[IMBD 2021] Pytorch-2021-IMBD-Reggression

• 2021全國智慧製造大數據分析競賽

└── README.md 

主要訓練程式碼
├── train.py                執行訓練檔
├── training.yaml           調整訓練參數
├── dataset.py              讀取訓練驗證資料
├── model.py                網路架構
└── checkpoint              訓練完成存模型及log的資料夾

主要測試程式碼   
├── demo.py                 執行預測並匯出結果csv
└── test.py                 執行預測並計算成績(有答案)

其他程式碼
├── utils
|    ├── csv_utils          csv檔相關函式
|    ├── dir_utils          路徑相關函式
|    ├── model_utils        網路模型相關函式
|    └── score_utils        計算分數相關函式
├── csv_data
|    ├── testing            測試csv資料夾    
|    └── training           訓練csv資料夾
└── colab ver_              Colab版本

0. Competition result

• Preliminary uploaded data

  • Explain File
  • Test Result
  • Test Report

• Preliminary result

  • Participating teams: 118
  • Official Preliminary list

• Finals uploaded data

  • Confidentiality agreement which couldn't publish the finals' data!

• Finals result

  • Participating teams: 31
  • Official Finalists

1. Training

1.1 Prepair training data

• Official training data: 98072

• Data format (Our objective is entering the input features F1 ~ F13 and predict the final Output)

  Data number	F1	F2	F3	F4	F5	F6	F7	F8	F9	F10	F11	F12	F13	Output
  1	0	23.5	23.6	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	0
  2	0	23.5	23.6	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	-0.6
  3	0	23.5	23.5	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	0.6
  4	0	23.5	23.5	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	-0.6
  5	0	23.5	23.6	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	-0.3
  ......	...	...	...	...	...	...	...	...	...	...	...	...	...	...

• CSV to independent data
  Because the training data has some deviations which the same input feature values get different output results as showed below:

  Data number	F1	F2	F3	F4	F5	F6	F7	F8	F9	F10	F11	F12	F13	Output
  1	0	23.5	23.6	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	0
  2	0	23.5	23.6	23.6	23.6	23.8	24.3	23.6	23.5	22.6	23.3	23.1	22.3	-0.6
  ......	...	...	...	...	...	...	...	...	...	...	...	...	...	...

  You could run csv_utils.py to let all training data are independent with mean or mediam value of output.

1.2 Set hyperparameters and train

• Configuration file: training.yaml

    TRAINING:
      Network: 'MLP'
      EPOCH: 1000
      LR: 0.01
      LR_MIN: 0.0001
      GPU: true
      BATCH: 1000
      VAL_RATE: 0.8  # split validation set from training set
      VAL_AFTER_EVERY: 1  # save the model per ? epoch
      TRAIN_DIR: './csv_data/training/independent_mean.csv'  # path to training data
      SAVE_DIR: './checkpoints'  # path to save models and images
  
  

• Start training: train.py

  python train.py
  

1.3 Training and validation loss curve

• log file direction: checkpoints -> log folder

  tensorboard --logdir [log path]
  

2. Testing

2.1 Prepair preliminary testing data

• Official preliminary testing data: 7222

2.2 Load the model and test

• Pretrained model:

  • MLP 13-128-256-128-1 mean 0.8
  • MLP 13-128-256-128-1 mean 0.9
  • MLP 13-128-256-128-1 median 0.8
  • MLP 13-128-256-256-256-128-1 mean 0.8

• Model weight file direction: checkpoints -> model folder

• Start testing: demo.py

  python demo.py  
  

2.3 Score

• Official score calculate rule:
  

3. Reference

• https://lulaoshi.info/machine-learning/neural-network/pytorch-kaggle-house-prices.html
• https://shashikachamod4u.medium.com/excel-csv-to-pytorch-dataset-def496b6bcc1
• https://blog.csdn.net/just_sort/article/details/103110806