Flower classification model that classifies flowers in 10 classes.

  • Training and validation are done using a pre-anotated dataset from Kaggle (https://www.kaggle.com/olgabelitskaya/flower-color-images).
  • Dataset is split into training and validation sub-sets (80-20). Test set is not sampled because in Kaggle challenges, test set is avaliable to you only when you are testing the model on the platform.
  • Model is made using transfer learning with an InceptionV3 model with one added fully-connected layer (1024 neurons) and a softmax exit layer.

Dataset is very small but data is evenly distributed so accuracy can be used as a valid metric. In my implementation, accuracy is about 85% while F1 value is about 60%. This does not seem like a lot but this can be greatly improved with a larger dataset and a longer training time.


View Github