[2:],
            stride=self.c._stride,
            padding=self.c._padding,
            dilation=self.c._dilation,
            groups=self.c._groups,
        )
        m.weight.set_value(w)
        m.bias.set_value(b)
        return m


class Residual(nn.Layer):
    def __init__(self, m, drop=0.0):
        super().__init__()
        self.m = m
        self.drop = drop

    def forward(self, x):
        if self.training and self.drop > 0:
            return (
                x
                + self.m(x)
                * paddle.rand(x.size(0), 1, 1, 1)
                .ge_(self.drop)
                .div(1 - self.drop)
                .detach()
            )
        else:
            return x + self.m(x)

    @paddle.no_grad()
    def fuse(self):
        if isinstance(self.m, Conv2D_BN):
            m = self.m.fuse()
            assert m._groups == m.in_channels
            identity = paddle.ones([m.weight.shape[0], m.weight.shape[1], 1, 1])
            identity = nn.functional.pad(identity, [1, 1, 1, 1])
            m.weight += identity
            return m
        elif isinstance(self.m, nn.Conv2D):
            m = self.m
            assert m._groups != m.in_channels
            identity = paddle.ones([m.weight.shape[0], m.weight.shape[1], 1, 1])
            identity = nn.functional.pad(identity, [1, 1, 1, 1])
            m.weight += identity
            return m
        else:
            return self


class RepVGGDW(nn.Layer):
    def __init__(self, ed) -> None:
        super().__init__()
        self.conv = Conv2D_BN(ed, ed, 3, 1, 1, groups=ed)
        self.conv1 = nn.Conv2D(ed, ed, 1, 1, 0, groups=ed)
        self.dim = ed
        self.bn = nn.BatchNorm2D(ed)

    def forward(self, x):
        return self.bn((self.conv(x) + self.conv1(x)) + x)

    @paddle.no_grad()
    def fuse(self):
        conv = self.conv.fuse()
        conv1 = self.conv1

        conv_w = conv.weight
        conv_b = conv.bias
        conv1_w = conv1.weight
        conv1_b = conv1.bias

        conv1_w = nn.functional.pad(conv1_w, [1, 1, 1, 1])

        identity = nn.functional.pad(
            paddle.ones([conv1_w.shape[0], conv1_w.shape[1], 1, 1]), [1, 1, 1, 1]
        )

        final_conv_w = conv_w + conv1_w + identity
        final_conv_b = conv_b + conv1_b

        conv.weight.set_value(final_conv_w)
        conv.bias.set_value(final_conv_b)

        bn = self.bn
        w = bn.weight / (bn._variance + bn._epsilon) ** 0.5
        w = conv.weight * w[:, None, None, None]
        b = (
            bn.bias
            + (conv.bias - bn._mean) * bn.weight / (bn._variance + bn._epsilon) ** 0.5
        )
        conv.weight.set_value(w)
        conv.bias.set_value(b)
        return conv


class RepViTBlock(nn.Layer):
    def __init__(self, inp, hidden_dim, oup, kernel_size, stride, use_se, use_hs):
        super(RepViTBlock, self).__init__()

        self.identity = stride == 1 and inp == oup
        assert hidden_dim == 2 * inp

        if stride != 1:
            self.token_mixer = nn.Sequential(
                Conv2D_BN(
                    inp, inp, kernel_size, stride, (kernel_size - 1) // 2, groups=inp
                ),
                SEModule(inp, 0.25) if use_se else nn.Identity(),
                Conv2D_BN(inp, oup, ks=1, stride=1, pad=0),
            )
            self.channel_mixer = Residual(
                nn.Sequential(
                    # pw
                    Conv2D_BN(oup, 2 * oup, 1, 1, 0),
                    nn.GELU() if use_hs else nn.GELU(),
                    # pw-linear
                    Conv2D_BN(2 * oup, oup, 1, 1, 0, bn_weight_init=0),
                )
            )
        else:
            assert self.identity
            self.token_mixer = nn.Sequential(
                RepVGGDW(inp),
                SEModule(inp, 0.25) if use_se else nn.Identity(),
            )
            self.channel_mixer = Residual(
                nn.Sequential(
                    # pw
                    Conv2D_BN(inp, hidden_dim, 1, 1, 0),
                    nn.GELU() if use_hs else nn.GELU(),
                    # pw-linear
                    Conv2D_BN(hidden_dim, oup, 1, 1, 0, bn_weight_init=0),
                )
            )

    def forward(self, x):
        return self.channel_mixer(self.token_mixer(x))


class RepViT(nn.Layer):
    def __init__(self, cfgs, in_channels=3, out_indices=None):
        super(RepViT, self).__init__()
        # setting of inverted residual blocks
        self.cfgs = cfgs

        # building first layer
        input_channel = self.cfgs[0][2]
        patch_embed = nn.Sequential(
            Conv2D_BN(in_channels, input_channel // 2, 3, 2, 1),
            nn.GELU(),
            Conv2D_BN(input_channel // 2, input_channel, 3, 2, 1),
        )
        layers = [patch_embed]
        # building inverted residual blocks
        block = RepViTBlock
        for k, t, c, use_se, use_hs, s in self.cfgs:
            output_channel = _make_divisible(c, 8)
            exp_size = _make_divisible(input_channel * t, 8)
            layers.append(
                block(input_channel, exp_size, output_channel, k, s, use_se, use_hs)
            )
            input_channel = output_channel
        self.features = nn.LayerList(layers)
        self.out_indices = out_indices
        if out_indices is not None:
            self.out_channels = [self.cfgs[ids - 1][2] for ids in out_indices]
        else:
            self.out_channels = self.cfgs[-1][2]

    def forward(self, x):
        if self.out_indices is not None:
            return self.forward_det(x)
        return self.forward_rec(x)

    def forward_det(self, x):
        outs = []
        for i, f in enumerate(self.features):
            x = f(x)
            if i in self.out_indices:
                outs.append(x)
        return outs

    def forward_rec(self, x):
        for f in self.features:
            x = f(x)
        h = x.shape[2]
        x = nn.functional.avg_pool2d(x, [h, 2])
        return x


def RepSVTR(in_channels=3):
    """
    Constructs a MobileNetV3-Large model
    """
    # k, t, c, SE, HS, s
    cfgs = [
        [3, 2, 96, 1, 0, 1],
        [3, 2, 96, 0, 0, 1],
        [3, 2, 96, 0, 0, 1],
        [3, 2, 192, 0, 1, (2, 1)],
        [3, 2, 192, 1, 1, 1],
        [3, 2, 192, 0, 1, 1],
        [3, 2, 192, 1, 1, 1],
        [3, 2, 192, 0, 1, 1],
        [3, 2, 192, 1, 1, 1],
        [3, 2, 192, 0, 1, 1],
        [3, 2, 384, 0, 1, (2, 1)],
        [3, 2, 384, 1, 1, 1],
        [3, 2, 384, 0, 1, 1],
    ]
    return RepViT(cfgs, in_channels=in_channels)


def RepSVTR_det(in_channels=3, out_indices=[2, 5, 10, 13]):
    """
    Constructs a MobileNetV3-Large model
    """
    # k, t, c, SE, HS, s
    cfgs = [
        [3, 2, 48, 1, 0, 1],
        [3, 2, 48, 0, 0, 1],
        [3, 2, 96, 0, 0, 2],
        [3, 2, 96, 1, 0, 1],
        [3, 2, 96, 0, 0, 1],
        [3, 2, 192, 0, 1, 2],
        [3, 2, 192, 1, 1, 1],
        [3, 2, 192, 0, 1, 1],
        [3, 2, 192, 1, 1, 1],
        [3, 2, 192, 0, 1, 1],
        [3, 2, 384, 0, 1, 2],
        [3, 2, 384, 1, 1, 1],
        [3, 2, 384, 0, 1, 1],
    ]
    return RepViT(cfgs, in_channels=in_channels, out_indices=out_indices)