讀後摘要 DpLRN with TF 2/E – part 1

Supervised LRN

  • 使用 labeled/pre-defined training data
  • 缺點:problem of generalization
  • 典型 workflow:
    1. 切分資料為:training, development/validation, test sets
    2. 使用 training set 訓練模型
    3. 訓練時,使用 validation set 驗證 training set, 減輕 overfitting. 以及 regularization 模型
    4. 使用(模型未曾見過的) test set 評估模型效能
    5. Tuning model based on hyper-parameter optimization
    6. Deploy best model into a production-ready environment
  • 有「分類」以及「迴歸」兩種:依照輸出是否為連續值來判斷。
  • 分類問題中,要注意「Unbalanced (input) data」:
    * Little unbalanced: 60% for one class, 40% for the other class
    >> 此時需要將 data 隨機分為三份:50% training, 20% validation, 30% testing

    Pasted Graphic 2

Continue reading