使用pytorch构建resnet50-v2

• 本文为365天深度学习训练营 中的学习记录博客
• 原作者：K同学啊|接辅导、项目定制

resnet-v2改进点以及和v1差别

改进点：

(a)original表示原始的ResNet的残差结构，(b)proposed表示新的ResNet的残差结构。

主要差别就是

(a)结构先卷积后进行BN和激活函数计算，最后执行addition后再进行ReLU计算；(b)结构先进性BN和激活函数计算后卷积，把addition后的ReLU计算放到了残差结构内部。

改进结果：作者使用这两种不同的结构再CIFAR-10数据集上做测试，模型用的是1001层的ResNet模型。从图中的结果我们可以看出，(b)proposed的测试集错误率明显更低一些，达到了4.92%的错误率。(a)original的测试集错误率是7.61%

在相同的数据集来实现（之前是resnet50）

导入需要的包

import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
import os, PIL
import numpy as np
from torch.utils.data import DataLoader,Subset
from torchvision import transforms
from torchvision.datasets import ImageFolder

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
plt.rcParams["font.sans-serif"] = ["SimHei"]
plt.rcParams["axes.unicode_minus"] = False

遍历并展示图片数量

path = "./data/bird_photos"
f = []
for root, dirs, files in os.walk(path):
    for name in files:
        f.append(os.path.join(root, name))
print("图片总数：",len(f))

图片总数： 565

导入数据

transform = transforms.Compose([
    transforms.Resize(224), #统一图片大小
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406],[0.229, 0.224, 0.225]) #标准化
])
data = ImageFolder(path, transform = transform)

print(data.class_to_idx)
print(data.classes)

{‘Bananaquit’: 0, ‘Black Skimmer’: 1, ‘Black Throated Bushtiti’: 2, ‘Cockatoo’: 3}
[‘Bananaquit’, ‘Black Skimmer’, ‘Black Throated Bushtiti’, ‘Cockatoo’]

可视化数据

def imageshow(data, idx, norm = None, label = False):
    plt.figure(dpi=100,figsize=(12,4))
    for i in range(15):
        plt.subplot(3, 5, i + 1)
        img = data[idx[i]][0].numpy().transpose((1, 2, 0))
        if norm is not None:
            mean = norm[0]
            std = norm[1]
            img = img * std + mean
        img = np.clip(img, a_min = 0, a_max=1)
        plt.imshow(img)
    
        if label:
            plt.title(data.classes[data[idx[i]][1]])
        plt.axis('off')
        plt.tight_layout(pad=0.5)
    plt.show()

norm = [[0.485, 0.456, 0.406],[0.229, 0.224, 0.225]]
np.random.seed(22)
demo_img_ids = np.random.randint(564,size = 15)
imageshow(data, demo_img_ids, norm = norm, label = True) 

配置数据集

torch.manual_seed(123)
train_ds, test_ds = torch.utils.data.random_split(data, [452,113])
train_dl = DataLoader(train_ds, batch_size = 8, shuffle=True, prefetch_factor=2) #prefetch_factor 加速训练， shffule 打乱数据
test_dl = DataLoader(train_ds, batch_size = 8, shuffle=False, prefetch_factor=2 )

再次检查数据

dataiter = iter(train_dl)
print("batch img shape: ", dataiter.next()[0].shape)
print("batch label shape:", dataiter.next()[1].shape)

构建模型

import torch
import torch.nn as nn
class BottleBlock(nn.Module):
    def __init__(self, in_c, out_c, stride = 1, use_maxpool = False, K=False):
        super().__init__()
        if in_c != out_c * 4:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_c, out_c * 4, kernel_size=1,stride=stride),  
                nn.BatchNorm2d(out_c * 4)
                )
        else: self.downsample = nn.Identity()
        
        if use_maxpool:
            self.downsample = nn.MaxPool2d(kernel_size = 1,stride =2)
            stride = 2

        self.bn1 = nn.BatchNorm2d(in_c)
        self.ac = nn.ReLU(inplace=True)
        self.conv1 = nn.Conv2d(in_c, out_c, kernel_size=1)
        
        self.bn2 = nn.BatchNorm2d(out_c)
        self.conv2 = nn.Conv2d(out_c, out_c, kernel_size=3, padding=1,stride = stride)
        
        self.bn3 = nn.BatchNorm2d(out_c)
        self.conv3 = nn.Conv2d(out_c, out_c * 4, kernel_size=1)
        self.K = K


    def forward(self, x):
        out = self.bn1(x)
        out = self.ac(out)
        out = self.conv1(out)

        out = self.bn2(out)
        out = self.ac(out)
        out = self.conv2(out)
        
        out = self.bn3(out)
        out = self.ac(out)
        out = self.conv3(out)
        x = self.downsample(x)
        
        return x + out
    
class resnet_50_v2(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7,stride = 2,padding=3)
        self.bn = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2 ,padding=1)
        
        self.layer1 = self.make_layer(BottleBlock, [64,64], 3, use_maxpool_layer = 3,K=True)
        self.layer2 = self.make_layer(BottleBlock, [256,128], 4, use_maxpool_layer=4,K=True)
        self.layer3 = self.make_layer(BottleBlock, [512,256], 6, use_maxpool_layer=3,K=True)
        self.layer4 = self.make_layer(BottleBlock, [1024,512], 3,K=True)
        
        self.bn2 = nn.BatchNorm2d(2048)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Sequential(
            nn.Flatten(),
            nn.Linear(2048,4))
        
    #创建一个拼接的模块    
    def make_layer(self, module, filters, n_layer, use_maxpool_layer=0, K=False):
        filter1, filter2 = filters
        layers = nn.Sequential()
        layers.add_module('0', module(filter1, filter2,K=K))
        
        filter1 = filter2*4
        for i in range(1, n_layer):
            if i == use_maxpool_layer-1:
                layers.add_module(str(i), module(filter1, filter2,use_maxpool=True,K=False))
            else:
                layers.add_module(str(i), module(filter1, filter2, K=False))
        return layers

    def forward(self,x):
        
        out = self.conv1(x)
        out = self.bn(out)
        out = self.relu(out)
        
        out = self.pool1(out)
        
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)      
        
        out = self.bn2(out)
        out = self.relu(out)
        out = self.avgpool(out)
        return self.fc(out)
model = resnet_50_v2()
model

模型结构

resnet_50_v2(
(conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3))
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(pool1): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
(layer1): Sequential(
(0): BottleBlock(
(downsample): Sequential(
(0): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1))
)
(1): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1))
)
(2): BottleBlock(
(downsample): MaxPool2d(kernel_size=1, stride=2, padding=0, dilation=1, ceil_mode=False)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
(bn3): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1))
)
)
(layer2): Sequential(
(0): BottleBlock(
(downsample): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
)
(1): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
)
(2): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
)
(3): BottleBlock(
(downsample): MaxPool2d(kernel_size=1, stride=2, padding=0, dilation=1, ceil_mode=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
(bn3): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
)
)
(layer3): Sequential(
(0): BottleBlock(
(downsample): Sequential(
(0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
(1): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
(2): BottleBlock(
(downsample): MaxPool2d(kernel_size=1, stride=2, padding=0, dilation=1, ceil_mode=False)
(bn1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
(3): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
(4): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
(5): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1))
)
)
(layer4): Sequential(
(0): BottleBlock(
(downsample): Sequential(
(0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(bn1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1))
)
(1): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1))
)
(2): BottleBlock(
(downsample): Identity()
(bn1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(ac): ReLU(inplace=True)
(conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1))
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1))
)
)
(bn2): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(avgpool): AdaptiveAvgPool2d(output_size=(1, 1))
(fc): Sequential(
(0): Flatten(start_dim=1, end_dim=-1)
(1): Linear(in_features=2048, out_features=4, bias=True)
)
)

训练

import copy
def train_model(epoch, model, optim, loss_function, train_dl, test_dl=None,lr_scheduler = None):
    for k in range(epoch):
        model.train()
        train_acc,train_len = 0, 0 
        history={"train_loss":[],
                "train_acc":[],
                "val_loss":[],
                "val_acc":[],
                'best_val_acc': (0, 0)}
        best_acc = 0.0
        for step, (x, y) in enumerate(train_dl):
            #print(step)
            x,y = x.to(device), y.to(device)
            y_h = model(x)
            optim.zero_grad()
            loss = loss_function(y_h, y)
            loss.backward()
            optim.step()
            predict = torch.max(y_h,dim=1)[1]
            train_acc += (predict == y).sum().item()
            train_len += len(y)
            trainacc = train_acc/train_len*100
            history['train_loss'].append(loss)
            history['train_acc'].append(trainacc)
             
        val_loss = 0.0
        val_acc,val_len = 0, 0  
        if test_dl is not None:
            model.eval()
            with torch.no_grad():
                for _, (x,y) in enumerate(test_dl):
                    x,y = x.to(device), y.to(device)
                    y_h = model(x)
                    loss = loss_function(y_h, y)
                    val_loss += loss.item()
                    predict = torch.max(y_h,dim=1)[1]
                    val_acc += (predict == y).sum().item()
                    val_len += len(y)
            acc = val_acc/val_len*100
            history['val_loss'].append(val_loss)
            history['val_acc'].append(acc)
        if (k+1)%5 == 0:
            print(f"epoch: {k+1}/{epoch}\n   train_loss: {history['train_loss'][-1]},\n   train_acc: {trainacc},\n   val_acc: {acc},\n   val_loss: {history['val_loss'][-1]}")
        if lr_scheduler is not None: 
            lr_scheduler.step()    
        if best_acc< trainacc:
            best_acc = acc
            history['best val acc']= (epoch, acc)
            best_model_wts = copy.deepcopy(model.state_dict())
    print('=' * 80)
    print(f'Best val acc: {best_acc}')        
    return model, history, best_model_wts

model = model.to(device)
loss_fn = nn.CrossEntropyLoss()
optims = torch.optim.Adam(model.parameters(), lr = 1e-2)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optims, 50, last_epoch=-1)
model, history, best_model_wts = train_model(50, model, optims, loss_fn, train_dl, test_dl,lr_scheduler)

epoch: 5/50
train_loss: 1.3529443740844727,
train_acc: 60.84070796460177,
val_acc: 50.66371681415929,
val_loss: 67.96911036968231
epoch: 10/50
train_loss: 0.3679060935974121,
train_acc: 67.03539823008849,
val_acc: 74.5575221238938,
val_loss: 39.28975901007652
epoch: 15/50
train_loss: 1.636350393295288,
train_acc: 75.22123893805309,
val_acc: 80.53097345132744,
val_loss: 30.901596864685416
epoch: 20/50
train_loss: 0.5659410953521729,
train_acc: 77.21238938053098,
val_acc: 77.65486725663717,
val_loss: 30.538477830588818
epoch: 25/50
train_loss: 0.11142534017562866,
train_acc: 81.85840707964603,
val_acc: 88.93805309734513,
val_loss: 16.465759705752134
epoch: 30/50
train_loss: 0.14212661981582642,
train_acc: 86.72566371681415,
val_acc: 87.38938053097345,
val_loss: 18.4492209777236
epoch: 35/50
train_loss: 1.4149291515350342,
train_acc: 87.38938053097345,
val_acc: 97.34513274336283,
val_loss: 8.108955120667815
epoch: 40/50
train_loss: 0.651155412197113,
train_acc: 94.69026548672566,
val_acc: 97.78761061946902,
val_loss: 3.9437274279771373
epoch: 45/50
train_loss: 0.062520831823349,
train_acc: 96.01769911504425,
val_acc: 99.33628318584071,
val_loss: 2.148590993718244
epoch: 50/50
train_loss: 0.03985146805644035,
train_acc: 95.13274336283186,
val_acc: 99.77876106194691,
val_loss: 1.442910757730715
================================================================================
Best val acc: 99.77876106194691

在这个数据集上，效果确实比resnet原版好一些