EfficientNet 训练测试自己的分类数据集

EfficientNet 训练测试自己的分类数据集

Google团队19年的EfficientNet和EfficientDet在图像分类和目标检测方面都取得了惊艳的效果。EfficientNet既快又准，目测这两年将会成为比ResNet更广泛使用的Backbone。下面简单介绍在自己的分类数据集上运行EfficientNet。

1. 准备代码

Pytorch版本官方代码在这：EfficientNet-Pytorch
为了方便，你也可以下载这个完整代码：Code 来更快知道如何使用。

2. 准备数据集

数据集的文件位置大概是这样：

• dataset_name
  • model
  • train
    • 1
    • 2
    • …
  • test
    • 1
    • 2
    • …

为了便于理解，我提供了一个OxFlower17公开的花卉数据集，转换好格式供参考：

百度云链接: https://pan.baidu.com/s/1oBqSMuyuYmiCg7LVbqUtAQ
密码: cf19

3. 准备权重

这一步在完整代码里可以自行选择何种方式加载权重

（1）自动下载

model_ft = EfficientNet.from_pretrained('efficientnet-b0')

（2）手动下载

net_name = 'efficientnet-b0'
model_ft = EfficientNet.from_name(net_name)
net_weight = 'eff_weights/' + 'efficientnet-b0-355c32eb.pth'
state_dict = torch.load(net_weight)
model_ft.load_state_dict(state_dict)

4. 训练测试完整代码

修改最前面的一些参数符合你的数据集即可，另外需要根据你是使用git工程还是pip安装库决定EfficientNet的导入方式，代码里面已经标了一下，程序会自动保存最好的结果，并在最后进行测试。

from __future__ import print_function, division

import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
from torchvision import datasets, models, transforms
import time
import os
# 如果使用上面的Git工程的话这样导入
# from efficientnet.model import EfficientNet
# 如果使用pip安装的Efficient的话这样导入
from efficientnet_pytorch import EfficientNet

# some parameters
use_gpu = torch.cuda.is_available()
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
data_dir = 'Flowers'
batch_size = 64
lr = 0.01
momentum = 0.9
num_epochs = 60
input_size = 224
class_num = 17
net_name = 'efficientnet-b0'


def loaddata(data_dir, batch_size, set_name, shuffle):
    data_transforms = {
        'train': transforms.Compose([
            transforms.Resize(input_size),
            transforms.CenterCrop(input_size),
            transforms.RandomAffine(degrees=0, translate=(0.05, 0.05)),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ]),
        'test': transforms.Compose([
            transforms.Resize(input_size),
            transforms.CenterCrop(input_size),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ]),
    }

    image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in [set_name]}
    # num_workers=0 if CPU else =1
    dataset_loaders = {x: torch.utils.data.DataLoader(image_datasets[x],
                                                      batch_size=batch_size,
                                                      shuffle=shuffle, num_workers=1) for x in [set_name]}
    data_set_sizes = len(image_datasets[set_name])
    return dataset_loaders, data_set_sizes


def train_model(model_ft, criterion, optimizer, lr_scheduler, num_epochs=50):
    train_loss = []
    since = time.time()
    best_model_wts = model_ft.state_dict()
    best_acc = 0.0
    model_ft.train(True)
    for epoch in range(num_epochs):
        dset_loaders, dset_sizes = loaddata(data_dir=data_dir, batch_size=batch_size, set_name='train', shuffle=True)
        print('Data Size', dset_sizes)
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)
        optimizer = lr_scheduler(optimizer, epoch)

        running_loss = 0.0
        running_corrects = 0
        count = 0

        for data in dset_loaders['train']:
            inputs, labels = data
            labels = torch.squeeze(labels.type(torch.LongTensor))
            if use_gpu:
                inputs, labels = Variable(inputs.cuda()), Variable(labels.cuda())
            else:
                inputs, labels = Variable(inputs), Variable(labels)

            outputs = model_ft(inputs)
            loss = criterion(outputs, labels)
            _, preds = torch.max(outputs.data, 1)

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            count += 1
            if count % 30 == 0 or outputs.size()[0] < batch_size:
                print('Epoch:{}: loss:{:.3f}'.format(epoch, loss.item()))
                train_loss.append(loss.item())

            running_loss += loss.item() * inputs.size(0)
            running_corrects += torch.sum(preds == labels.data)

        epoch_loss = running_loss / dset_sizes
        epoch_acc = running_corrects.double() / dset_sizes

        print('Loss: {:.4f} Acc: {:.4f}'.format(
            epoch_loss, epoch_acc))

        if epoch_acc > best_acc:
            best_acc = epoch_acc
            best_model_wts = model_ft.state_dict()
        if epoch_acc > 0.999:
            break

    # save best model
    save_dir = data_dir + '/model'
    os.makedirs(save_dir, exist_ok=True)
    model_ft.load_state_dict(best_model_wts)
    model_out_path = save_dir + "/" + net_name + '.pth'
    torch.save(model_ft, model_out_path)

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))

    return train_loss, best_model_wts


def test_model(model, criterion):
    model.eval()
    running_loss = 0.0
    running_corrects = 0
    cont = 0
    outPre = []
    outLabel = []
    dset_loaders, dset_sizes = loaddata(data_dir=data_dir, batch_size=16, set_name='test', shuffle=False)
    for data in dset_loaders['test']:
        inputs, labels = data
        labels = torch.squeeze(labels.type(torch.LongTensor))
        inputs, labels = Variable(inputs.cuda()), Variable(labels.cuda())
        outputs = model(inputs)
        _, preds = torch.max(outputs.data, 1)
        loss = criterion(outputs, labels)
        if cont == 0:
            outPre = outputs.data.cpu()
            outLabel = labels.data.cpu()
        else:
            outPre = torch.cat((outPre, outputs.data.cpu()), 0)
            outLabel = torch.cat((outLabel, labels.data.cpu()), 0)
        running_loss += loss.item() * inputs.size(0)
        running_corrects += torch.sum(preds == labels.data)
        cont += 1
    print('Loss: {:.4f} Acc: {:.4f}'.format(running_loss / dset_sizes,
                                            running_corrects.double() / dset_sizes))


def exp_lr_scheduler(optimizer, epoch, init_lr=0.01, lr_decay_epoch=10):
    """Decay learning rate by a f#            model_out_path ="./model/W_epoch_{}.pth".format(epoch)
#            torch.save(model_W, model_out_path) actor of 0.1 every lr_decay_epoch epochs."""
    lr = init_lr * (0.8**(epoch // lr_decay_epoch))
    print('LR is set to {}'.format(lr))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

    return optimizer


# train
pth_map = {
    'efficientnet-b0': 'efficientnet-b0-355c32eb.pth',
    'efficientnet-b1': 'efficientnet-b1-f1951068.pth',
    'efficientnet-b2': 'efficientnet-b2-8bb594d6.pth',
    'efficientnet-b3': 'efficientnet-b3-5fb5a3c3.pth',
    'efficientnet-b4': 'efficientnet-b4-6ed6700e.pth',
    'efficientnet-b5': 'efficientnet-b5-b6417697.pth',
    'efficientnet-b6': 'efficientnet-b6-c76e70fd.pth',
    'efficientnet-b7': 'efficientnet-b7-dcc49843.pth',
}
# 自动下载到本地预训练
# model = EfficientNet.from_pretrained('efficientnet-b0')
# 离线加载预训练，需要事先下载好
model_ft = EfficientNet.from_name(net_name)
net_weight = 'eff_weights/' + pth_map[net_name]
state_dict = torch.load(net_weight)
model_ft.load_state_dict(state_dict)

# 修改全连接层
num_ftrs = model_ft._fc.in_features
model_ft._fc = nn.Linear(num_ftrs, class_num)

criterion = nn.CrossEntropyLoss()
if use_gpu:
    model_ft = model_ft.cuda()
    criterion = criterion.cuda()

optimizer = optim.SGD((model_ft.parameters()), lr=lr,
                      momentum=momentum, weight_decay=0.0004)

train_loss, best_model_wts = train_model(model_ft, criterion, optimizer, exp_lr_scheduler, num_epochs=num_epochs)

# test
print('-' * 10)
print('Test Accuracy:')
model_ft.load_state_dict(best_model_wts)
criterion = nn.CrossEntropyLoss().cuda()
test_model(model_ft, criterion)

5. 实验结果

没怎么调参，粗略地跑了一下：

Network	# Params	Accuracy
resnet-50	25.6M	95.88
resnet-101	44.5M	97.35
efficientnet-b0	5.3M	95.29
efficientnet-b1	7.8M	97.94
efficientnet-b2	9.2M	98.24
efficientnet-b3	12M	98.24
efficientnet-b4	19M	98.24

结果确实出彩，参数量比Resnet少性能却依旧强劲！