The code for "华为云人工智能创新应用大赛". The main tasks of the competition are:

The task of this competition is to classify the pictures of popular scenic spots, delicacies, specialties, folklore and handicrafts in Xi'an.

Many strategies were tried during the competition, with a final score of 0.979 and a ranking of 32 / 732. Although I didn't get a high ranking in the end, I have a deeper understanding of the classification problem, thanks to Huawei Cloud for organizing the competition.

This repositorie is suitable for solving all kinds of classification problems, supports a variety of training skills, and has high expansibility.

Requirements

• Pytorch 1.3.0
• Torchvision 0.4.0
• Python3.7
• pretrainedmodels
• efficientnet-pytorch
• Chinese font-simhei.ttf, Download Link. Put simhei in project_root/font.

How to run

Clone Our Project

git clone https://github.com/XiangqianMa/AI-Competition-HuaWei.git
cd AI-Competition-HuaWei

Prepare Dataset

Download dataset, unzip and put them into your_data_path directory.

Structure of the your_data_path folder can be like:

train_data
label_id_name.json

Create soft links of datasets in the following directories:

cd AI-Competition-HuaWei
mkdir data
cd data
ln -s your_data_path huawei_data

It is important to note that the dataset should be organized in such a way that the following conditions are met:

• The pictures and annotation files are stored in the same folder, as shown in the following figure:

  

• The callout file is in txt format, and its contents are as follows:

  img_1.jpg, 0
  

If it is your own dataset, please process it to be consistent with the above format, otherwise it may cause an error in the running of the program.

Train

cd AI-Competition-HuaWei
python train_classifier.py

HuaWei Cloud Online Test

We use online-service folder to accomplish online test on HuaWei ModelArts platform. The struct of this folder is like the following:

You should put your dependencies into online-service/model like this:

If you want to import customized package in your code, import sentence should be like this:

from model import xxx

Also, you must specified your dependencies in config.json:

"dependencies": [
    {
        "installer": "pip",
        "packages": [
            {
                "package_name": "Pillow",
                "package_version": "5.0.0",
                "restraint": "EXACT"
            },
            {
                "package_name": "torchvision",
                "package_version": "0.2.1",
                "restraint": "EXACT"
            },
            {
                "package_name": "tqdm"
            }                
        ]
    }
]

Demo

First, manually create the following folder:

mkdir data/demo_data

After the training is complete, execute the following code to get a visual classification result, and the test will be carried out on the validation set:

python demo.py

After the test is completed, the sample of the prediction error will be stored in the data/demo_data/results folder. Examples of demos are as follows:

How to use obsutil

Prepare obsutil

Download obsutil from here. Unzip it to ./online-service. After unzipping, ./online-service folder will contain obsutil and setup.sh files

cd online-service

Replace the following -i=xxxxxxx to -i=Access Key Id?AK? and replace the following -k=xxxxxxxxxxxxxxxxx to -k=Secret Access Key?SK?

./obsutil config -i=xxxxxxx -k=xxxxxxxxxxxxxxxxx -e=obs.cn-north-4.myhuaweicloud.com

Then, you will see feedback similar to the one below

Config file url:
  /home/zdkit/.obsutilconfig

Update config file successfully!

Then, execute the following code to check connectivity

./obsutil ls -s

If the returned result contains "Bucket number is:", the configuration is correct. Just like these:

Start at 2019-11-22 15:43:11.060016056 +0000 UTC

obs://ai-competition-zdaiot
obs://test-zdaiot
Bucket number is: 2

Use obsutil

Please Manually change the selected model in online-service/model/customize_service.py file. And manually change the model parameters in online-service/upload.sh. The script will automatically find the newly modified folder from the checkpoints/${model} directory and copy the model_best.pth from it to the online-service/model directory.

If you want to place the model_best.pth file manually, please comment out the cp ../checkpoints/${model}/${filename}/model_best.pth online-service/model line in the online-service/upload.sh file. Then manually place the weight file model_best.pth to the online-service/model directory.

Notice that, when you copy pyfiles from models to online-service/model/deploy_models, you should pay attention to these

• change import path of Python package. For example, change from models.custom_model import CustomModel to from model.deploy_models.custom_model import CustomModel in online-service/model/deploy_models/custom_model.py.
• modify the parameters of prepare_model.create_model function in the online-service/model/customize_service.py and online-service/offline_service.py according to the actual needs.

When you first use it, you should run this:

cd online-service
./upload.sh 0

Then, this script will create a bucket called ai-competition-$USER in your OBS, create a new folder called model_snapshots, and upload the ./model folder to the model_snapshots folder.

$USER is your Linux username

If you want to update the weight file or python files. Please manually put new python files in the corresponding directory. And run this:

./upload.sh

How to train online

Suppose you have built a new bucket called ai-competition-zdaiot

upload dataset

Create a new folder named data in ai-competition-zdaiot bucket . And upload dataset to data. Don't fotget upload label_id_name.json. For example:

upload project

Create a new folder named project inai-competition-zdaiot bucket . And upload Our Project to project. Don't fotget upload font. For example:

upload model_zoo

Create a new folder named model_zoo in ai-competition-zdaiot bucket . And upload pretrained model to model_zoo.

For example, I uploaded a pre-training model called se_resnext101_32x4d-3b2fe3d8.pth.

prepare logs folder

Create a new folder named logs in ai-competition-zdaiot bucket .

Run

Open ModelArts->训练管理->训练作业->创建，Fill in the following parameters

If necessary, add custom parameters in the 训练作业, and then click OK.

During the training process, you can view the running logs through the logs folder under the ai-competition-zdaiot bucket root directory.

After training, the inference scripts and the optimal weight will be saved in the model_snapshots folder under the ai-competition-zdaiot bucket root directory.

Tricks Tried

Dataset

• train_data
• combine
• huge
• adversial_samples（Not work，When the samples and their adversial samples are in training set and verification set respectively, the accuracy of the verification set is too high.）

Data clean

Delete pictures with the following characteristics

• ambiguity
• wrong
• image resolution is too low
• There are no such data characteristics, for example, there are only 石子 in 石子饼 and only 浆水 in 浆水面.
• The target does not occupy the main body of the image. This is especially important for image classification tasks.
• Tourism promotional images, false advertising images, graffiti

Augmentation

• CenterCrop(Not well，only 0.969)
• erase_prob
• gray_prob
• cut_mix
• Use more data augmentation, see here. (Use Resize, RandomHorizontalFlip, ToTensor, jittering, lighting, Normalize. And use cutmix 0.967)

Cross validation

• StratifiedKFold
• KFold

Training strategy

1. Train only on Single image scale
2. The training is divided into two stages, the first stage is training at lower resolution（256）, and the second stage is fine-tuning at higher resolution（416）.
3. The training is divided into three stages, the first stage is trained at a lower resolution（256）, the second stage is fine-tuned at a medium resolution（336）, and the third stage is at a higher resolution（416）.
4. Multi-scale training, switching resolution every several iterations.
5. The training is divided into two stages, the first stage is based on all data sets, and the second stage is only fine-tuning on food. In the second stage of training, get rid of all data enhancement methods.（Not work）
6. Save the optimal model using loss on the validation set（Not work, only 0.972）

Losses

• CrossEntropy
• SmoothCrossEntropy（best）
• FocalLoss
• CB_Sigmoid
• CB_Focal
• CB_Softmax
• CB_Smooth_Softmax（Not work，0.964）

Super parameter

• Lr: If the initial learning rates of backbone and fc layers are the same, the convergence speed is too slow.
• Weight decay: Not work
• Epoch： Try more epochs

Optimizer

• Adam
• SGD

Lr scheduler

Init Lr and weight decay

• StepLR

• CosineLR

• ReduceLR

• MultiStepLR（18/25，Not work 0.9710）

Model type

• densenet201

• efficientnet-b5

• se_resnext101_32x4d

Contributors

• XiangqianMa
• Zdaiot