Skip to content

AI Track's Exercise to create a basic CNN to recognize different animals from the Chinese Zodiac

License

Notifications You must be signed in to change notification settings

vsay01/Chinese_Zodiac_Sign_Kaggle

Repository files navigation

AI Track's Exercise to create a basic CNN to recognize different animals from the Chinese Zodiac

Inference Results:

  • MobileV2

    • Test Loss: 0.301537
    • Test Accuracy: 90% (595/658)
    • Model size = 23 MB
  • ShuffleNetV2

    • Test Loss: 1.068708
    • Test Accuracy: 70% (465/658)
    • Model size = 17 MB
  • MNASNET

    • Test Loss: 0.368574
    • Test Accuracy: 89% (587/658)
    • Model size = 12 MB
  • DenseNet

    • Test Loss: 0.196492
    • Test Accuracy: 93% (616/658)
    • Model size = 102 MB
  • Vgg16

    • Test Loss: 0.941419
    • Test Accuracy: 86% (569/658)
    • Model size = 1 GB

Training and Validation loss comparisons

Image of Training and Validation loss comparisons

Deploy to Android application:

  • I deployed MNASNET trained model to a sample android application to test this out. This model give acceptable accuracy and smaller size.

Here is the demo video:

Chinese_Zodiac_Sign Classification