mobilenetv3.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
from paddle.utils.download import get_weights_path_from_url
from functools import partial

from .utils import _make_divisible
from ..ops import ConvNormActivation

__all__ = []

model_urls = {
    "mobilenet_v3_small_x1.0":
    ("https://paddle-hapi.bj.bcebos.com/models/mobilenet_v3_small_x1.0.pdparams",
     "34fe0e7c1f8b00b2b056ad6788d0590c"),
    "mobilenet_v3_large_x1.0":
    ("https://paddle-hapi.bj.bcebos.com/models/mobilenet_v3_large_x1.0.pdparams",
     "118db5792b4e183b925d8e8e334db3df"),
}


class SqueezeExcitation(nn.Layer):
    """
    This block implements the Squeeze-and-Excitation block from https://arxiv.org/abs/1709.01507 (see Fig. 1).
    Parameters ``activation``, and ``scale_activation`` correspond to ``delta`` and ``sigma`` in eq. 3.
    This code is based on the torchvision code with modifications.
    You can also see at https://github.com/pytorch/vision/blob/main/torchvision/ops/misc.py#L127
    Args:
        input_channels (int): Number of channels in the input image
        squeeze_channels (int): Number of squeeze channels
        activation (Callable[..., paddle.nn.Layer], optional): ``delta`` activation. Default: ``paddle.nn.ReLU``
        scale_activation (Callable[..., paddle.nn.Layer]): ``sigma`` activation. Default: ``paddle.nn.Sigmoid``
    """

    def __init__(self,
                 input_channels,
                 squeeze_channels,
                 activation=nn.ReLU,
                 scale_activation=nn.Sigmoid):
        super().__init__()
        self.avgpool = nn.AdaptiveAvgPool2D(1)
        self.fc1 = nn.Conv2D(input_channels, squeeze_channels, 1)
        self.fc2 = nn.Conv2D(squeeze_channels, input_channels, 1)
        self.activation = activation()
        self.scale_activation = scale_activation()

    def _scale(self, input):
        scale = self.avgpool(input)
        scale = self.fc1(scale)
        scale = self.activation(scale)
        scale = self.fc2(scale)
        return self.scale_activation(scale)

    def forward(self, input):
        scale = self._scale(input)
        return scale * input


class InvertedResidualConfig:
    def __init__(self,
                 in_channels,
                 kernel,
                 expanded_channels,
                 out_channels,
                 use_se,
                 activation,
                 stride,
                 scale=1.0):
        self.in_channels = self.adjust_channels(in_channels, scale=scale)
        self.kernel = kernel
        self.expanded_channels = self.adjust_channels(
            expanded_channels, scale=scale)
        self.out_channels = self.adjust_channels(out_channels, scale=scale)
        self.use_se = use_se
        if activation is None:
            self.activation_layer = None
        elif activation == "relu":
            self.activation_layer = nn.ReLU
        elif activation == "hardswish":
            self.activation_layer = nn.Hardswish
        else:
            raise RuntimeError("The activation function is not supported: {}".
                               format(activation))
        self.stride = stride

    @staticmethod
    def adjust_channels(channels, scale=1.0):
        return _make_divisible(channels * scale, 8)


class InvertedResidual(nn.Layer):
    def __init__(self, in_channels, expanded_channels, out_channels,
                 filter_size, stride, use_se, activation_layer, norm_layer):
        super().__init__()
        self.use_res_connect = stride == 1 and in_channels == out_channels
        self.use_se = use_se
        self.expand = in_channels != expanded_channels

        if self.expand:
            self.expand_conv = ConvNormActivation(
                in_channels=in_channels,
                out_channels=expanded_channels,
                kernel_size=1,
                stride=1,
                padding=0,
                norm_layer=norm_layer,
                activation_layer=activation_layer)

        self.bottleneck_conv = ConvNormActivation(
            in_channels=expanded_channels,
            out_channels=expanded_channels,
            kernel_size=filter_size,
            stride=stride,
            padding=int((filter_size - 1) // 2),
            groups=expanded_channels,
            norm_layer=norm_layer,
            activation_layer=activation_layer)

        if self.use_se:
            self.mid_se = SqueezeExcitation(
                expanded_channels,
                _make_divisible(expanded_channels // 4),
                scale_activation=nn.Hardsigmoid)

        self.linear_conv = ConvNormActivation(
            in_channels=expanded_channels,
            out_channels=out_channels,
            kernel_size=1,
            stride=1,
            padding=0,
            norm_layer=norm_layer,
            activation_layer=None)

    def forward(self, x):
        identity = x
        if self.expand:
            x = self.expand_conv(x)
        x = self.bottleneck_conv(x)
        if self.use_se:
            x = self.mid_se(x)
        x = self.linear_conv(x)
        if self.use_res_connect:
            x = paddle.add(identity, x)
        return x


class MobileNetV3(nn.Layer):
    """MobileNetV3 model from
    `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.

    Args:
        config (list[InvertedResidualConfig]): MobileNetV3 depthwise blocks config.
        last_channel (int): The number of channels on the penultimate layer.
        scale (float, optional): Scale of channels in each layer. Default: 1.0.
        num_classes (int, optional): Output dim of last fc layer. If num_classes <=0, last fc layer
                            will not be defined. Default: 1000.
        with_pool (bool, optional): Use pool before the last fc layer or not. Default: True.
    """

    def __init__(self,
                 config,
                 last_channel,
                 scale=1.0,
                 num_classes=1000,
                 with_pool=True):
        super().__init__()

        self.config = config
        self.scale = scale
        self.last_channel = last_channel
        self.num_classes = num_classes
        self.with_pool = with_pool
        self.firstconv_in_channels = config[0].in_channels
        self.lastconv_in_channels = config[-1].in_channels
        self.lastconv_out_channels = self.lastconv_in_channels * 6
        norm_layer = partial(nn.BatchNorm2D, epsilon=0.001, momentum=0.99)

        self.conv = ConvNormActivation(
            in_channels=3,
            out_channels=self.firstconv_in_channels,
            kernel_size=3,
            stride=2,
            padding=1,
            groups=1,
            activation_layer=nn.Hardswish,
            norm_layer=norm_layer)

        self.blocks = nn.Sequential(*[
            InvertedResidual(
                in_channels=cfg.in_channels,
                expanded_channels=cfg.expanded_channels,
                out_channels=cfg.out_channels,
                filter_size=cfg.kernel,
                stride=cfg.stride,
                use_se=cfg.use_se,
                activation_layer=cfg.activation_layer,
                norm_layer=norm_layer) for cfg in self.config
        ])

        self.lastconv = ConvNormActivation(
            in_channels=self.lastconv_in_channels,
            out_channels=self.lastconv_out_channels,
            kernel_size=1,
            stride=1,
            padding=0,
            groups=1,
            norm_layer=norm_layer,
            activation_layer=nn.Hardswish)

        if with_pool:
            self.avgpool = nn.AdaptiveAvgPool2D(1)

        if num_classes > 0:
            self.classifier = nn.Sequential(
                nn.Linear(self.lastconv_out_channels, self.last_channel),
                nn.Hardswish(),
                nn.Dropout(p=0.2),
                nn.Linear(self.last_channel, num_classes))

    def forward(self, x):
        x = self.conv(x)
        x = self.blocks(x)
        x = self.lastconv(x)

        if self.with_pool:
            x = self.avgpool(x)

        if self.num_classes > 0:
            x = paddle.flatten(x, 1)
            x = self.classifier(x)

        return x


class MobileNetV3Small(MobileNetV3):
    """MobileNetV3 Small architecture model from
    `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.

    Args:
        scale (float, optional): Scale of channels in each layer. Default: 1.0.
        num_classes (int, optional): Output dim of last fc layer. If num_classes <=0, last fc layer
                            will not be defined. Default: 1000.
        with_pool (bool, optional): Use pool before the last fc layer or not. Default: True.

    Examples:
        .. code-block:: python

            import paddle
            from paddle.vision.models import MobileNetV3Small

            # build model
            model = MobileNetV3Small(scale=1.0)

            x = paddle.rand([1, 3, 224, 224])
            out = model(x)

            print(out.shape)
    """

    def __init__(self, scale=1.0, num_classes=1000, with_pool=True):
        config = [
            InvertedResidualConfig(16, 3, 16, 16, True, "relu", 2, scale),
            InvertedResidualConfig(16, 3, 72, 24, False, "relu", 2, scale),
            InvertedResidualConfig(24, 3, 88, 24, False, "relu", 1, scale),
            InvertedResidualConfig(24, 5, 96, 40, True, "hardswish", 2, scale),
            InvertedResidualConfig(40, 5, 240, 40, True, "hardswish", 1, scale),
            InvertedResidualConfig(40, 5, 240, 40, True, "hardswish", 1, scale),
            InvertedResidualConfig(40, 5, 120, 48, True, "hardswish", 1, scale),
            InvertedResidualConfig(48, 5, 144, 48, True, "hardswish", 1, scale),
            InvertedResidualConfig(48, 5, 288, 96, True, "hardswish", 2, scale),
            InvertedResidualConfig(96, 5, 576, 96, True, "hardswish", 1, scale),
            InvertedResidualConfig(96, 5, 576, 96, True, "hardswish", 1, scale),
        ]
        last_channel = _make_divisible(1024 * scale, 8)
        super().__init__(
            config,
            last_channel=last_channel,
            scale=scale,
            with_pool=with_pool,
            num_classes=num_classes)


class MobileNetV3Large(MobileNetV3):
    """MobileNetV3 Large architecture model from
    `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.

    Args:
        scale (float, optional): Scale of channels in each layer. Default: 1.0.
        num_classes (int, optional): Output dim of last fc layer. If num_classes <=0, last fc layer
                            will not be defined. Default: 1000.
        with_pool (bool, optional): Use pool before the last fc layer or not. Default: True.

    Examples:
        .. code-block:: python

            import paddle
            from paddle.vision.models import MobileNetV3Large

            # build model
            model = MobileNetV3Large(scale=1.0)

            x = paddle.rand([1, 3, 224, 224])
            out = model(x)

            print(out.shape)
    """

    def __init__(self, scale=1.0, num_classes=1000, with_pool=True):
        config = [
            InvertedResidualConfig(16, 3, 16, 16, False, "relu", 1, scale),
            InvertedResidualConfig(16, 3, 64, 24, False, "relu", 2, scale),
            InvertedResidualConfig(24, 3, 72, 24, False, "relu", 1, scale),
            InvertedResidualConfig(24, 5, 72, 40, True, "relu", 2, scale),
            InvertedResidualConfig(40, 5, 120, 40, True, "relu", 1, scale),
            InvertedResidualConfig(40, 5, 120, 40, True, "relu", 1, scale),
            InvertedResidualConfig(40, 3, 240, 80, False, "hardswish", 2,
                                   scale),
            InvertedResidualConfig(80, 3, 200, 80, False, "hardswish", 1,
                                   scale),
            InvertedResidualConfig(80, 3, 184, 80, False, "hardswish", 1,
                                   scale),
            InvertedResidualConfig(80, 3, 184, 80, False, "hardswish", 1,
                                   scale),
            InvertedResidualConfig(80, 3, 480, 112, True, "hardswish", 1,
                                   scale),
            InvertedResidualConfig(112, 3, 672, 112, True, "hardswish", 1,
                                   scale),
            InvertedResidualConfig(112, 5, 672, 160, True, "hardswish", 2,
                                   scale),
            InvertedResidualConfig(160, 5, 960, 160, True, "hardswish", 1,
                                   scale),
            InvertedResidualConfig(160, 5, 960, 160, True, "hardswish", 1,
                                   scale),
        ]
        last_channel = _make_divisible(1280 * scale, 8)
        super().__init__(
            config,
            last_channel=last_channel,
            scale=scale,
            with_pool=with_pool,
            num_classes=num_classes)


def _mobilenet_v3(arch, pretrained=False, scale=1.0, **kwargs):
    if arch == "mobilenet_v3_large":
        model = MobileNetV3Large(scale=scale, **kwargs)
    else:
        model = MobileNetV3Small(scale=scale, **kwargs)
    if pretrained:
        arch = "{}_x{}".format(arch, scale)
        assert (
            arch in model_urls
        ), "{} model do not have a pretrained model now, you should set pretrained=False".format(
            arch)
        weight_path = get_weights_path_from_url(model_urls[arch][0],
                                                model_urls[arch][1])

        param = paddle.load(weight_path)
        model.set_dict(param)
    return model


def mobilenet_v3_small(pretrained=False, scale=1.0, **kwargs):
    """MobileNetV3 Small architecture model from
    `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet. Default: False.
        scale (float, optional): Scale of channels in each layer. Default: 1.0.

    Examples:
        .. code-block:: python

            import paddle
            from paddle.vision.models import mobilenet_v3_small

            # build model
            model = mobilenet_v3_small()

            # build model and load imagenet pretrained weight
            # model = mobilenet_v3_small(pretrained=True)

            # build mobilenet v3 small model with scale=0.5
            model = mobilenet_v3_small(scale=0.5)

            x = paddle.rand([1, 3, 224, 224])
            out = model(x)

            print(out.shape)

    """
    model = _mobilenet_v3(
        "mobilenet_v3_small", scale=scale, pretrained=pretrained, **kwargs)
    return model


def mobilenet_v3_large(pretrained=False, scale=1.0, **kwargs):
    """MobileNetV3 Large architecture model from
    `"Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>`_.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet. Default: False.
        scale (float, optional): Scale of channels in each layer. Default: 1.0.

    Examples:
        .. code-block:: python

            import paddle
            from paddle.vision.models import mobilenet_v3_large

            # build model
            model = mobilenet_v3_large()

            # build model and load imagenet pretrained weight
            # model = mobilenet_v3_large(pretrained=True)

            # build mobilenet v3 large model with scale=0.5
            model = mobilenet_v3_large(scale=0.5)

            x = paddle.rand([1, 3, 224, 224])
            out = model(x)

            print(out.shape)

    """
    model = _mobilenet_v3(
        "mobilenet_v3_large", scale=scale, pretrained=pretrained, **kwargs)
    return model