How to optimize TensorFlow Image classification?

I’ve been using Image classification provided by TensowFlow to classify binary images in my dataset. There are 80 classes in my dataset and only one sample per class. However, I’m planning to increase my dataset in the next steps.
I haven’t been able to achieve a good accuracy. When I studied the documentation, under Create the model, they’ve mentioned:

This model has not been tuned for high accuracy; the goal of this tutorial is to show a standard approach.

When I think about it, there are many ways I can tune the model. For example,

  • Use another model building approach apart from Sequential
  • Use different optimizer class
  • Use different number of Conv/FC layers and and/or neurons
  • Use different activation functions

Overall, there many possibilities of tuning the model. What is the best configuration that I can apply to achieve a good accuracy?

Thank You!

Hi @Hamza, As you said you have a dataset with only one sample per class. Did you train the model with the same dataset or you have trained the model with a dataset having more images per class?

We cannot say the particular method is able to increase the accuracy. It’s an experimental thing. Mainly even if you have a perfect model the accuracy purely depends on the data used for training the model. i.e. the quality of the data directly proportional to the performance of the model. Better training data gives best performance model. Thank You.

Thanks a lot for your reply @Kiran_Sai_Ramineni :slight_smile:

Actually, I’m only concerned about their statement “This model has not been tuned for high accuracy”. I just want to confirm if I’m using the best suited model with the right configuration. After finalizing the model, I’ll play around with the dataset.

Hi @Hamza, The statement mentioned was due to the aim of the tutorial is to show how an image classification can be done. But it is not guaranteed that if you use the same model with the same hyper parameters you can get high accuracy. Also in the tutorial you can see the model is overfitting (having high training accuracy and low validation accuracy). so they mention that statement. Thank You.