【机器学习】详解 EfficientNet

何处闻韶

已于 2022-05-11 15:22:48 修改

阅读量8.1k

点赞数 15

分类专栏：【机器学习与深度学习】文章标签：机器学习深度学习人工智能

于 2021-07-08 01:08:42 首次发布

本文链接：https://blog.csdn.net/qq_39478403/article/details/117414535

版权

【机器学习与深度学习】专栏收录该内容

39 篇文章 141 订阅

订阅专栏

Paper 地址：https://arxiv.org/abs/1905.11946
PyTorch 实现：https://github.com/lukemelas/EfficientNet-PyTorch
TensorFlow 实现：https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet

一、摘要

卷积神经网络 (ConvNets) 通常是在固定资源预算下开发的，若有更多资源可用，则可扩大规模以提高准确率。本文系统地研究了模型放缩，并证实了 仔细地平衡网络深度、网络宽度和图像分辨率可带来更好性能。基于这一观察，本文提出了一种新的放缩方法，该方法使用了一个简单但高效的 复合系数 (compound coeffificient) 统一地放缩深度/宽度/分辨率的所有维度。本文证明了这种方法在扩大 MobileNets 和 ResNet 方面的有效性。

更进一步地，本文使用 神经架构搜索 (neural architecture search) 来设计一个新的基线网络，并对其扩展以获得一系列称为 EfficientNets 的模型，它比以前的 ConvNets 的准确性和效率更佳。特别地，EfficientNet-B7 在 ImageNet 上达到了 SOTA 的 84.3% 的 top-1 准确率，同时比现有最好的 ConvNet 小 8.4 倍、推理速度快 6.1 倍。EfficientNets 在 CIFAR-100 (91.7%)、Flowers (98.8%) 和其他 3 个迁移学习数据集上也能很好地迁移，并达到 SOTA 的准确率，而参数少了一个数量级。

二、介绍

扩大 ConvNets 被广泛用于获得更佳精度。例如，ResNet 可通过使用更多层从 ResNet-18 扩大到 ResNet-200；最近，GPipe 通过扩大基线模型 4 倍，获得了 84.3% 的ImageNet top-1 准确度。然而，扩大 ConvNets 的过程从未被很好地理解过，目前有很多方法可以做到这一点。最常见的方法是通过深度或宽度扩大 ConvNets。另一种不太常见但越来越流行的方法是通过图像分辨率扩大模型。以往，通常只按深度/宽度/图像尺寸这三个维度之一来放缩。尽管可以任意放缩两个或三个维度，但需要繁琐的手工调优，且常无法达到最佳的准确度和效率。

在本文中，我们想研究和重新思考扩大 ConvNets 的过程。特别地，我们调查中心问题：是否存在一个原则性方法来扩大 ConvNets，以达到更好的准确率和效率？我们的实证研究表明，平衡网络宽度/深度/分辨率的所有维度是至关重要的，且令人惊讶的是，这种平衡可通过简单地使用固定比例缩放来实现。基于此观察，我们提出了一种简单而有效的复合放缩方法。有别于传统实践中的任意放缩这些因子，我们的方法均匀/统一地缩放网络的宽度、深度及一组固定放缩系数的分辨率。例如，如果我们想使用 $2^N$ 倍以上的计算资源，那么我们可以简单地将网络深度增加 $\alpha ^ N$ ，宽度增加 $\beta ^ N$ ，图像大小增加 $\gamma ^ N$ ，其中 $\alpha$ ， $\beta$ ， $\gamma$ 是在原始小模型上进行小网格搜索所确定的常系数。图 2 阐述了我们的放缩方法和传统方法之间的区别。

直观上，复合放缩方法是有意义的，因为若输入图像更大，那么网络需要更多层来增加感受野和更多的通道来捕捉更细粒度的模式。事实上，之前的理论和实证结果均表明网络宽度和深度之间存在一定的关系，但据我们所知，我们是第一个对网络宽度、深度和分辨率三个维度之间的关系进行实证量化的人。

本文证明了上述复合缩放方法在 MobileNets 和 ResNet 上工作得很好。值得注意的是，模型缩放的有效性在很大程度上依赖于基线网络。更进一步地，我们使用 神经架构搜索 来开发一个新的基线网络，并将其扩展以获得一系列称为 EfficientNets 的模型。图 1 总结了 ImageNet 的性能，其中 EfficientNets 明显优于其他 ConvNets。特别地，EfficientNets-B7 超过了现有的最佳 GPipe 准确度，却减少了 8.4 倍参数、提高了 6.1倍推理速度。与广泛使用的 ResNet-50 相比，EffcientNet-B4 在类似的 FLOPs 下将 top-1 准确度从 76.3% 提高到 83.0%(+6.7%)。除 ImageNet，EfficientNets 也有很好的迁移能力，在 8 个广泛使用的数据集中的 5 个上实现了 SOTA 的准确度，同时减少了比现有 ConvNets 至多 21 倍的参数。

三、网络放缩

3.1 问题构造

3.2 维度放缩

问题 2 的主要难点是最优的 $d$ ， $w$ ， $r$ 相互依赖，在不同的资源约束下，其值会发生变化。由于该困难，传统方法主要在以下维度之一中扩大 ConvNets：

深度 $d$

最常用的方法是放缩网络深度，这被许多 ConvNets 使用。直观地，更深的 ConvNets 可以捕获更丰富、更复杂的特征，并很好地泛化与新的任务。然而，由于梯度消失，深度网络的训练也更加困难。虽然诸如 Skip Connection 和 BN 等技术缓解了训练问题，极深网络的准确率增益仍会降低。例如，ResNet-1000 和 ResNet-101 有相似的准确度，尽管前者有更多的层。图3 (中) 显示了对不同深度系数 $d$ 的基线模型的放缩实证研究，进一步表明 极深的 ConvNets 的准确率回报减少了。

宽度 $w$

放缩网络宽度广泛用于小型模型。更宽的网络往往能够捕捉到更细粒度的特征，也更容易训练。然而。极宽但浅的网络在捕获更高层次的特征时往往存在困难。我们的经验结果显示于 图3 (左) ，当网络因具有更大的 $w$ 而变得更宽时，准确率很快饱和。

分辨率 $r$

有了更高分辨率的输入图像，ConvNets 可以潜在地捕获更细粒度的模式。从早期 ConvNets 的 224x224 开始，现代的 ConvNets 倾向于使用 299x299 或 331x331 以提高准确率。最近，GPipe 实现了 SOTA 的 ImageNet 准确率，分辨率为 480x480。如 600x600 等更高分辨率，也被广泛应用于目标检测 ConvNets 中。图3 (右) 显示了网络分辨率放缩的结果，更高的分辨率确实提高了准确率，但是对于非常高的分辨率，准确率增益会降低 (分辨率 $r = 1.0$ 表示 224x224， $r = 2.5$ 表示 560x560)。

以上分析使我们得出第一个结论:

观察结论 1：扩大网络宽度、深度或分辨率的任意维度都可以提高准确度，但对于更大的模型，准确度增益会降低。

3.3 复合放缩

我们从经验上观察到，不同的放缩维度不是独立的。直观地，对于分辨率较高的图像，应增加网络深度，从而更大的感受野可以在更大图像中有助于捕获包含更多像素的相似特征。相应地，当分辨率较高时，也应增加网络宽度，以便于在高分辨率图像中捕获具有更多像素的更多细粒度模式。这些直觉表明，我们需要协调和平衡不同的放缩维度，而非传统的一维放缩。

为验证我们的直觉，我们比较了不同网络深度和分辨率下的宽度放缩，如图4所示。如果只放缩网络宽度 $w$ 而不改变深度 ( $d = 1.0$ ) 和分辨率 ( $r = 1.0$ )，准确率会很快饱和。具有更深的 ( $d = 2.0$ ) 和更高的分辨率 ( $r = 2.0$ )，宽度缩放在相同的 FLOPs 代价下达到更好的准确率。

以上结果使我们得出第二个结论：

观察结论2：为追求更高的准确率和效率，在放缩 ConVnet 时，关键是要平衡网络宽度、深度和分辨率的各个维度。

事实上，之前的一些工作已尝试过任意平衡网络宽度和深度，但均需繁琐的手动调整。

本文提出了一种新的复合放缩方法，利用复合系数有原则地均匀/统一放缩网络宽度、深度和分辨率：

其中， $\alpha$ ， $\beta$ ， $\gamma$ 均为常数，可由小网格搜索决定。直观地， $\phi$ 是一个用户特定的系数，用于控制多少资源可用于模型放缩，而 $\alpha$ ， $\beta$ ， $\gamma$ 表明如何分配这些额外资源以分别用于网络宽度、深度和分辨率。值得注意的是，常规卷积操作的 FLOPS 与 $d$ ， $w^2$ ， $r^2$ 成正比，即令网络深度加倍将使 FLOPS 加倍，但使网络宽度或分辨率加倍则将使 FLOPS 增加 4 倍。因为卷积操作通常决定了 ConvNets 的运算成本，根据上式放缩一个 ConvNet 将近似增加 $(\alpha \cdot \beta^2 \cdot \gamma^2)^\phi$ 的总 FLOPS。在本文，我们约束 $\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$ 以便于对任意新 $\phi$ ，总 FLOPS 将近似增加 $2^\phi$ 。

四、网络架构

由于模型缩放不会改变基线网络中的层算子 $\hat{F_i}$ ，因此具有一个良好的基线网络也是至关重要的。我们将使用现有的 ConyNets 来评估我们的放缩方法，但为更好地证明我们的放缩方法的有效性。我们还开发了一种新的手机尺寸的基准 —— EfficientNet。

受到启发，我们通过利用 多目标 NAS 来优化基线网络，这种搜索既优化了准确度，也优化了 FLOPS。具体地，使用相同的搜索空间并用 $ACC(m) \times [FLOPS(m)/T]^w$ 作为优化目标，其中 $ACC(m)$ 和 $FLOPS(m)$ 表示模型 $m$ 的准确率和 FLOPS， $T$ 是目标 FLOPS 且 $w=-0.07$ 是一个超参数，用于控制准确率和 FLOPS 之间的权衡。不同于之前的工作，这里我们优化的是 FLOPS 而非延迟，因为我们不针对任何特定的硬件设备。我们的搜索产生了一个高效的网络 —— EfficientNet - B0。由于我们使用相同的搜索空间，架构类似于 MnasNet，但 EfficientNet - B0 稍微大一些，因其 FLOPS 目标更大 (我们的 FLOPS 目标是 400M)。表 1 显示了 EfficientNet - B0 的架构。其主要构造块 (Building Block) 为 mobile inverted bottleneck MBConv，我们也对其加入了 通道注意力 进行优化。

从基线 EfficientNet - B0 开始，我们通过两个步骤应用我们的复合放缩方法来扩展它：

步骤1：首先固定 $\phi = 1$ ，假设有两倍以上的资源可用，则基于式 2 和 3 对 $\alpha$ ， $\beta$ ， $\gamma$ 进行小网格搜索。特别地，在 $\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$ 的约束下，我们发现 EfficientNet - B0 的最佳值为 $\alpha=1.2, \beta=1.1, \gamma=1.15$ 。
步骤2：然后将 $\alpha$ ， $\beta$ ， $\gamma$ 固定为常数，并根据式 3 使用不同的 $\phi$ 扩大基线网络，从而得到 EfficientNet - B1 ~ EfficientNet - B7 (详见表2)。

值得注意的是，通过直接围绕大模型搜索 $\alpha$ ， $\beta$ ， $\gamma$ 有可能获得更好的性能，但在大模型上的搜索成本会变得昂贵得令人难以接受。我们的方法解决了该问题，即 只在小的基线网络上先进行一次搜索 (步骤1)，然后对所有其他模型使用相同的放缩系数 (步骤2)。

五、对比实验

六、架构细节

6.1 图示

preview

任何网络的 Stem 都是最关键的，确定后才会进行后续实验，上图的 头部/尾部架构 在 EfficientNet - B1 ~ EfficientNet - B7 都有出现，这些块还有不同数量的子块，子块数随版本的增加而增加。以下代码可用于查看模型架构：

import tensorflow as tf
IMG_SHAPE = (224, 224, 3)
model0 = tf.keras.applications.EfficientNetB0(input_shape=IMG_SHAPE, include_top=False, weights="imagenet")
tf.keras.utils.plot_model(model0)  # to draw and visualize
model0.summary()  # to see the list of layers and parameters

其中，EfficientNet-B0 总层数为 237、EfficientNet-B7 总层数为 813，但其实都可由以下 5 个模块和上述 2 个头部/尾部架构构成。

preview

由上述 5 个模块 (Module) 构成以下三种块 (Block)：

至此，所有用于构建 EfficientNet 的块已经确定，以下是各模型的架构 (×n means that modules inside the bracket are repeated n times)：

EfficientNet-B0

preview

EfficientNet-B1

preview

EfficientNet-B2

结构与 EfficientNet-B1 相同，唯一的区别是通道数不同，从而增加了参数量。

EfficientNet-B3

preview

EfficientNet-B4

preview

EfficientNet-B5

preview

EfficientNet-B6

preview

EfficientNet-B7

preview

综上，容易看出各模型之间的差异 —— 子块数量逐渐增多。下表显示了 EfficientNet-B0 中卷积核尺寸、分辨率大小、通道数和层数：

preview

下表展示了不同网络架构下，各阶段的通道数 (各阶段分辨率大小一一对应)：

preview

6.2 代码

EfficientNet-PyTorch/efficientnet_pytorch/model.py

"""model.py - Model and module class for EfficientNet.
   They are built to mirror those in the official TensorFlow implementation.
"""

# Author: lukemelas (github username)
# Github repo: https://github.com/lukemelas/EfficientNet-PyTorch
# With adjustments and added comments by workingcoder (github username).

import torch
from torch import nn
from torch.nn import functional as F
from .utils import (
    round_filters,
    round_repeats,
    drop_connect,
    get_same_padding_conv2d,
    get_model_params,
    efficientnet_params,
    load_pretrained_weights,
    Swish,
    MemoryEfficientSwish,
    calculate_output_image_size
)


VALID_MODELS = (
    'efficientnet-b0', 'efficientnet-b1', 'efficientnet-b2', 'efficientnet-b3',
    'efficientnet-b4', 'efficientnet-b5', 'efficientnet-b6', 'efficientnet-b7',
    'efficientnet-b8',

    # Support the construction of 'efficientnet-l2' without pretrained weights
    'efficientnet-l2'
)


class MBConvBlock(nn.Module):
    """Mobile Inverted Residual Bottleneck Block.
    Args:
        block_args (namedtuple): BlockArgs, defined in utils.py.
        global_params (namedtuple): GlobalParam, defined in utils.py.
        image_size (tuple or list): [image_height, image_width].
    References:
        [1] https://arxiv.org/abs/1704.04861 (MobileNet v1)
        [2] https://arxiv.org/abs/1801.04381 (MobileNet v2)
        [3] https://arxiv.org/abs/1905.02244 (MobileNet v3)
    """

    def __init__(self, block_args, global_params, image_size=None):
        super().__init__()
        self._block_args = block_args
        self._bn_mom = 1 - global_params.batch_norm_momentum  # pytorch's difference from tensorflow
        self._bn_eps = global_params.batch_norm_epsilon
        self.has_se = (self._block_args.se_ratio is not None) and (0 < self._block_args.se_ratio <= 1)
        self.id_skip = block_args.id_skip  # whether to use skip connection and drop connect

        # Expansion phase (Inverted Bottleneck)
        inp = self._block_args.input_filters  # number of input channels
        oup = self._block_args.input_filters * self._block_args.expand_ratio  # number of output channels
        if self._block_args.expand_ratio != 1:
            Conv2d = get_same_padding_conv2d(image_size=image_size)
            self._expand_conv = Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)
            self._bn0 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
            # image_size = calculate_output_image_size(image_size, 1) <-- this wouldn't modify image_size

        # Depthwise convolution phase
        k = self._block_args.kernel_size
        s = self._block_args.stride
        Conv2d = get_same_padding_conv2d(image_size=image_size)
        self._depthwise_conv = Conv2d(
            in_channels=oup, out_channels=oup, groups=oup,  # groups makes it depthwise
            kernel_size=k, stride=s, bias=False)
        self._bn1 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
        image_size = calculate_output_image_size(image_size, s)

        # Squeeze and Excitation layer, if desired
        if self.has_se:
            Conv2d = get_same_padding_conv2d(image_size=(1, 1))
            num_squeezed_channels = max(1, int(self._block_args.input_filters * self._block_args.se_ratio))
            self._se_reduce = Conv2d(in_channels=oup, out_channels=num_squeezed_channels, kernel_size=1)
            self._se_expand = Conv2d(in_channels=num_squeezed_channels, out_channels=oup, kernel_size=1)

        # Pointwise convolution phase
        final_oup = self._block_args.output_filters
        Conv2d = get_same_padding_conv2d(image_size=image_size)
        self._project_conv = Conv2d(in_channels=oup, out_channels=final_oup, kernel_size=1, bias=False)
        self._bn2 = nn.BatchNorm2d(num_features=final_oup, momentum=self._bn_mom, eps=self._bn_eps)
        self._swish = MemoryEfficientSwish()

    def forward(self, inputs, drop_connect_rate=None):
        """MBConvBlock's forward function.
        Args:
            inputs (tensor): Input tensor.
            drop_connect_rate (bool): Drop connect rate (float, between 0 and 1).
        Returns:
            Output of this block after processing.
        """

        # Expansion and Depthwise Convolution
        x = inputs
        if self._block_args.expand_ratio != 1:
            x = self._expand_conv(inputs)
            x = self._bn0(x)
            x = self._swish(x)

        x = self._depthwise_conv(x)
        x = self._bn1(x)
        x = self._swish(x)

        # Squeeze and Excitation
        if self.has_se:
            x_squeezed = F.adaptive_avg_pool2d(x, 1)
            x_squeezed = self._se_reduce(x_squeezed)
            x_squeezed = self._swish(x_squeezed)
            x_squeezed = self._se_expand(x_squeezed)
            x = torch.sigmoid(x_squeezed) * x

        # Pointwise Convolution
        x = self._project_conv(x)
        x = self._bn2(x)

        # Skip connection and drop connect
        input_filters, output_filters = self._block_args.input_filters, self._block_args.output_filters
        if self.id_skip and self._block_args.stride == 1 and input_filters == output_filters:
            # The combination of skip connection and drop connect brings about stochastic depth.
            if drop_connect_rate:
                x = drop_connect(x, p=drop_connect_rate, training=self.training)
            x = x + inputs  # skip connection
        return x

    def set_swish(self, memory_efficient=True):
        """Sets swish function as memory efficient (for training) or standard (for export).
        Args:
            memory_efficient (bool): Whether to use memory-efficient version of swish.
        """
        self._swish = MemoryEfficientSwish() if memory_efficient else Swish()


class EfficientNet(nn.Module):
    """EfficientNet model.
       Most easily loaded with the .from_name or .from_pretrained methods.
    Args:
        blocks_args (list[namedtuple]): A list of BlockArgs to construct blocks.
        global_params (namedtuple): A set of GlobalParams shared between blocks.
    References:
        [1] https://arxiv.org/abs/1905.11946 (EfficientNet)
    Example:
        >>> import torch
        >>> from efficientnet.model import EfficientNet
        >>> inputs = torch.rand(1, 3, 224, 224)
        >>> model = EfficientNet.from_pretrained('efficientnet-b0')
        >>> model.eval()
        >>> outputs = model(inputs)
    """

    def __init__(self, blocks_args=None, global_params=None):
        super().__init__()
        assert isinstance(blocks_args, list), 'blocks_args should be a list'
        assert len(blocks_args) > 0, 'block args must be greater than 0'
        self._global_params = global_params
        self._blocks_args = blocks_args

        # Batch norm parameters
        bn_mom = 1 - self._global_params.batch_norm_momentum
        bn_eps = self._global_params.batch_norm_epsilon

        # Get stem static or dynamic convolution depending on image size
        image_size = global_params.image_size
        Conv2d = get_same_padding_conv2d(image_size=image_size)

        # Stem
        in_channels = 3  # rgb
        out_channels = round_filters(32, self._global_params)  # number of output channels
        self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)
        self._bn0 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
        image_size = calculate_output_image_size(image_size, 2)

        # Build blocks
        self._blocks = nn.ModuleList([])
        for block_args in self._blocks_args:

            # Update block input and output filters based on depth multiplier.
            block_args = block_args._replace(
                input_filters=round_filters(block_args.input_filters, self._global_params),
                output_filters=round_filters(block_args.output_filters, self._global_params),
                num_repeat=round_repeats(block_args.num_repeat, self._global_params)
            )

            # The first block needs to take care of stride and filter size increase.
            self._blocks.append(MBConvBlock(block_args, self._global_params, image_size=image_size))
            image_size = calculate_output_image_size(image_size, block_args.stride)
            if block_args.num_repeat > 1:  # modify block_args to keep same output size
                block_args = block_args._replace(input_filters=block_args.output_filters, stride=1)
            for _ in range(block_args.num_repeat - 1):
                self._blocks.append(MBConvBlock(block_args, self._global_params, image_size=image_size))
                # image_size = calculate_output_image_size(image_size, block_args.stride)  # stride = 1

        # Head
        in_channels = block_args.output_filters  # output of final block
        out_channels = round_filters(1280, self._global_params)
        Conv2d = get_same_padding_conv2d(image_size=image_size)
        self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        self._bn1 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)

        # Final linear layer
        self._avg_pooling = nn.AdaptiveAvgPool2d(1)
        if self._global_params.include_top:
            self._dropout = nn.Dropout(self._global_params.dropout_rate)
            self._fc = nn.Linear(out_channels, self._global_params.num_classes)

        # set activation to memory efficient swish by default
        self._swish = MemoryEfficientSwish()

    def set_swish(self, memory_efficient=True):
        """Sets swish function as memory efficient (for training) or standard (for export).
        Args:
            memory_efficient (bool): Whether to use memory-efficient version of swish.
        """
        self._swish = MemoryEfficientSwish() if memory_efficient else Swish()
        for block in self._blocks:
            block.set_swish(memory_efficient)

    def extract_endpoints(self, inputs):
        """Use convolution layer to extract features
        from reduction levels i in [1, 2, 3, 4, 5].
        Args:
            inputs (tensor): Input tensor.
        Returns:
            Dictionary of last intermediate features
            with reduction levels i in [1, 2, 3, 4, 5].
            Example:
                >>> import torch
                >>> from efficientnet.model import EfficientNet
                >>> inputs = torch.rand(1, 3, 224, 224)
                >>> model = EfficientNet.from_pretrained('efficientnet-b0')
                >>> endpoints = model.extract_endpoints(inputs)
                >>> print(endpoints['reduction_1'].shape)  # torch.Size([1, 16, 112, 112])
                >>> print(endpoints['reduction_2'].shape)  # torch.Size([1, 24, 56, 56])
                >>> print(endpoints['reduction_3'].shape)  # torch.Size([1, 40, 28, 28])
                >>> print(endpoints['reduction_4'].shape)  # torch.Size([1, 112, 14, 14])
                >>> print(endpoints['reduction_5'].shape)  # torch.Size([1, 320, 7, 7])
                >>> print(endpoints['reduction_6'].shape)  # torch.Size([1, 1280, 7, 7])
        """
        endpoints = dict()

        # Stem
        x = self._swish(self._bn0(self._conv_stem(inputs)))
        prev_x = x

        # Blocks
        for idx, block in enumerate(self._blocks):
            drop_connect_rate = self._global_params.drop_connect_rate
            if drop_connect_rate:
                drop_connect_rate *= float(idx) / len(self._blocks)  # scale drop connect_rate
            x = block(x, drop_connect_rate=drop_connect_rate)
            if prev_x.size(2) > x.size(2):
                endpoints['reduction_{}'.format(len(endpoints) + 1)] = prev_x
            elif idx == len(self._blocks) - 1:
                endpoints['reduction_{}'.format(len(endpoints) + 1)] = x
            prev_x = x

        # Head
        x = self._swish(self._bn1(self._conv_head(x)))
        endpoints['reduction_{}'.format(len(endpoints) + 1)] = x

        return endpoints

    def extract_features(self, inputs):
        """use convolution layer to extract feature .
        Args:
            inputs (tensor): Input tensor.
        Returns:
            Output of the final convolution
            layer in the efficientnet model.
        """
        # Stem
        x = self._swish(self._bn0(self._conv_stem(inputs)))

        # Blocks
        for idx, block in enumerate(self._blocks):
            drop_connect_rate = self._global_params.drop_connect_rate
            if drop_connect_rate:
                drop_connect_rate *= float(idx) / len(self._blocks)  # scale drop connect_rate
            x = block(x, drop_connect_rate=drop_connect_rate)

        # Head
        x = self._swish(self._bn1(self._conv_head(x)))

        return x

    def forward(self, inputs):
        """EfficientNet's forward function.
           Calls extract_features to extract features, applies final linear layer, and returns logits.
        Args:
            inputs (tensor): Input tensor.
        Returns:
            Output of this model after processing.
        """
        # Convolution layers
        x = self.extract_features(inputs)
        # Pooling and final linear layer
        x = self._avg_pooling(x)
        if self._global_params.include_top:
            x = x.flatten(start_dim=1)
            x = self._dropout(x)
            x = self._fc(x)
        return x

    @classmethod
    def from_name(cls, model_name, in_channels=3, **override_params):
        """Create an efficientnet model according to name.
        Args:
            model_name (str): Name for efficientnet.
            in_channels (int): Input data's channel number.
            override_params (other key word params):
                Params to override model's global_params.
                Optional key:
                    'width_coefficient', 'depth_coefficient',
                    'image_size', 'dropout_rate',
                    'num_classes', 'batch_norm_momentum',
                    'batch_norm_epsilon', 'drop_connect_rate',
                    'depth_divisor', 'min_depth'
        Returns:
            An efficientnet model.
        """
        cls._check_model_name_is_valid(model_name)
        blocks_args, global_params = get_model_params(model_name, override_params)
        model = cls(blocks_args, global_params)
        model._change_in_channels(in_channels)
        return model

    @classmethod
    def from_pretrained(cls, model_name, weights_path=None, advprop=False,
                        in_channels=3, num_classes=1000, **override_params):
        """Create an efficientnet model according to name.
        Args:
            model_name (str): Name for efficientnet.
            weights_path (None or str):
                str: path to pretrained weights file on the local disk.
                None: use pretrained weights downloaded from the Internet.
            advprop (bool):
                Whether to load pretrained weights
                trained with advprop (valid when weights_path is None).
            in_channels (int): Input data's channel number.
            num_classes (int):
                Number of categories for classification.
                It controls the output size for final linear layer.
            override_params (other key word params):
                Params to override model's global_params.
                Optional key:
                    'width_coefficient', 'depth_coefficient',
                    'image_size', 'dropout_rate',
                    'batch_norm_momentum',
                    'batch_norm_epsilon', 'drop_connect_rate',
                    'depth_divisor', 'min_depth'
        Returns:
            A pretrained efficientnet model.
        """
        model = cls.from_name(model_name, num_classes=num_classes, **override_params)
        load_pretrained_weights(model, model_name, weights_path=weights_path,
                                load_fc=(num_classes == 1000), advprop=advprop)
        model._change_in_channels(in_channels)
        return model

    @classmethod
    def get_image_size(cls, model_name):
        """Get the input image size for a given efficientnet model.
        Args:
            model_name (str): Name for efficientnet.
        Returns:
            Input image size (resolution).
        """
        cls._check_model_name_is_valid(model_name)
        _, _, res, _ = efficientnet_params(model_name)
        return res

    @classmethod
    def _check_model_name_is_valid(cls, model_name):
        """Validates model name.
        Args:
            model_name (str): Name for efficientnet.
        Returns:
            bool: Is a valid name or not.
        """
        if model_name not in VALID_MODELS:
            raise ValueError('model_name should be one of: ' + ', '.join(VALID_MODELS))

    def _change_in_channels(self, in_channels):
        """Adjust model's first convolution layer to in_channels, if in_channels not equals 3.
        Args:
            in_channels (int): Input data's channel number.
        """
        if in_channels != 3:
            Conv2d = get_same_padding_conv2d(image_size=self._global_params.image_size)
            out_channels = round_filters(32, self._global_params)
            self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)

EfficientNet-PyTorch/efficientnet_pytorch/utils.py

"""utils.py - Helper functions for building the model and for loading model parameters.
   These helper functions are built to mirror those in the official TensorFlow implementation.
"""

# Author: lukemelas (github username)
# Github repo: https://github.com/lukemelas/EfficientNet-PyTorch
# With adjustments and added comments by workingcoder (github username).

import re
import math
import collections
from functools import partial
import torch
from torch import nn
from torch.nn import functional as F
from torch.utils import model_zoo


################################################################################
# Help functions for model architecture
################################################################################

# GlobalParams and BlockArgs: Two namedtuples
# Swish and MemoryEfficientSwish: Two implementations of the method
# round_filters and round_repeats:
#     Functions to calculate params for scaling model width and depth ! ! !
# get_width_and_height_from_size and calculate_output_image_size
# drop_connect: A structural design
# get_same_padding_conv2d:
#     Conv2dDynamicSamePadding
#     Conv2dStaticSamePadding
# get_same_padding_maxPool2d:
#     MaxPool2dDynamicSamePadding
#     MaxPool2dStaticSamePadding
#     It's an additional function, not used in EfficientNet,
#     but can be used in other model (such as EfficientDet).

# Parameters for the entire model (stem, all blocks, and head)
GlobalParams = collections.namedtuple('GlobalParams', [
    'width_coefficient', 'depth_coefficient', 'image_size', 'dropout_rate',
    'num_classes', 'batch_norm_momentum', 'batch_norm_epsilon',
    'drop_connect_rate', 'depth_divisor', 'min_depth', 'include_top'])

# Parameters for an individual model block
BlockArgs = collections.namedtuple('BlockArgs', [
    'num_repeat', 'kernel_size', 'stride', 'expand_ratio',
    'input_filters', 'output_filters', 'se_ratio', 'id_skip'])

# Set GlobalParams and BlockArgs's defaults
GlobalParams.__new__.__defaults__ = (None,) * len(GlobalParams._fields)
BlockArgs.__new__.__defaults__ = (None,) * len(BlockArgs._fields)

# Swish activation function
if hasattr(nn, 'SiLU'):
    Swish = nn.SiLU
else:
    # For compatibility with old PyTorch versions
    class Swish(nn.Module):
        def forward(self, x):
            return x * torch.sigmoid(x)


# A memory-efficient implementation of Swish function
class SwishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, i):
        result = i * torch.sigmoid(i)
        ctx.save_for_backward(i)
        return result

    @staticmethod
    def backward(ctx, grad_output):
        i = ctx.saved_tensors[0]
        sigmoid_i = torch.sigmoid(i)
        return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))


class MemoryEfficientSwish(nn.Module):
    def forward(self, x):
        return SwishImplementation.apply(x)


def round_filters(filters, global_params):
    """Calculate and round number of filters based on width multiplier.
       Use width_coefficient, depth_divisor and min_depth of global_params.
    Args:
        filters (int): Filters number to be calculated.
        global_params (namedtuple): Global params of the model.
    Returns:
        new_filters: New filters number after calculating.
    """
    multiplier = global_params.width_coefficient
    if not multiplier:
        return filters
    # TODO: modify the params names.
    #       maybe the names (width_divisor,min_width)
    #       are more suitable than (depth_divisor,min_depth).
    divisor = global_params.depth_divisor
    min_depth = global_params.min_depth
    filters *= multiplier
    min_depth = min_depth or divisor  # pay attention to this line when using min_depth
    # follow the formula transferred from official TensorFlow implementation
    new_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor)
    if new_filters < 0.9 * filters:  # prevent rounding by more than 10%
        new_filters += divisor
    return int(new_filters)


def round_repeats(repeats, global_params):
    """Calculate module's repeat number of a block based on depth multiplier.
       Use depth_coefficient of global_params.
    Args:
        repeats (int): num_repeat to be calculated.
        global_params (namedtuple): Global params of the model.
    Returns:
        new repeat: New repeat number after calculating.
    """
    multiplier = global_params.depth_coefficient
    if not multiplier:
        return repeats
    # follow the formula transferred from official TensorFlow implementation
    return int(math.ceil(multiplier * repeats))


def drop_connect(inputs, p, training):
    """Drop connect.
    Args:
        input (tensor: BCWH): Input of this structure.
        p (float: 0.0~1.0): Probability of drop connection.
        training (bool): The running mode.
    Returns:
        output: Output after drop connection.
    """
    assert 0 <= p <= 1, 'p must be in range of [0,1]'

    if not training:
        return inputs

    batch_size = inputs.shape[0]
    keep_prob = 1 - p

    # generate binary_tensor mask according to probability (p for 0, 1-p for 1)
    random_tensor = keep_prob
    random_tensor += torch.rand([batch_size, 1, 1, 1], dtype=inputs.dtype, device=inputs.device)
    binary_tensor = torch.floor(random_tensor)

    output = inputs / keep_prob * binary_tensor
    return output


def get_width_and_height_from_size(x):
    """Obtain height and width from x.
    Args:
        x (int, tuple or list): Data size.
    Returns:
        size: A tuple or list (H,W).
    """
    if isinstance(x, int):
        return x, x
    if isinstance(x, list) or isinstance(x, tuple):
        return x
    else:
        raise TypeError()


def calculate_output_image_size(input_image_size, stride):
    """Calculates the output image size when using Conv2dSamePadding with a stride.
       Necessary for static padding. Thanks to mannatsingh for pointing this out.
    Args:
        input_image_size (int, tuple or list): Size of input image.
        stride (int, tuple or list): Conv2d operation's stride.
    Returns:
        output_image_size: A list [H,W].
    """
    if input_image_size is None:
        return None
    image_height, image_width = get_width_and_height_from_size(input_image_size)
    stride = stride if isinstance(stride, int) else stride[0]
    image_height = int(math.ceil(image_height / stride))
    image_width = int(math.ceil(image_width / stride))
    return [image_height, image_width]


# Note:
# The following 'SamePadding' functions make output size equal ceil(input size/stride).
# Only when stride equals 1, can the output size be the same as input size.
# Don't be confused by their function names ! ! !

def get_same_padding_conv2d(image_size=None):
    """Chooses static padding if you have specified an image size, and dynamic padding otherwise.
       Static padding is necessary for ONNX exporting of models.
    Args:
        image_size (int or tuple): Size of the image.
    Returns:
        Conv2dDynamicSamePadding or Conv2dStaticSamePadding.
    """
    if image_size is None:
        return Conv2dDynamicSamePadding
    else:
        return partial(Conv2dStaticSamePadding, image_size=image_size)


class Conv2dDynamicSamePadding(nn.Conv2d):
    """2D Convolutions like TensorFlow, for a dynamic image size.
       The padding is operated in forward function by calculating dynamically.
    """

    # Tips for 'SAME' mode padding.
    #     Given the following:
    #         i: width or height
    #         s: stride
    #         k: kernel size
    #         d: dilation
    #         p: padding
    #     Output after Conv2d:
    #         o = floor((i+p-((k-1)*d+1))/s+1)
    # If o equals i, i = floor((i+p-((k-1)*d+1))/s+1),
    # => p = (i-1)*s+((k-1)*d+1)-i

    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, groups=1, bias=True):
        super().__init__(in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias)
        self.stride = self.stride if len(self.stride) == 2 else [self.stride[0]] * 2

    def forward(self, x):
        ih, iw = x.size()[-2:]
        kh, kw = self.weight.size()[-2:]
        sh, sw = self.stride
        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)  # change the output size according to stride ! ! !
        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
        if pad_h > 0 or pad_w > 0:
            x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2])
        return F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)


class Conv2dStaticSamePadding(nn.Conv2d):
    """2D Convolutions like TensorFlow's 'SAME' mode, with the given input image size.
       The padding mudule is calculated in construction function, then used in forward.
    """

    # With the same calculation as Conv2dDynamicSamePadding

    def __init__(self, in_channels, out_channels, kernel_size, stride=1, image_size=None, **kwargs):
        super().__init__(in_channels, out_channels, kernel_size, stride, **kwargs)
        self.stride = self.stride if len(self.stride) == 2 else [self.stride[0]] * 2

        # Calculate padding based on image size and save it
        assert image_size is not None
        ih, iw = (image_size, image_size) if isinstance(image_size, int) else image_size
        kh, kw = self.weight.size()[-2:]
        sh, sw = self.stride
        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)
        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
        if pad_h > 0 or pad_w > 0:
            self.static_padding = nn.ZeroPad2d((pad_w // 2, pad_w - pad_w // 2,
                                                pad_h // 2, pad_h - pad_h // 2))
        else:
            self.static_padding = nn.Identity()

    def forward(self, x):
        x = self.static_padding(x)
        x = F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
        return x


def get_same_padding_maxPool2d(image_size=None):
    """Chooses static padding if you have specified an image size, and dynamic padding otherwise.
       Static padding is necessary for ONNX exporting of models.
    Args:
        image_size (int or tuple): Size of the image.
    Returns:
        MaxPool2dDynamicSamePadding or MaxPool2dStaticSamePadding.
    """
    if image_size is None:
        return MaxPool2dDynamicSamePadding
    else:
        return partial(MaxPool2dStaticSamePadding, image_size=image_size)


class MaxPool2dDynamicSamePadding(nn.MaxPool2d):
    """2D MaxPooling like TensorFlow's 'SAME' mode, with a dynamic image size.
       The padding is operated in forward function by calculating dynamically.
    """

    def __init__(self, kernel_size, stride, padding=0, dilation=1, return_indices=False, ceil_mode=False):
        super().__init__(kernel_size, stride, padding, dilation, return_indices, ceil_mode)
        self.stride = [self.stride] * 2 if isinstance(self.stride, int) else self.stride
        self.kernel_size = [self.kernel_size] * 2 if isinstance(self.kernel_size, int) else self.kernel_size
        self.dilation = [self.dilation] * 2 if isinstance(self.dilation, int) else self.dilation

    def forward(self, x):
        ih, iw = x.size()[-2:]
        kh, kw = self.kernel_size
        sh, sw = self.stride
        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)
        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
        if pad_h > 0 or pad_w > 0:
            x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2])
        return F.max_pool2d(x, self.kernel_size, self.stride, self.padding,
                            self.dilation, self.ceil_mode, self.return_indices)


class MaxPool2dStaticSamePadding(nn.MaxPool2d):
    """2D MaxPooling like TensorFlow's 'SAME' mode, with the given input image size.
       The padding mudule is calculated in construction function, then used in forward.
    """

    def __init__(self, kernel_size, stride, image_size=None, **kwargs):
        super().__init__(kernel_size, stride, **kwargs)
        self.stride = [self.stride] * 2 if isinstance(self.stride, int) else self.stride
        self.kernel_size = [self.kernel_size] * 2 if isinstance(self.kernel_size, int) else self.kernel_size
        self.dilation = [self.dilation] * 2 if isinstance(self.dilation, int) else self.dilation

        # Calculate padding based on image size and save it
        assert image_size is not None
        ih, iw = (image_size, image_size) if isinstance(image_size, int) else image_size
        kh, kw = self.kernel_size
        sh, sw = self.stride
        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)
        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
        if pad_h > 0 or pad_w > 0:
            self.static_padding = nn.ZeroPad2d((pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2))
        else:
            self.static_padding = nn.Identity()

    def forward(self, x):
        x = self.static_padding(x)
        x = F.max_pool2d(x, self.kernel_size, self.stride, self.padding,
                         self.dilation, self.ceil_mode, self.return_indices)
        return x


################################################################################
# Helper functions for loading model params
################################################################################

# BlockDecoder: A Class for encoding and decoding BlockArgs
# efficientnet_params: A function to query compound coefficient
# get_model_params and efficientnet:
#     Functions to get BlockArgs and GlobalParams for efficientnet
# url_map and url_map_advprop: Dicts of url_map for pretrained weights
# load_pretrained_weights: A function to load pretrained weights

class BlockDecoder(object):
    """Block Decoder for readability,
       straight from the official TensorFlow repository.
    """

    @staticmethod
    def _decode_block_string(block_string):
        """Get a block through a string notation of arguments.
        Args:
            block_string (str): A string notation of arguments.
                                Examples: 'r1_k3_s11_e1_i32_o16_se0.25_noskip'.
        Returns:
            BlockArgs: The namedtuple defined at the top of this file.
        """
        assert isinstance(block_string, str)

        ops = block_string.split('_')
        options = {}
        for op in ops:
            splits = re.split(r'(\d.*)', op)
            if len(splits) >= 2:
                key, value = splits[:2]
                options[key] = value

        # Check stride
        assert (('s' in options and len(options['s']) == 1) or
                (len(options['s']) == 2 and options['s'][0] == options['s'][1]))

        return BlockArgs(
            num_repeat=int(options['r']),
            kernel_size=int(options['k']),
            stride=[int(options['s'][0])],
            expand_ratio=int(options['e']),
            input_filters=int(options['i']),
            output_filters=int(options['o']),
            se_ratio=float(options['se']) if 'se' in options else None,
            id_skip=('noskip' not in block_string))

    @staticmethod
    def _encode_block_string(block):
        """Encode a block to a string.
        Args:
            block (namedtuple): A BlockArgs type argument.
        Returns:
            block_string: A String form of BlockArgs.
        """
        args = [
            'r%d' % block.num_repeat,
            'k%d' % block.kernel_size,
            's%d%d' % (block.strides[0], block.strides[1]),
            'e%s' % block.expand_ratio,
            'i%d' % block.input_filters,
            'o%d' % block.output_filters
        ]
        if 0 < block.se_ratio <= 1:
            args.append('se%s' % block.se_ratio)
        if block.id_skip is False:
            args.append('noskip')
        return '_'.join(args)

    @staticmethod
    def decode(string_list):
        """Decode a list of string notations to specify blocks inside the network.
        Args:
            string_list (list[str]): A list of strings, each string is a notation of block.
        Returns:
            blocks_args: A list of BlockArgs namedtuples of block args.
        """
        assert isinstance(string_list, list)
        blocks_args = []
        for block_string in string_list:
            blocks_args.append(BlockDecoder._decode_block_string(block_string))
        return blocks_args

    @staticmethod
    def encode(blocks_args):
        """Encode a list of BlockArgs to a list of strings.
        Args:
            blocks_args (list[namedtuples]): A list of BlockArgs namedtuples of block args.
        Returns:
            block_strings: A list of strings, each string is a notation of block.
        """
        block_strings = []
        for block in blocks_args:
            block_strings.append(BlockDecoder._encode_block_string(block))
        return block_strings


def efficientnet_params(model_name):
    """Map EfficientNet model name to parameter coefficients.
    Args:
        model_name (str): Model name to be queried.
    Returns:
        params_dict[model_name]: A (width,depth,res,dropout) tuple.
    """
    params_dict = {
        # Coefficients:   width,depth,res,dropout
        'efficientnet-b0': (1.0, 1.0, 224, 0.2),
        'efficientnet-b1': (1.0, 1.1, 240, 0.2),
        'efficientnet-b2': (1.1, 1.2, 260, 0.3),
        'efficientnet-b3': (1.2, 1.4, 300, 0.3),
        'efficientnet-b4': (1.4, 1.8, 380, 0.4),
        'efficientnet-b5': (1.6, 2.2, 456, 0.4),
        'efficientnet-b6': (1.8, 2.6, 528, 0.5),
        'efficientnet-b7': (2.0, 3.1, 600, 0.5),
        'efficientnet-b8': (2.2, 3.6, 672, 0.5),
        'efficientnet-l2': (4.3, 5.3, 800, 0.5),
    }
    return params_dict[model_name]


def efficientnet(width_coefficient=None, depth_coefficient=None, image_size=None,
                 dropout_rate=0.2, drop_connect_rate=0.2, num_classes=1000, include_top=True):
    """Create BlockArgs and GlobalParams for efficientnet model.
    Args:
        width_coefficient (float)
        depth_coefficient (float)
        image_size (int)
        dropout_rate (float)
        drop_connect_rate (float)
        num_classes (int)
        Meaning as the name suggests.
    Returns:
        blocks_args, global_params.
    """

    # Blocks args for the whole model(efficientnet-b0 by default)
    # It will be modified in the construction of EfficientNet Class according to model
    blocks_args = [
        'r1_k3_s11_e1_i32_o16_se0.25',
        'r2_k3_s22_e6_i16_o24_se0.25',
        'r2_k5_s22_e6_i24_o40_se0.25',
        'r3_k3_s22_e6_i40_o80_se0.25',
        'r3_k5_s11_e6_i80_o112_se0.25',
        'r4_k5_s22_e6_i112_o192_se0.25',
        'r1_k3_s11_e6_i192_o320_se0.25',
    ]
    blocks_args = BlockDecoder.decode(blocks_args)

    global_params = GlobalParams(
        width_coefficient=width_coefficient,
        depth_coefficient=depth_coefficient,
        image_size=image_size,
        dropout_rate=dropout_rate,

        num_classes=num_classes,
        batch_norm_momentum=0.99,
        batch_norm_epsilon=1e-3,
        drop_connect_rate=drop_connect_rate,
        depth_divisor=8,
        min_depth=None,
        include_top=include_top,
    )

    return blocks_args, global_params


def get_model_params(model_name, override_params):
    """Get the block args and global params for a given model name.
    Args:
        model_name (str): Model's name.
        override_params (dict): A dict to modify global_params.
    Returns:
        blocks_args, global_params
    """
    if model_name.startswith('efficientnet'):
        w, d, s, p = efficientnet_params(model_name)
        # note: all models have drop connect rate = 0.2
        blocks_args, global_params = efficientnet(
            width_coefficient=w, depth_coefficient=d, dropout_rate=p, image_size=s)
    else:
        raise NotImplementedError('model name is not pre-defined: {}'.format(model_name))
    if override_params:
        # ValueError will be raised here if override_params has fields not included in global_params.
        global_params = global_params._replace(**override_params)
    return blocks_args, global_params


# train with Standard methods
# check more details in paper(EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks)
url_map = {
    'efficientnet-b0': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b0-355c32eb.pth',
    'efficientnet-b1': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b1-f1951068.pth',
    'efficientnet-b2': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b2-8bb594d6.pth',
    'efficientnet-b3': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b3-5fb5a3c3.pth',
    'efficientnet-b4': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b4-6ed6700e.pth',
    'efficientnet-b5': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b5-b6417697.pth',
    'efficientnet-b6': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b6-c76e70fd.pth',
    'efficientnet-b7': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b7-dcc49843.pth',
}

# train with Adversarial Examples(AdvProp)
# check more details in paper(Adversarial Examples Improve Image Recognition)
url_map_advprop = {
    'efficientnet-b0': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b0-b64d5a18.pth',
    'efficientnet-b1': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b1-0f3ce85a.pth',
    'efficientnet-b2': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b2-6e9d97e5.pth',
    'efficientnet-b3': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b3-cdd7c0f4.pth',
    'efficientnet-b4': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b4-44fb3a87.pth',
    'efficientnet-b5': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b5-86493f6b.pth',
    'efficientnet-b6': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b6-ac80338e.pth',
    'efficientnet-b7': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b7-4652b6dd.pth',
    'efficientnet-b8': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b8-22a8fe65.pth',
}

# TODO: add the petrained weights url map of 'efficientnet-l2'


def load_pretrained_weights(model, model_name, weights_path=None, load_fc=True, advprop=False, verbose=True):
    """Loads pretrained weights from weights path or download using url.
    Args:
        model (Module): The whole model of efficientnet.
        model_name (str): Model name of efficientnet.
        weights_path (None or str):
            str: path to pretrained weights file on the local disk.
            None: use pretrained weights downloaded from the Internet.
        load_fc (bool): Whether to load pretrained weights for fc layer at the end of the model.
        advprop (bool): Whether to load pretrained weights
                        trained with advprop (valid when weights_path is None).
    """
    if isinstance(weights_path, str):
        state_dict = torch.load(weights_path)
    else:
        # AutoAugment or Advprop (different preprocessing)
        url_map_ = url_map_advprop if advprop else url_map
        state_dict = model_zoo.load_url(url_map_[model_name])

    if load_fc:
        ret = model.load_state_dict(state_dict, strict=False)
        assert not ret.missing_keys, 'Missing keys when loading pretrained weights: {}'.format(ret.missing_keys)
    else:
        state_dict.pop('_fc.weight')
        state_dict.pop('_fc.bias')
        ret = model.load_state_dict(state_dict, strict=False)
        assert set(ret.missing_keys) == set(
            ['_fc.weight', '_fc.bias']), 'Missing keys when loading pretrained weights: {}'.format(ret.missing_keys)
    assert not ret.unexpected_keys, 'Missing keys when loading pretrained weights: {}'.format(ret.unexpected_keys)

    if verbose:
        print('Loaded pretrained weights for {}'.format(model_name))