Pytorch CIFAR10图像分类 EfficientNet v1篇

Pytorch CIFAR10图像分类 EfficientNet v1篇

再次介绍一下我的专栏，很适合大家初入深度学习或者是Pytorch和Keras，希望这能够帮助初学深度学习的同学一个入门Pytorch或者Keras的项目和在这之中更加了解Pytorch&Keras和各个图像分类的模型。

他有比较清晰的可视化结构和架构，除此之外，我是用jupyter写的，所以说在文章整体架构可以说是非常清晰，可以帮助你快速学习到各个模块的知识，而不是通过python脚本一行一行的看，这样的方式是符合初学者的。

除此之外，如果你需要变成脚本形式，也是很简单的。
这里贴一下汇总篇：汇总篇

4. 定义网络（EfficientNet）

EfficientNet介绍

EfficientNet源自Google Brain的论文EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. 从标题也可以看出，这篇论文最主要的创新点是Model Scaling. 论文提出了compound scaling，混合缩放，把网络缩放的三种方式：深度、宽度、分辨率，组合起来按照一定规则缩放，从而提高网络的效果。EfficientNet在网络变大时效果提升明显，把精度上限进一步提升，成为了当前最强网络。EfficientNet-B7在ImageNet上获得了最先进的 84.4%的top-1精度 和 97.1%的top-5精度，比之前最好的卷积网络（GPipe, Top-1: 84.3%, Top-5: 97.0%）大小缩小8.4倍、速度提升6.1倍。

在一般情况下，我们知道，增加网络参数可以获得更好的精度（有足够的数据，不过拟合的条件下），例如ResNet可以加深从ResNet-18到ResNet-200，GPipe将baseline模型放大四倍在ImageNet数据集上获得了84.3%的top-1精度。增加网络参数的方式有三种：深度、宽度和分辨率。

深度是指网络的层数，宽度指网络中卷积的channel数（例如wide resnet中通过增加channel数获得精度收益），分辨率是指通过网络输入大小（例如从112x112到224x224）。在EfficientNet之前，没有研究工作只是针对这三个维度中的某一个维度进行调整，因为没钱啊！！有限的计算能力，很少有研究对这三个维度进行综合调整的。

直观上来讲，这三种缩放方式并不不独立。对于分辨率高的图像，应该用更深的网络，因为需要更大的感受野，同时也应该增加网络宽度来获得更细粒度的特征。

EfficientNet性能比较

之前增加网络参数都是单独放大这三种方式中的一种，并没有同时调整，也没有调整方式的研究。EfficientNet使用了compound scaling方法，统一缩放网络深度、宽度和分辨率。类似于靠强大的搜索能力和计算能力，EfficientNet的主要创新点并不是结构，不像ResNet、SENet发明了shortcut或attention机制，EfficientNet的base结构是利用结构搜索搜出来的，然后使用compound scaling规则放缩，得到一系列表现优异的网络：B0~B7.下面两幅图分别是ImageNet的Top-1 Accuracy随参数量和flops变化关系图，可以看到EfficientNet饱和值高，并且到达速度快。可以这样说：对于ImageNet历史上的各种网络而言，可以说EfficientNet在效果上实现了碾压

EfficientNet的baseline

EfficientNet使用了MobileNet V2中的MBCConv作为模型的主干网络，同时也是用了SENet中的squeeze and excitation方法对网络结构进行了优化。MBConv在MobileNet V2中已经介绍过了，SENet会单独在之后的博文中进行详细讲解。

总之呢，综合了MBConv和squeeze and excitation方法的EfficientNet-B0的网络结构如下表所示：

这里找了一个更形象的图，方便我们理解和进行构建EfficientNet网络。

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-JmYLDSsa-1673061774717)(https://www.researchgate.net/publication/344410350/figure/fig4/AS:1022373302128641@1620764198841/Architecture-of-EfficientNet-B0-with-MBConv-as-Basic-building-blocks.png)]

EfficientNet模型混合缩放方法

EfficientNet的规范化混合调参方法使用了一个复合系数 $\varphi$ ，来对三个参数进行符合调整：使用了compound scaling方法，统一缩放网络深度、宽度和分辨率。

其中的$\alpha ,\beta ,\gamma$都是常数，可以通过网格搜索获得。复合系数通过人工调节。考虑到如果网络深度翻番那么对应的计算量翻番，网络宽度和图像分辨率翻番对应的计算量会翻4番，卷积操作的计算量与d ,$w^2,r^2$成正比，。在这个约束下，网络的计算量大约是之前的$2^{\varphi }$。

后续就可以利用混合缩放的方法进行对baseline网络进行改进，如下图所示，(a)为baseline网络，(b)、©、(d)为单独通过增加width，depth以及resolution使得网络变大的方式，(e)为compound scaling的方式。

这样就衍生出了EfficientB1-B7，从上面的图我们可以看到使用了compound scaling后，效果非常显著，在不同参数量和计算量都取得了多倍的提升。EfficientNet在ImageNet上的效果碾压，而且模型规模比此前的GPipe小了8.4倍。

其他版本的EfficientNet(B1-B7)

Model	input_size	width_coefficient	depth_coefficient	drop_connect_rate	dropout_rate
EfficientNetB0	224x224	1.0	1.0	0.2	0.2
EfficientNetB1	240x240	1.0	1.1	0.2	0.2
EfficientNetB2	260x260	1.1	1.2	0.2	0.3
EfficientNetB3	300x300	1.2	1.4	0.2	0.3
EfficientNetB4	380x380	1.4	1.8	0.2	0.4
EfficientNetB5	456x456	1.6	2.2	0.2	0.4
EfficientNetB6	528x528	1.8	2.6	0.2	0.5
EfficientNetB7	600x600	2.0	3.1	0.2	0.5

• input_size代表训练网络时输入网络的图像大小

• width_coefficient代表channel维度上的倍率因子，比如在 EfficientNetB0中Stage1的3x3卷积层所使用的卷积核个数是32，那么在B6中就是$32\times 1.8=57.6$接着取整到离它最近的8的整数倍即56，其它Stage同理。

• depth_coefficient代表depth维度上的倍率因子（仅针对Stage2到Stage8），比如在EfficientNetB0中Stage7的${\hat{L}}_i=4$,那么在B6中就是$4\times 2.6=10.4$接着向上取整即11。

• drop_connect_rate是在MBConv结构中dropout层使用的drop_rate，在官方keras模块的实现中MBConv结构的drop_rate是从0递增到drop_connect_rate的（注意，在源码实现中只有使用shortcut的时候才有Dropout层），还需要注意的是，这里的Dropout层是Stochastic Depth，即会随机丢掉整个block的主分支（只剩捷径分支，相当于直接跳过了这个block）也可以理解为减少了网络的深度。

• dropout_rate是最后一个全连接层前的dropout层（在stage9的Pooling与FC之间）的dropout_rate。

  最后给出原论文中关于EfficientNet与当时主流网络的性能参数对比：

  

判断是否使用GPU

首先我们还是得判断是否可以利用GPU，因为GPU的速度可能会比我们用CPU的速度快20-100倍左右，特别是对卷积神经网络来说，更是提升特别明显。

device = 'cuda' if torch.cuda.is_available() else 'cpu'

我们首先定义一下激活函数Swish和DropConnect方法，DropConnect是一种正则化方法，它在训练过程中随机将网络中的某些权重设置为0，在有些实现中，大家会使用DropPath进行，也是可以的。

# 激活函数
def swish(x):
    return x * x.sigmoid()

# DropConnect是一种正则化方法，它在训练过程中随机将网络中的某些权重设置为0
# DropConnect是对网络中每一个权重进行随机设置的
def drop_connect(x, drop_ratio):
    keep_ratio = 1.0 - drop_ratio
    mask = torch.empty([x.shape[0], 1, 1, 1], dtype=x.dtype, device=x.device)
    mask.bernoulli_(keep_ratio) 
    x.div_(keep_ratio)
    x.mul_(mask)
    return x

SE模块

有时候对于SE模块来说，我们会用1x1的卷积层实现，因为在这里，实际上1x1的卷积层是等价于全连接层，也就是Linear层的

SE模块如下所示，由一个全局平均池化，两个全连接层组成。第一个全连接层的节点个数是输入该MBConv特征矩阵channels的$ \frac{1}{4} $，且使用Swish激活函数。第二个全连接层的节点个数等于Depthwise Conv层输出的特征矩阵channels，且使用Sigmoid激活函数。

# SE模块实现
class SE(nn.Module):
    '''Squeeze-and-Excitation MBconv with Swish.'''

    def __init__(self, in_channels, se_channels):
        super(SE, self).__init__()
        # 这里的1x1的卷积核与全连接层是等价的
        self.se1 = nn.Conv2d(in_channels, se_channels,
                             kernel_size=1, bias=True) 
        self.se2 = nn.Conv2d(se_channels, in_channels,
                             kernel_size=1, bias=True)

    def forward(self, x):
        out = F.adaptive_avg_pool2d(x, (1, 1)) # 平均池化
        out = swish(self.se1(out))
        out = self.se2(out).sigmoid()
        out = x * out # 进行连接相乘
        return out

MBConv 结构

如图所示，MBConv 结构主要由一个 1x1 的普通卷积（升维作用），一个 kxk 的 Depthwise Conv 卷积。k 的具体值主要有 3x3 和 5x5 两种情况，一个 SE 模块，一个 1x1 的普通卷积（降维作用），一个 Droupout 层构成

• 第一个升维的 1x1 卷积层，它的卷积核个数是输入特征矩阵 channel 的 n 倍，n ∈ { 1 , 6 }

• 当 n = 1 时，不要第一个升维的 1x1 卷积层，即 Stage2 中的 MBConv 结构都没有第一个升维的 1x1 卷积层（这和MobileNetV3网络类似）。

• 仅当输入 MBConv 结构的特征矩阵与输出的特征矩阵 shape 相同时才存在 shortcut 连接（代码中可通过stride==1 and inputc_channels==output_channels条件来判断）。

• 在源码实现中只有使用 shortcut 的时候才有 Dropout 层

# MBconv模块
class MBconv(nn.Module):
    '''expansion + depthwise + pointwise + squeeze-excitation'''

    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size,
                 stride,
                 expand_ratio=1,
                 se_ratio=0.25, # 也就是1/4
                 drop_rate=0.):
        super(MBconv, self).__init__()
        self.stride = stride
        self.drop_rate = drop_rate
        self.expand_ratio = expand_ratio

        # Expansion 第一个升维的 1x1 卷积层
        channels = expand_ratio * in_channels
        self.conv1 = nn.Conv2d(in_channels,
                               channels,
                               kernel_size=1,
                               stride=1,
                               padding=0,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(channels)

        # Depthwise conv
        self.conv2 = nn.Conv2d(channels,
                               channels,
                               kernel_size=kernel_size,
                               stride=stride,
                               padding=(1 if kernel_size == 3 else 2),
                               groups=channels,
                               bias=False)
        self.bn2 = nn.BatchNorm2d(channels)

        # SE layers SE模块
        se_channels = int(in_channels * se_ratio)
        self.se = SE(channels, se_channels)

        # Output
        self.conv3 = nn.Conv2d(channels,
                               out_channels,
                               kernel_size=1,
                               stride=1,
                               padding=0,
                               bias=False)
        self.bn3 = nn.BatchNorm2d(out_channels)

        # Skip connection if in and out shapes are the same (MV-V2 style)
        # 仅当输入 MBConv 结构的特征矩阵与输出的特征矩阵 shape 相同时才存在 shortcut 连接
        self.has_skip = (stride == 1) and (in_channels == out_channels)

    def forward(self, x):
        # 当n = 1的时候，是不要第一个升维的1x1的卷积层
        out = x if self.expand_ratio == 1 else swish(self.bn1(self.conv1(x)))
        out = swish(self.bn2(self.conv2(out)))
        out = self.se(out)
        out = self.bn3(self.conv3(out))
        if self.has_skip:
            if self.training and self.drop_rate > 0: # 有shortcut的时候用Dropout
                out = drop_connect(out, self.drop_rate)
            out = out + x
        return out

定义EfficientNet的网络

这样我们就可以根据EFficientNet的论文的架构定义我们的网络，用到上述的SE模块和MBconv模块进行

# 定义EfficientNet方法
class EfficientNet(nn.Module):
    # 默认识别的类别为10
    def __init__(self,
                 width_coefficient=1.0,
                 depth_coefficient=1.0,
                 dropout_rate = 0.2,
                 num_classes=10):
        super(EfficientNet, self).__init__()
        # kernel_size, in_channel, out_channel, exp_ratio, strides, use_SE, drop_connect_rate, repeats
        cfg = {
    
    
            'num_blocks': [1, 2, 2, 3, 3, 4, 1], # repeats
            'expansion': [1, 6, 6, 6, 6, 6, 6], # expansion
            'out_channels': [16, 24, 40, 80, 112, 192, 320], # out_channels
            'kernel_size': [3, 3, 5, 3, 5, 5, 3], # kernel_size
            'stride': [1, 2, 2, 2, 1, 2, 1], # stride
            'dropout_rate': dropout_rate,
            'drop_connect_rate': 0.2,
        }
        # 配置文件
        self.cfg = cfg
        self.width_coefficient = width_coefficient
        self.depth_coefficient = depth_coefficient
        # 第一个3x3的卷积层
        self.conv1 = nn.Conv2d(3,
                               32,
                               kernel_size=3,
                               stride=1,
                               padding=1,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(32)
        self.layers = self._make_layers(in_channels=32)
        self.linear = nn.Linear(cfg['out_channels'][-1]*int(self.width_coefficient), num_classes)

    def _make_layers(self, in_channels):
        layers = []
        cfg = [self.cfg[k] for k in ['expansion', 'out_channels', 'num_blocks', 'kernel_size', 'stride']]
        b = 0
        blocks = sum(self.cfg['num_blocks'])
        for expansion, out_channels, num_blocks, kernel_size, stride in zip(*cfg):
            import math
            num_blocks = math.floor(self.depth_coefficient) * num_blocks
            strides = [stride] + [1] * (num_blocks - 1)
            for stride in strides:
                drop_rate = self.cfg['drop_connect_rate'] * b / blocks
                layers.append(
                    MBconv(in_channels,
                          out_channels*int(self.width_coefficient),
                          kernel_size,
                          stride,
                          expansion,
                          se_ratio=0.25,
                          drop_rate=drop_rate))
                in_channels = out_channels*int(self.width_coefficient)
        return nn.Sequential(*layers)

    def forward(self, x):
        out = swish(self.bn1(self.conv1(x)))
        out = self.layers(out)
        out = F.adaptive_avg_pool2d(out, 1)
        out = out.view(out.size(0), -1)
        dropout_rate = self.cfg['dropout_rate']
        if self.training and dropout_rate > 0:
            out = F.dropout(out, p=dropout_rate)
        out = self.linear(out)
        return out

定义EfficientNetB0~B7

在上面我们已经定义了两个参数，我们可以根据参数的配置，进行设置，这样就可以定义出我们的EfficientNetB0~B7的模型

def EfficientNetB0(num_classes = 10):
    # input image size 224x224
    return EfficientNet(width_coefficient=1.0,
                        depth_coefficient=1.0,
                        num_classes=num_classes)

def EfficientNetB1(num_classes = 10):
    # input image size 240x240
    return EfficientNet(width_coefficient=1.0,
                        depth_coefficient=1.1,
                        num_classes=num_classes)

def EfficientNetB2(num_classes = 10):
    # input image size 260x260
    return EfficientNet(width_coefficient=1.1,
                        depth_coefficient=1.2,
                        dropout_rate=0.3,
                        num_classes=num_classes)

def EfficientNetB3(num_classes = 10):
    # input image size 300x300
    return EfficientNet(width_coefficient=1.2,
                        depth_coefficient=1.4,
                        dropout_rate=0.3,
                        num_classes=num_classes)
    
def EfficientNetB4(num_classes = 10):
    # input image size 380x380
    return EfficientNet(width_coefficient=1.4,
                        depth_coefficient=1.8,
                        dropout_rate=0.4,
                        num_classes=num_classes)
    
def EfficientNetB5(num_classes = 10):
    # input image size 456x456
    return EfficientNet(width_coefficient=1.6,
                        depth_coefficient=2.2,
                        dropout_rate=0.4,
                        num_classes=num_classes)

def EfficientNetB6(num_classes = 10):
    # input image size 528x528
    return EfficientNet(width_coefficient=1.8,
                        depth_coefficient=2.6,
                        dropout_rate=0.5,
                        num_classes=num_classes)

def EfficientNetB7(num_classes = 10):
    # input image size 600x600
    return EfficientNet(width_coefficient=2.0,
                        depth_coefficient=3.1,
                        dropout_rate=0.5,
                        num_classes=num_classes)

net = EfficientNetB0(num_classes = 10).to(device)

summary查看网络

我们可以通过summary来看到，模型的维度的变化，经过层后shape的变化，是否最后也是输出(batch,shape)

summary(net,(2,3,32,32))

==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
EfficientNet                             [2, 10]                   --
├─Conv2d: 1-1                            [2, 32, 32, 32]           864
├─BatchNorm2d: 1-2                       [2, 32, 32, 32]           64
├─Sequential: 1-3                        [2, 320, 2, 2]            --
│    └─MBconv: 2-1                       [2, 16, 32, 32]           1,088
│    │    └─Conv2d: 3-1                  [2, 32, 32, 32]           288
│    │    └─BatchNorm2d: 3-2             [2, 32, 32, 32]           64
│    │    └─SE: 3-3                      [2, 32, 32, 32]           552
│    │    └─Conv2d: 3-4                  [2, 16, 32, 32]           512
│    │    └─BatchNorm2d: 3-5             [2, 16, 32, 32]           32
│    └─MBconv: 2-2                       [2, 24, 16, 16]           --
│    │    └─Conv2d: 3-6                  [2, 96, 32, 32]           1,536
│    │    └─BatchNorm2d: 3-7             [2, 96, 32, 32]           192
│    │    └─Conv2d: 3-8                  [2, 96, 16, 16]           864
│    │    └─BatchNorm2d: 3-9             [2, 96, 16, 16]           192
│    │    └─SE: 3-10                     [2, 96, 16, 16]           868
│    │    └─Conv2d: 3-11                 [2, 24, 16, 16]           2,304
│    │    └─BatchNorm2d: 3-12            [2, 24, 16, 16]           48
│    └─MBconv: 2-3                       [2, 24, 16, 16]           --
│    │    └─Conv2d: 3-13                 [2, 144, 16, 16]          3,456
│    │    └─BatchNorm2d: 3-14            [2, 144, 16, 16]          288
│    │    └─Conv2d: 3-15                 [2, 144, 16, 16]          1,296
│    │    └─BatchNorm2d: 3-16            [2, 144, 16, 16]          288
│    │    └─SE: 3-17                     [2, 144, 16, 16]          1,878
│    │    └─Conv2d: 3-18                 [2, 24, 16, 16]           3,456
│    │    └─BatchNorm2d: 3-19            [2, 24, 16, 16]           48
│    └─MBconv: 2-4                       [2, 40, 8, 8]             --
│    │    └─Conv2d: 3-20                 [2, 144, 16, 16]          3,456
│    │    └─BatchNorm2d: 3-21            [2, 144, 16, 16]          288
│    │    └─Conv2d: 3-22                 [2, 144, 8, 8]            3,600
│    │    └─BatchNorm2d: 3-23            [2, 144, 8, 8]            288
│    │    └─SE: 3-24                     [2, 144, 8, 8]            1,878
│    │    └─Conv2d: 3-25                 [2, 40, 8, 8]             5,760
│    │    └─BatchNorm2d: 3-26            [2, 40, 8, 8]             80
│    └─MBconv: 2-5                       [2, 40, 8, 8]             --
│    │    └─Conv2d: 3-27                 [2, 240, 8, 8]            9,600
│    │    └─BatchNorm2d: 3-28            [2, 240, 8, 8]            480
│    │    └─Conv2d: 3-29                 [2, 240, 8, 8]            6,000
│    │    └─BatchNorm2d: 3-30            [2, 240, 8, 8]            480
│    │    └─SE: 3-31                     [2, 240, 8, 8]            5,050
│    │    └─Conv2d: 3-32                 [2, 40, 8, 8]             9,600
│    │    └─BatchNorm2d: 3-33            [2, 40, 8, 8]             80
│    └─MBconv: 2-6                       [2, 80, 4, 4]             --
│    │    └─Conv2d: 3-34                 [2, 240, 8, 8]            9,600
│    │    └─BatchNorm2d: 3-35            [2, 240, 8, 8]            480
│    │    └─Conv2d: 3-36                 [2, 240, 4, 4]            2,160
│    │    └─BatchNorm2d: 3-37            [2, 240, 4, 4]            480
│    │    └─SE: 3-38                     [2, 240, 4, 4]            5,050
│    │    └─Conv2d: 3-39                 [2, 80, 4, 4]             19,200
│    │    └─BatchNorm2d: 3-40            [2, 80, 4, 4]             160
│    └─MBconv: 2-7                       [2, 80, 4, 4]             --
│    │    └─Conv2d: 3-41                 [2, 480, 4, 4]            38,400
│    │    └─BatchNorm2d: 3-42            [2, 480, 4, 4]            960
│    │    └─Conv2d: 3-43                 [2, 480, 4, 4]            4,320
│    │    └─BatchNorm2d: 3-44            [2, 480, 4, 4]            960
│    │    └─SE: 3-45                     [2, 480, 4, 4]            19,700
│    │    └─Conv2d: 3-46                 [2, 80, 4, 4]             38,400
│    │    └─BatchNorm2d: 3-47            [2, 80, 4, 4]             160
│    └─MBconv: 2-8                       [2, 80, 4, 4]             --
│    │    └─Conv2d: 3-48                 [2, 480, 4, 4]            38,400
│    │    └─BatchNorm2d: 3-49            [2, 480, 4, 4]            960
│    │    └─Conv2d: 3-50                 [2, 480, 4, 4]            4,320
│    │    └─BatchNorm2d: 3-51            [2, 480, 4, 4]            960
│    │    └─SE: 3-52                     [2, 480, 4, 4]            19,700
│    │    └─Conv2d: 3-53                 [2, 80, 4, 4]             38,400
│    │    └─BatchNorm2d: 3-54            [2, 80, 4, 4]             160
│    └─MBconv: 2-9                       [2, 112, 4, 4]            --
│    │    └─Conv2d: 3-55                 [2, 480, 4, 4]            38,400
│    │    └─BatchNorm2d: 3-56            [2, 480, 4, 4]            960
│    │    └─Conv2d: 3-57                 [2, 480, 4, 4]            12,000
│    │    └─BatchNorm2d: 3-58            [2, 480, 4, 4]            960
│    │    └─SE: 3-59                     [2, 480, 4, 4]            19,700
│    │    └─Conv2d: 3-60                 [2, 112, 4, 4]            53,760
│    │    └─BatchNorm2d: 3-61            [2, 112, 4, 4]            224
│    └─MBconv: 2-10                      [2, 112, 4, 4]            --
│    │    └─Conv2d: 3-62                 [2, 672, 4, 4]            75,264
│    │    └─BatchNorm2d: 3-63            [2, 672, 4, 4]            1,344
│    │    └─Conv2d: 3-64                 [2, 672, 4, 4]            16,800
│    │    └─BatchNorm2d: 3-65            [2, 672, 4, 4]            1,344
│    │    └─SE: 3-66                     [2, 672, 4, 4]            38,332
│    │    └─Conv2d: 3-67                 [2, 112, 4, 4]            75,264
│    │    └─BatchNorm2d: 3-68            [2, 112, 4, 4]            224
│    └─MBconv: 2-11                      [2, 112, 4, 4]            --
│    │    └─Conv2d: 3-69                 [2, 672, 4, 4]            75,264
│    │    └─BatchNorm2d: 3-70            [2, 672, 4, 4]            1,344
│    │    └─Conv2d: 3-71                 [2, 672, 4, 4]            16,800
│    │    └─BatchNorm2d: 3-72            [2, 672, 4, 4]            1,344
│    │    └─SE: 3-73                     [2, 672, 4, 4]            38,332
│    │    └─Conv2d: 3-74                 [2, 112, 4, 4]            75,264
│    │    └─BatchNorm2d: 3-75            [2, 112, 4, 4]            224
│    └─MBconv: 2-12                      [2, 192, 2, 2]            --
│    │    └─Conv2d: 3-76                 [2, 672, 4, 4]            75,264
│    │    └─BatchNorm2d: 3-77            [2, 672, 4, 4]            1,344
│    │    └─Conv2d: 3-78                 [2, 672, 2, 2]            16,800
│    │    └─BatchNorm2d: 3-79            [2, 672, 2, 2]            1,344
│    │    └─SE: 3-80                     [2, 672, 2, 2]            38,332
│    │    └─Conv2d: 3-81                 [2, 192, 2, 2]            129,024
│    │    └─BatchNorm2d: 3-82            [2, 192, 2, 2]            384
│    └─MBconv: 2-13                      [2, 192, 2, 2]            --
│    │    └─Conv2d: 3-83                 [2, 1152, 2, 2]           221,184
│    │    └─BatchNorm2d: 3-84            [2, 1152, 2, 2]           2,304
│    │    └─Conv2d: 3-85                 [2, 1152, 2, 2]           28,800
│    │    └─BatchNorm2d: 3-86            [2, 1152, 2, 2]           2,304
│    │    └─SE: 3-87                     [2, 1152, 2, 2]           111,792
│    │    └─Conv2d: 3-88                 [2, 192, 2, 2]            221,184
│    │    └─BatchNorm2d: 3-89            [2, 192, 2, 2]            384
│    └─MBconv: 2-14                      [2, 192, 2, 2]            --
│    │    └─Conv2d: 3-90                 [2, 1152, 2, 2]           221,184
│    │    └─BatchNorm2d: 3-91            [2, 1152, 2, 2]           2,304
│    │    └─Conv2d: 3-92                 [2, 1152, 2, 2]           28,800
│    │    └─BatchNorm2d: 3-93            [2, 1152, 2, 2]           2,304
│    │    └─SE: 3-94                     [2, 1152, 2, 2]           111,792
│    │    └─Conv2d: 3-95                 [2, 192, 2, 2]            221,184
│    │    └─BatchNorm2d: 3-96            [2, 192, 2, 2]            384
│    └─MBconv: 2-15                      [2, 192, 2, 2]            --
│    │    └─Conv2d: 3-97                 [2, 1152, 2, 2]           221,184
│    │    └─BatchNorm2d: 3-98            [2, 1152, 2, 2]           2,304
│    │    └─Conv2d: 3-99                 [2, 1152, 2, 2]           28,800
│    │    └─BatchNorm2d: 3-100           [2, 1152, 2, 2]           2,304
│    │    └─SE: 3-101                    [2, 1152, 2, 2]           111,792
│    │    └─Conv2d: 3-102                [2, 192, 2, 2]            221,184
│    │    └─BatchNorm2d: 3-103           [2, 192, 2, 2]            384
│    └─MBconv: 2-16                      [2, 320, 2, 2]            --
│    │    └─Conv2d: 3-104                [2, 1152, 2, 2]           221,184
│    │    └─BatchNorm2d: 3-105           [2, 1152, 2, 2]           2,304
│    │    └─Conv2d: 3-106                [2, 1152, 2, 2]           10,368
│    │    └─BatchNorm2d: 3-107           [2, 1152, 2, 2]           2,304
│    │    └─SE: 3-108                    [2, 1152, 2, 2]           111,792
│    │    └─Conv2d: 3-109                [2, 320, 2, 2]            368,640
│    │    └─BatchNorm2d: 3-110           [2, 320, 2, 2]            640
├─Linear: 1-4                            [2, 10]                   3,210
==========================================================================================
Total params: 3,599,686
Trainable params: 3,599,686
Non-trainable params: 0
Total mult-adds (M): 60.76
==========================================================================================
Input size (MB): 0.02
Forward/backward pass size (MB): 17.59
Params size (MB): 14.39
Estimated Total Size (MB): 32.01
==========================================================================================

测试和定义网络

我们也可以打印出我们的模型观察一下

print(net)

EfficientNet(
  (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
  (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (layers): Sequential(
    (0): MBconv(
      (conv1): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False)
      (bn2): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(32, 8, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(8, 32, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(32, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (1): MBconv(
      (conv1): Conv2d(16, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(96, 96, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=96, bias=False)
      (bn2): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(96, 4, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(4, 96, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(96, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (2): MBconv(
      (conv1): Conv2d(24, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(144, 144, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=144, bias=False)
      (bn2): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(144, 6, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(6, 144, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(144, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (3): MBconv(
      (conv1): Conv2d(24, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(144, 144, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=144, bias=False)
      (bn2): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(144, 6, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(6, 144, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(144, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (4): MBconv(
      (conv1): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(240, 240, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=240, bias=False)
      (bn2): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(240, 10, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(10, 240, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(240, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (5): MBconv(
      (conv1): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(240, 240, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=240, bias=False)
      (bn2): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(240, 10, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(10, 240, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(240, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (6): MBconv(
      (conv1): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(480, 480, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=480, bias=False)
      (bn2): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(480, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (7): MBconv(
      (conv1): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(480, 480, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=480, bias=False)
      (bn2): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(480, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (8): MBconv(
      (conv1): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(480, 480, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=480, bias=False)
      (bn2): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(480, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (9): MBconv(
      (conv1): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(672, 672, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=672, bias=False)
      (bn2): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(672, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (10): MBconv(
      (conv1): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(672, 672, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=672, bias=False)
      (bn2): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(672, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (11): MBconv(
      (conv1): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(672, 672, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=672, bias=False)
      (bn2): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(672, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (12): MBconv(
      (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(1152, 1152, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=1152, bias=False)
      (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (13): MBconv(
      (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(1152, 1152, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=1152, bias=False)
      (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (14): MBconv(
      (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(1152, 1152, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=1152, bias=False)
      (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
    (15): MBconv(
      (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1152, bias=False)
      (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (se): SE(
        (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
        (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
      )
      (conv3): Conv2d(1152, 320, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(320, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    )
  )
  (linear): Linear(in_features=320, out_features=10, bias=True)
)

接下来可以简单测试一下，是否输入后能得到我们的正确的维度shape

test_x = torch.randn(2,3,32,32).to(device)
test_y = net(test_x)
print(test_y.shape)

torch.Size([2, 10])

定义网络和设置类别

net = EfficientNetB0(num_classes=10)

5. 定义损失函数和优化器

pytorch将深度学习中常用的优化方法全部封装在torch.optim之中，所有的优化方法都是继承基类optim.Optimizier

损失函数是封装在神经网络工具箱nn中的,包含很多损失函数

这里我使用的是SGD + momentum算法，并且我们损失函数定义为交叉熵函数，除此之外学习策略定义为动态更新学习率，如果5次迭代后，训练的损失并没有下降，那么我们便会更改学习率，会变为原来的0.5倍，最小降低到0.00001

如果想更加了解优化器和学习率策略的话，可以参考以下资料

• Pytorch Note15 优化算法1 梯度下降（Gradient descent varients）
• Pytorch Note16 优化算法2 动量法(Momentum)
• Pytorch Note34 学习率衰减

这里决定迭代10次

import torch.optim as optim
optimizer = optim.SGD(net.parameters(), lr=1e-3, momentum=0.9, weight_decay=5e-4)
optimizer = optim.AdamW(net.parameters(), lr=1e-2)
criterion = nn.CrossEntropyLoss()
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', factor=0.94 ,patience = 1,min_lr = 0.000001) # 动态更新学习率
# scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[75, 150], gamma=0.5)
import time
epoch = 10

6. 训练及可视化（增加TensorBoard可视化）

首先定义模型保存的位置

import os
if not os.path.exists('./model'):
    os.makedirs('./model')
else:
    print('文件已存在')
save_path = './model/EfficientNet-B0.pth'

文件已存在

这次更新了tensorboard的可视化，可以得到更好看的图片，并且能可视化出不错的结果

# 使用tensorboard
from torch.utils.tensorboard import SummaryWriter
os.makedirs("./logs", exist_ok=True)
tbwriter = SummaryWriter(log_dir='./logs/EfficientNet-B0', comment='EfficientNet-B0')  # 使用tensorboard记录中间输出
tbwriter.add_graph(model= net, input_to_model=torch.randn(size=(2, 3, 32, 32)))

如果存在GPU可以选择使用GPU进行运行，并且可以设置并行运算

if device == 'cuda':
    net.to(device)
    net = nn.DataParallel(net) # 使用并行运算

开始训练

我定义了一个train函数，在train函数中进行一个训练，并保存我们训练后的模型，这一部分一定要注意，这里的utils文件是我个人写的，所以需要下载下来

或者可以参考我们的工具函数篇，我还更新了结果和方法。

from utils import plot_history
from utils import train
Acc, Loss, Lr = train(net, trainloader, testloader, epoch, optimizer, criterion, scheduler, save_path, tbwriter, verbose = True)

Train Epoch 1/10: 100%|██████████| 391/391 [01:14<00:00,  5.28it/s, Train Acc=0.377, Train Loss=1.67]
Test Epoch 1/10: 100%|██████████| 79/79 [00:03<00:00, 25.05it/s, Test Acc=0.458, Test Loss=1.65]


Epoch [  1/ 10]  Train Loss:1.670216  Train Acc:37.75% Test Loss:1.651851  Test Acc:45.83%  Learning Rate:0.010000


Train Epoch 2/10: 100%|██████████| 391/391 [01:12<00:00,  5.38it/s, Train Acc=0.601, Train Loss=1.12]
Test Epoch 2/10: 100%|██████████| 79/79 [00:03<00:00, 19.84it/s, Test Acc=0.634, Test Loss=1.05]


Epoch [  2/ 10]  Train Loss:1.115703  Train Acc:60.13% Test Loss:1.049351  Test Acc:63.43%  Learning Rate:0.010000


Train Epoch 3/10: 100%|██████████| 391/391 [01:12<00:00,  5.40it/s, Train Acc=0.688, Train Loss=0.888]
Test Epoch 3/10: 100%|██████████| 79/79 [00:03<00:00, 24.33it/s, Test Acc=0.703, Test Loss=0.856]


Epoch [  3/ 10]  Train Loss:0.888298  Train Acc:68.83% Test Loss:0.856442  Test Acc:70.28%  Learning Rate:0.010000


Train Epoch 4/10: 100%|██████████| 391/391 [01:11<00:00,  5.45it/s, Train Acc=0.735, Train Loss=0.763]
Test Epoch 4/10: 100%|██████████| 79/79 [00:03<00:00, 25.79it/s, Test Acc=0.743, Test Loss=0.738]


Epoch [  4/ 10]  Train Loss:0.762550  Train Acc:73.53% Test Loss:0.738470  Test Acc:74.34%  Learning Rate:0.010000


Train Epoch 5/10: 100%|██████████| 391/391 [01:12<00:00,  5.36it/s, Train Acc=0.771, Train Loss=0.667]
Test Epoch 5/10: 100%|██████████| 79/79 [00:03<00:00, 22.48it/s, Test Acc=0.75, Test Loss=0.713] 


Epoch [  5/ 10]  Train Loss:0.666550  Train Acc:77.13% Test Loss:0.712513  Test Acc:75.05%  Learning Rate:0.010000


Train Epoch 6/10: 100%|██████████| 391/391 [01:10<00:00,  5.54it/s, Train Acc=0.794, Train Loss=0.602]
Test Epoch 6/10: 100%|██████████| 79/79 [00:03<00:00, 24.28it/s, Test Acc=0.772, Test Loss=0.693]


Epoch [  6/ 10]  Train Loss:0.601911  Train Acc:79.35% Test Loss:0.693019  Test Acc:77.20%  Learning Rate:0.010000


Train Epoch 7/10: 100%|██████████| 391/391 [01:12<00:00,  5.42it/s, Train Acc=0.809, Train Loss=0.56] 
Test Epoch 7/10: 100%|██████████| 79/79 [00:03<00:00, 25.42it/s, Test Acc=0.775, Test Loss=0.655]


Epoch [  7/ 10]  Train Loss:0.559702  Train Acc:80.86% Test Loss:0.654783  Test Acc:77.52%  Learning Rate:0.010000


Train Epoch 8/10: 100%|██████████| 391/391 [01:11<00:00,  5.48it/s, Train Acc=0.825, Train Loss=0.511]
Test Epoch 8/10: 100%|██████████| 79/79 [00:03<00:00, 23.75it/s, Test Acc=0.795, Test Loss=0.614]


Epoch [  8/ 10]  Train Loss:0.510548  Train Acc:82.48% Test Loss:0.614094  Test Acc:79.46%  Learning Rate:0.010000


Train Epoch 9/10: 100%|██████████| 391/391 [01:11<00:00,  5.44it/s, Train Acc=0.834, Train Loss=0.482]
Test Epoch 9/10: 100%|██████████| 79/79 [00:03<00:00, 25.49it/s, Test Acc=0.789, Test Loss=0.634]


Epoch [  9/ 10]  Train Loss:0.482006  Train Acc:83.40% Test Loss:0.634270  Test Acc:78.88%  Learning Rate:0.010000


Train Epoch 10/10: 100%|██████████| 391/391 [01:11<00:00,  5.46it/s, Train Acc=0.844, Train Loss=0.456]
Test Epoch 10/10: 100%|██████████| 79/79 [00:04<00:00, 19.49it/s, Test Acc=0.76, Test Loss=0.742] 

Epoch [ 10/ 10]  Train Loss:0.455699  Train Acc:84.36% Test Loss:0.742326  Test Acc:76.05%  Learning Rate:0.010000

​
​

训练曲线可视化

plot_history(epoch ,Acc, Loss, Lr)

可以运行以下代码进行tensorboard可视化

tensorboard --logdir logs

7. 测试

查看准确率

correct = 0   # 定义预测正确的图片数，初始化为0
total = 0     # 总共参与测试的图片数，也初始化为0

for data in testloader:  # 循环每一个batch
    images, labels = data
    images = images.to(device)
    labels = labels.to(device)
    net.eval()  # 把模型转为test模式
    if hasattr(torch.cuda, 'empty_cache'):
        torch.cuda.empty_cache()
    outputs = net(images)  # 输入网络进行测试
    
    # outputs.data是一个4x10张量，将每一行的最大的那一列的值和序号各自组成一个一维张量返回，第一个是值的张量，第二个是序号的张量。
    _, predicted = torch.max(outputs.data, 1)
    total += labels.size(0)          # 更新测试图片的数量
    correct += (predicted == labels).sum() # 更新正确分类的图片的数量

print('Accuracy of the network on the 10000 test images: %.2f %%' % (100 * correct / total))
 

Accuracy of the network on the 10000 test images: 76.13 %

程序中的 torch.max(outputs.data, 1) ，返回一个tuple (元组)

而这里很明显，这个返回的元组的第一个元素是image data，即是最大的 值，第二个元素是label， 即是最大的值 的 索引！我们只需要label（最大值的索引），所以就会有_,predicted这样的赋值语句，表示忽略第一个返回值，把它赋值给 _， 就是舍弃它的意思；

查看每一类的准确率

 # 定义2个存储每类中测试正确的个数的 列表，初始化为0
class_correct = list(0. for i in range(10))
class_total = list(0. for i in range(10))

net.eval()
with torch.no_grad():
    for data in testloader:
        images, labels = data
        images = images.to(device)
        labels = labels.to(device)
        if hasattr(torch.cuda, 'empty_cache'):
            torch.cuda.empty_cache()
        outputs = net(images)

        _, predicted = torch.max(outputs.data, 1)
        #(batch_size)数据中，输出于label相同的，标记为1，否则为0
        c = (predicted == labels).squeeze()
        for i in range(len(images)):      # 因为每个batch都有len(iamges)张图片，所以还需要一个len(iamges)的小循环
            label = labels[i]   # 对各个类的进行各自累加
            class_correct[label] += c[i]
            class_total[label] += 1
 
 
for i in range(10):
    print('Accuracy of %5s : %.2f %%' % (classes[i], 100 * class_correct[i] / class_total[i]))

Accuracy of airplane : 80.40 %
Accuracy of automobile : 92.20 %
Accuracy of  bird : 75.70 %
Accuracy of   cat : 52.30 %
Accuracy of  deer : 50.00 %
Accuracy of   dog : 75.80 %
Accuracy of  frog : 64.40 %
Accuracy of horse : 92.60 %
Accuracy of  ship : 89.10 %
Accuracy of truck : 88.80 %

抽样测试并可视化一部分结果

dataiter = iter(testloader)
images, labels = dataiter.next()

images_,labels = images.to(device),labels.to(device)
val_output = net(images_)

_, val_preds = torch.max(val_output, 1)
correct = torch.sum(val_preds == labels.data).item()
val_preds,labels = val_preds.cpu(),labels.cpu()

print("Accuracy Rate = {}%".format(correct/len(images) * 100))

fig = plt.figure(figsize=(25,25))
for idx in np.arange(64):    
    ax = fig.add_subplot(8, 8, idx+1, xticks=[], yticks=[])
    imshow(images[idx])
    ax.set_title("{}, ({})".format(classes[val_preds[idx].item()], classes[labels[idx].item()]), 
                 color = ("green" if val_preds[idx].item()==labels[idx].item() else "red"))

Accuracy Rate = 78.125%

8. 保存模型

这里保存成前面设定的模型的名字

torch.save(net,save_path[:-4]+'_'+str(epoch)+'.pth')

9. 预测

import torch
from PIL import Image
from torch.autograd import Variable
import torch.nn.functional as F
from torchvision import datasets, transforms
import numpy as np
 
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

model = EfficientNetB0(num_classes=10)

model = torch.load(save_path, map_location="cpu")  # 加载模型
model.to(device)
model.eval()  # 把模型转为test模式

DataParallel(
  (module): EfficientNet(
    (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (layers): Sequential(
      (0): MBconv(
        (conv1): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=32, bias=False)
        (bn2): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(32, 8, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(8, 32, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(32, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (1): MBconv(
        (conv1): Conv2d(16, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(96, 96, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=96, bias=False)
        (bn2): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(96, 4, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(4, 96, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(96, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (2): MBconv(
        (conv1): Conv2d(24, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(144, 144, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=144, bias=False)
        (bn2): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(144, 6, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(6, 144, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(144, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (3): MBconv(
        (conv1): Conv2d(24, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(144, 144, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=144, bias=False)
        (bn2): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(144, 6, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(6, 144, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(144, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (4): MBconv(
        (conv1): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(240, 240, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=240, bias=False)
        (bn2): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(240, 10, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(10, 240, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(240, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (5): MBconv(
        (conv1): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(240, 240, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=240, bias=False)
        (bn2): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(240, 10, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(10, 240, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(240, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (6): MBconv(
        (conv1): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(480, 480, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=480, bias=False)
        (bn2): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(480, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (7): MBconv(
        (conv1): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(480, 480, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=480, bias=False)
        (bn2): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(480, 80, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(80, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (8): MBconv(
        (conv1): Conv2d(80, 480, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(480, 480, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=480, bias=False)
        (bn2): BatchNorm2d(480, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(480, 20, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(20, 480, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(480, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (9): MBconv(
        (conv1): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(672, 672, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=672, bias=False)
        (bn2): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(672, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (10): MBconv(
        (conv1): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(672, 672, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=672, bias=False)
        (bn2): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(672, 112, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(112, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (11): MBconv(
        (conv1): Conv2d(112, 672, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(672, 672, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=672, bias=False)
        (bn2): BatchNorm2d(672, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(672, 28, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(28, 672, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(672, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (12): MBconv(
        (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(1152, 1152, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=1152, bias=False)
        (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (13): MBconv(
        (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(1152, 1152, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=1152, bias=False)
        (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (14): MBconv(
        (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(1152, 1152, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=1152, bias=False)
        (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(1152, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(192, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (15): MBconv(
        (conv1): Conv2d(192, 1152, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn1): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (conv2): Conv2d(1152, 1152, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1152, bias=False)
        (bn2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (se): SE(
          (se1): Conv2d(1152, 48, kernel_size=(1, 1), stride=(1, 1))
          (se2): Conv2d(48, 1152, kernel_size=(1, 1), stride=(1, 1))
        )
        (conv3): Conv2d(1152, 320, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn3): BatchNorm2d(320, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (linear): Linear(in_features=320, out_features=10, bias=True)
  )
)

并且为了方便，定义了一个predict函数，简单思想就是，先resize成网络使用的shape，然后进行变化tensor输入即可，不过这里有一个点，我们需要对我们的图片也进行transforms，因为我们的训练的时候，对每个图像也是进行了transforms的，所以我们需要保持一致

def predict(model, img):
    trans = transforms.Compose([transforms.Resize((32,32)),
                            transforms.ToTensor(),
                            transforms.Normalize(mean=(0.5, 0.5, 0.5), 
                                                 std=(0.5, 0.5, 0.5)),
                           ])
 
    img = trans(img)
    img = img.to(device)
    # 图片扩展多一维,因为输入到保存的模型中是4维的[batch_size,通道,长，宽]，而普通图片只有三维，[通道,长，宽]
    img = img.unsqueeze(0)  # 扩展后，为[1，3，32，32]
    output = model(img)
    prob = F.softmax(output,dim=1) #prob是10个分类的概率
    print("概率",prob)
    value, predicted = torch.max(output.data, 1)
    print("类别",predicted.item())
    print(value)
    pred_class = classes[predicted.item()]
    print("分类",pred_class)

读取本地图片进行预测

# 读取要预测的图片
img = Image.open("./airplane.jpg").convert('RGB') # 读取图像
img

​

​

predict(model, img)

概率 tensor([[9.9339e-01, 1.9589e-05, 4.6372e-04, 2.1349e-06, 3.1111e-05, 6.9647e-07,
         1.7551e-06, 4.4248e-06, 5.8533e-03, 2.3146e-04]], device='cuda:0',
       grad_fn=<SoftmaxBackward>)
类别 0
tensor([9.4914], device='cuda:0')
分类 airplane

读取图片地址进行预测

我们也可以通过读取图片的url地址进行预测，这里我找了多个不同的图片进行预测

import requests
from PIL import Image
url = 'https://dss2.bdstatic.com/70cFvnSh_Q1YnxGkpoWK1HF6hhy/it/u=947072664,3925280208&fm=26&gp=0.jpg'
url = 'https://ss0.bdstatic.com/70cFuHSh_Q1YnxGkpoWK1HF6hhy/it/u=2952045457,215279295&fm=26&gp=0.jpg'
url = 'https://ss0.bdstatic.com/70cFvHSh_Q1YnxGkpoWK1HF6hhy/it/u=2838383012,1815030248&fm=26&gp=0.jpg'
url = 'https://gimg2.baidu.com/image_search/src=http%3A%2F%2Fwww.goupuzi.com%2Fnewatt%2FMon_1809%2F1_179223_7463b117c8a2c76.jpg&refer=http%3A%2F%2Fwww.goupuzi.com&app=2002&size=f9999,10000&q=a80&n=0&g=0n&fmt=jpeg?sec=1624346733&t=36ba18326a1e010737f530976201326d'
url = 'https://img2.baidu.com/it/u=3210145212,1163645841&fm=253&fmt=auto&app=120&f=JPEG?w=875&h=537' # horse
response = requests.get(url, stream=True)
print (response)
img = Image.open(response.raw)
img

<Response [200]>

predict(model, img)

概率 tensor([[1.5856e-03, 2.1046e-04, 3.9647e-03, 1.0489e-02, 7.3925e-02, 4.8408e-02,
         3.3223e-04, 8.6051e-01, 1.6616e-04, 4.1264e-04]], device='cuda:0',
       grad_fn=<SoftmaxBackward>)
类别 7
tensor([5.2634], device='cuda:0')
分类 horse

可以看到，分类还是比较正确的，正确分类了，并且他的置信度大概有86%，还是比较不错的。

10.总结

在这个模型中，其实我们只迭代了10次，EfficientNet实际上已经在10个迭代中已经达到了一个相当不错的结果，之前也被誉为是2020年最好的卷积神经网络，大概在训练集达到了84%的准确率，在测试集有76%的准确率，并且我也暂时没有做任何的数据增强。在原本的论文中，EfficientNet-B0进行分类是用224x224的图片，不过我这里用了32x32的原始CIFAR10图片，所以这个也是可以改进的，并且用更大的分辨率，不过这样可能会导致显存占的比较多，所以可以个人权衡一下资源来进行学习。

顺带提一句，我们的数据和代码都在我的汇总篇里有说明，如果需要，可以自取

这里再贴一下汇总篇：汇总篇