Flask深度学习模型服务端部署

1、Flask框架简介

Flask是一个使用Python编写的轻量级Web应用框架，可扩展性很强，相较于Django框架，灵活度很高，开发成本底。它仅仅实现了Web应用的核心功能，Flask由两个主要依赖组成，提供路由、调试、Web服务器网关接口的Werkzeug 实现的和模板语言依赖的jinja2，其他的一切都可以由第三方库来完成。

2、Flask框架安装

在使用Flask之前需要安装一下，安装Flask非常简单只需要在在命令行输入

pip install flask即可

3、Flask实现 Hello World案例

# 导入 Flask 类
from flask import Flask
# 创建了这个类的实例。第一个参数是应用模块或者包的名称。
app = Flask(__name__)# 使用 route() 装饰器来告诉 Flask 触发函数的 URL
@app.route("/")
def hello():return "Hello World!"if __name__ == "__main__":# 使用 run() 函数来运行本地服务器和我们的应用app.run()

4、Flask深度学习模型部署

本文通过使用轻量级的WEB框架Flask来实现Python在服务端的部署CIFAR-10的图像分类。效果如下：

CIFAR-10是一个小型图像分类数据集，数据格式类似于MNIST手写数字数据集，在CIFAR-10数据中图片共有10个类别，分别为airplane、automobile、bird、cat、deer、dog、frog、horse、ship、truck。

4.1 数据加载

对于CIFAR-10分类任务，PyTorch里的torchvision库提供了专门数据处理函数torchvision.datasets.CIFAR10,构建DataLoader代码如下：

import torchvision
from torchvision import transforms
import torch
from config import data_folder, batch_sizedef create_dataset(data_folder, transform_train=None, transform_test=None):if transform_train is None:transform_train = transforms.Compose([transforms.RandomCrop(32, padding=4),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465),(0.2023, 0.1994, 0.2010))])if transform_test is None:transform_test = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465),(0.2023, 0.1994, 0.2010))])trainset = torchvision.datasets.CIFAR10(root=data_folder, train=True, download=True, transform=transform_train)trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=2)testset = torchvision.datasets.CIFAR10(root=data_folder, train=False, download=True, transform=transform_test)testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, shuffle=False, num_workers=2)return trainloader, testloader

4.2 构建模型resent18实现分类

from torch import nn
import torch.nn.functional as F# 定义残差块ResBlock
class ResBlock(nn.Module):def __init__(self, inchannel, outchannel, stride=1):super(ResBlock, self).__init__()# 这里定义了残差块内连续的2个卷积层self.left = nn.Sequential(nn.Conv2d(inchannel, outchannel, kernel_size=3, stride=stride, padding=1, bias=False),nn.BatchNorm2d(outchannel),nn.ReLU(inplace=True),nn.Conv2d(outchannel, outchannel, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(outchannel))self.shortcut = nn.Sequential()if stride != 1 or inchannel != outchannel:# shortcut，这里为了跟2个卷积层的结果结构一致，要做处理self.shortcut = nn.Sequential(nn.Conv2d(inchannel, outchannel, kernel_size=1, stride=stride, bias=False),nn.BatchNorm2d(outchannel))def forward(self, x):out = self.left(x)# 将2个卷积层的输出跟处理过的x相加，实现ResNet的基本结构out = out + self.shortcut(x)out = F.relu(out)return outclass ResNet(nn.Module):def __init__(self, ResBlock, num_blocks, num_classes=10):super(ResNet, self).__init__()self.inchannel = 64self.conv1 = nn.Sequential(nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(64),nn.ReLU())self.layer1 = self.make_layer(ResBlock, 64, num_blocks[0], stride=1)self.layer2 = self.make_layer(ResBlock, 128, num_blocks[1], stride=2)self.layer3 = self.make_layer(ResBlock, 256, num_blocks[2], stride=2)self.layer4 = self.make_layer(ResBlock, 512, num_blocks[3], stride=2)self.fc = nn.Linear(512, num_classes)# 这个函数主要是用来，重复同一个残差块def make_layer(self, block, channels, num_blocks, stride):strides = [stride] + [1] * (num_blocks - 1)layers = []for stride in strides:layers.append(block(self.inchannel, channels, stride))self.inchannel = channelsreturn nn.Sequential(*layers)def forward(self, x):# 在这里，整个ResNet18的结构就很清晰了out = self.conv1(x)out = self.layer1(out)out = self.layer2(out)out = self.layer3(out)out = self.layer4(out)out = F.avg_pool2d(out, 4)out = out.view(out.size(0), -1)out = self.fc(out)return outdef resnet18():return ResNet(ResBlock, [2, 2, 2, 2])if __name__ == '__main__':from torchsummary import summaryfrom config import device# vggnet = vgg11().to(device)resnet = resnet18().to(device)# summary(vggnet, (3, 32, 32))summary(resnet, (3, 32, 32))----------------------------------------------------------------Layer (type)               Output Shape         Param #
================================================================Conv2d-1           [-1, 64, 32, 32]           1,728BatchNorm2d-2           [-1, 64, 32, 32]             128ReLU-3           [-1, 64, 32, 32]               0Conv2d-4           [-1, 64, 32, 32]          36,864BatchNorm2d-5           [-1, 64, 32, 32]             128ReLU-6           [-1, 64, 32, 32]               0Conv2d-7           [-1, 64, 32, 32]          36,864BatchNorm2d-8           [-1, 64, 32, 32]             128ResBlock-9           [-1, 64, 32, 32]               0Conv2d-10           [-1, 64, 32, 32]          36,864BatchNorm2d-11           [-1, 64, 32, 32]             128ReLU-12           [-1, 64, 32, 32]               0Conv2d-13           [-1, 64, 32, 32]          36,864BatchNorm2d-14           [-1, 64, 32, 32]             128ResBlock-15           [-1, 64, 32, 32]               0Conv2d-16          [-1, 128, 16, 16]          73,728BatchNorm2d-17          [-1, 128, 16, 16]             256ReLU-18          [-1, 128, 16, 16]               0Conv2d-19          [-1, 128, 16, 16]         147,456BatchNorm2d-20          [-1, 128, 16, 16]             256Conv2d-21          [-1, 128, 16, 16]           8,192BatchNorm2d-22          [-1, 128, 16, 16]             256ResBlock-23          [-1, 128, 16, 16]               0Conv2d-24          [-1, 128, 16, 16]         147,456BatchNorm2d-25          [-1, 128, 16, 16]             256ReLU-26          [-1, 128, 16, 16]               0Conv2d-27          [-1, 128, 16, 16]         147,456BatchNorm2d-28          [-1, 128, 16, 16]             256ResBlock-29          [-1, 128, 16, 16]               0Conv2d-30            [-1, 256, 8, 8]         294,912BatchNorm2d-31            [-1, 256, 8, 8]             512ReLU-32            [-1, 256, 8, 8]               0Conv2d-33            [-1, 256, 8, 8]         589,824BatchNorm2d-34            [-1, 256, 8, 8]             512Conv2d-35            [-1, 256, 8, 8]          32,768BatchNorm2d-36            [-1, 256, 8, 8]             512ResBlock-37            [-1, 256, 8, 8]               0Conv2d-38            [-1, 256, 8, 8]         589,824BatchNorm2d-39            [-1, 256, 8, 8]             512ReLU-40            [-1, 256, 8, 8]               0Conv2d-41            [-1, 256, 8, 8]         589,824BatchNorm2d-42            [-1, 256, 8, 8]             512ResBlock-43            [-1, 256, 8, 8]               0Conv2d-44            [-1, 512, 4, 4]       1,179,648BatchNorm2d-45            [-1, 512, 4, 4]           1,024ReLU-46            [-1, 512, 4, 4]               0Conv2d-47            [-1, 512, 4, 4]       2,359,296BatchNorm2d-48            [-1, 512, 4, 4]           1,024Conv2d-49            [-1, 512, 4, 4]         131,072BatchNorm2d-50            [-1, 512, 4, 4]           1,024ResBlock-51            [-1, 512, 4, 4]               0Conv2d-52            [-1, 512, 4, 4]       2,359,296BatchNorm2d-53            [-1, 512, 4, 4]           1,024ReLU-54            [-1, 512, 4, 4]               0Conv2d-55            [-1, 512, 4, 4]       2,359,296BatchNorm2d-56            [-1, 512, 4, 4]           1,024ResBlock-57            [-1, 512, 4, 4]               0Linear-58                   [-1, 10]           5,130
================================================================
Total params: 11,173,962
Trainable params: 11,173,962
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.01
Forward/backward pass size (MB): 13.63
Params size (MB): 42.63
Estimated Total Size (MB): 56.26
----------------------------------------------------------------

4.3 模型训练

1）损失函数

criterion = nn.CrossEntropyLoss()

2）优化器

optimizer= optim.SGD(net.parameters(), lr=lr, weight_decay=5e-4, momentum=0.9)

3）学习率

一般来说使用SGD优化器时初始学习率一般设置为0.01～0.1，Adam优化器的初始学习率一般设置为0.001～0.01.

4）训练与验证

模型训练过程：

① 梯度清零- optimizer.zero_grad()

② 前向传播- output = net(img)

③ 计算损失- loss = criterion(output, label)

④ 反向传播- loss.backward()

⑤ 更新参数- optimizer.step()

⑥ 验证模型。with torch.no_grad():

net.eval()

from torch import optim, nn
import torch
import os.path as osp
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriterfrom config import epochs, device, data_folder, checkpoint_folder
from data import create_dataset
from model import resnet18def train(net, trainloader, valloader,criteron, epochs, device, model_name='cls'
):best_acc = 0.0best_loss = 1e9writer = SummaryWriter('../CIFAR-10图像分类/log')if osp.exists(osp.join(checkpoint_folder, model_name+'.pth')):net.load_state_dict(torch.load(osp.join(checkpoint_folder, model_name+'.pth')))print('model loading')for n, (num_epochs, lr) in enumerate(epochs):optimizer = optim.SGD(net.parameters(), lr=lr, weight_decay=5e-4, momentum=0.9)for epoch in range(num_epochs):net.train()epoch_loss = 0.0epoch_acc = 0.0for i, (img, label) in tqdm(enumerate(trainloader), total=len(trainloader)):print(trainloader.shape, label.shape)img, label = img.to(device), label.to(device)output = net(img)optimizer.zero_grad()loss = criteron(output, label)loss.backward()optimizer.step()if model_name == 'cls':pred = torch.argmax(output, dim=1)acc = torch.sum(pred == label)epoch_acc += acc.item()epoch_loss += loss.item() * img.shape[0]epoch_loss /= len(trainloader.dataset)if model_name == 'cls':epoch_acc /= len(trainloader.dataset)print('epoch loss: {:.2f} epoch accuracy: {:.2f}'.format(epoch_loss, epoch_acc))writer.add_scalar('epoch_loss_{}'.format(model_name),epoch_loss,sum([e[0] for e in epochs[:n]]) + epoch)writer.add_scalar('epoch_acc_{}'.format(model_name),epoch_acc,sum([e[0] for e in epochs[:n]]) + epoch)else:print('epoch loss: {:.2f}'.format(epoch_loss))writer.add_scalar('epoch_loss_{}'.format(model_name),epoch_loss,sum([e[0] for e in epochs[:n]]) + epoch)with torch.no_grad():net.eval()val_loss = 0.0val_acc = 0.0for i, (img, label) in tqdm(enumerate(valloader), total=len(valloader)):img, label = img.to(device), label.to(device)output = net(img)loss = criteron(output, label)if model_name == 'cls':pred = torch.argmax(output, dim=1)acc = torch.sum(pred == label)val_acc += acc.item()val_loss += loss.item() * img.shape[0]val_loss /= len(valloader.dataset)val_acc /= len(valloader.dataset)if model_name == 'cls':if val_acc > best_acc:best_acc = val_acctorch.save(net.state_dict(),osp.join(checkpoint_folder, model_name + '.pth'))print('validation loss: {:.2f} validation acc: {:.2f}'.format(val_loss, val_acc))writer.add_scalar('validation_loss_{}'.format(model_name),val_loss,sum([e[0] for e in epochs[:n]]) + epoch)writer.add_scalar('validation_acc_{}'.format(model_name),val_acc,sum([e[0] for e in epochs[:n]]) + epoch)writer.close()if __name__ == '__main__':trainloader, valloader = create_dataset(data_folder)net = resnet18().to(device)criteron = nn.CrossEntropyLoss()train(net, trainloader, valloader, criteron, epochs, device)

模型训练完成在checkpoint文件夹下会生成分类模型文件cls.pth

4.4 使用Flask搭建服务接口

from flask import (Flask, render_template, request, flash, redirect, url_for, make_response, jsonify, send_from_directory)
from werkzeug.utils import secure_filename
import os
import time
from datetime import timedelta
from predict import pred
import jsonALLOWED_EXTENSIONS = set(["png","jpg","JPG","PNG", "bmp","jpeg"])def is_allowed_file(filename):return '.' in filename and filename.rsplit('.', 1)[1] in ALLOWED_EXTENSIONSapp = Flask(__name__)
app.config['UPLOAD_FOLDER'] = 'imgs/'
#设置编码
app.config['JSON_AS_ASCII'] = False
# 静态文件缓存过期时间
app.send_file_max_age_default = timedelta(seconds=1)@app.route('/uploads/<filename>')
def uploaded_file(filename):return send_from_directory(app.config['UPLOAD_FOLDER'],filename)@app.route('/result/')
def show_json(filename,class_name):return ({"name":filename, "class_name":class_name})# @app.route("/uploads",methods = ['POST', 'GET'])
@app.route("/",methods = ['POST', 'GET'])
def uploads():if request.method == "POST":#file = request.files['file']uploaded_files = request.files.getlist("file[]")weight_path2 = os.path.join("static", "cls.pth")filenames = {"airplane":[],"automobile":[],"bird":[],"cat":[],"deer":[],"dog":[],"frog":[],"horse":[],"ship":[],"truck":[],}for file in uploaded_files:print(file)if file and is_allowed_file(file.filename):# basepath = os.path.dirname(__file__)filename = secure_filename(file.filename)file.save(os.path.join(app.config['UPLOAD_FOLDER'], filename))img_path = os.path.join(app.config['UPLOAD_FOLDER'], filename)class_index,class_name = pred(img_path, weight_path2)filenames[class_name].append(filename)print(class_index,class_name)result = show_json(filename,class_name)print(result)#data.append(result)#res_json = json.dumps({"status": "200", "msg": "success","data":data})return render_template('upload_more_ok.html',filenames=filenames,)return render_template("upload_more.html", path="./images/test.jpg")if __name__ == "__main__":# 0.0.0.0表示你监听本机的所有IP地址上,通过任何一个IP地址都可以访问到.# port为端口号# debug=Fasle表示不开启调试模式app.run(host='127.0.0.1', port=5000, debug=True)

通过访问以下地址进行图像识别

点击选取文件，dog.jpeg 进行上传

点击识别图片结果如下：

通过以上代码可以实现在服务端进行CIFAR-10图像识别了，也可以通过训练自己的数据集，修改识别的类别数，加载训练好的模型，即可完成指定任务的图像识别。