Skip to content

Latest commit

 

History

History
102 lines (68 loc) · 4.29 KB

README.md

File metadata and controls

102 lines (68 loc) · 4.29 KB

Question 1: deep learning

(last updated: 3/24)

Description

  1. Setup model via KERAS methods
  2. Deep learning model option: DNN(Deep neural network) or CNN(convolution neural network)
  3. Evaluate performance & compare 10 prediction results.
  4. Details of question:

MY SOLUTION

  1. Environment (settings)

    • Miniconda virtual environment
    • M1 PRO (MPS)
    • Tensorflow version == 2.10.0
    • Python version == 3.8.13

  2. Workflow

    • Datasets split ratio is 7:2:1

      • Training: 3942, Testing: 830, Validation: 1569 (numbers of image)
      • via splitfolder
      • split datasets into three folders respectively
      • splitfolders.ratio('./datasets/', './split_datasets/', seed=1337, ratio=(0.7,0.2,0.1))

    • Model Ariitecture & compile parameters

      • Epochs, Batch size = (200, 32)

        Epochs: the number times that the learning algorithm will work through the entire training dataset.

        Batch size: a number of samples processed before the model is updated

      • optimizer: adam

      • loss function: categorical_crossentropy

      • Convolution neural network (CNN) aritechure

      • Training log (Accuracy vs loss)

        During the training process, I add dropout layer and image augmentation to avoid overfitting, which can observed by the trend of either accuracy or loss curve.

  3. Multi-class classification proformance evalulation

    • Avg accuracy: 89.27%

    • Random pick 15 of the test set, return 15 prediction correct

      Most of the picture can be predicted correctly!

    • Confusion matrix

      Result calculated from my CNN model. We can observed among four classes classification, every classes have no significant difference.

    • Performance (tensorflow labraries built-in library)

      precision recall f1-score support
      daisy(0) 0.96 0.87 0.91 147
      dendelion(1) 0.89 0.92 0.90 200
      rose(2) 0.85 0.87 0.86 151
      sunflower(3) 0.85 0.92 0.89 140
      tulip(4) 0.91 0.89 0.90 192
      accuracy 0.89 830
      macro avg 0.89 0.89 0.89 830
      weighted avg 0.89 0.88 0.89 830

    • Remarks (Definitions):

      Accuracy Score = (TP + TN) / (TP + TN + FP + FN)
Precision = TP / (TP + FP) 
Recall = TP / (TP + FN)
F1 = 2*Precision*Recall / (Precision + Recall)

      • Macro averaged precision: calculate precision for all classes individually and then average them

      • Micro averaged precision: calculate class wise true positive and false positive and then use that to calculate overall precision

      Cited from Kaggle

  4. Reference