I am using keras for binary image classification. I have attached the architecture of the generic CNN model. Please give your suggestions. Is there any mathematical benchmarking or visual representation?
Hello Nakib,
If you are looking to get the best performance out of your own model, then I would suggest "hyper-parameter tuning" while evaluating using some sort of cross-validation scheme. Tensorflow (which works well with keras) has tools for hyperparameter searching https://www.tensorflow.org/tensorboard/hyperparameter_tuning_with_hparams with visualizations available in tensorboard. Additionally Tensorflow Extended has tools for automated model tuning as more data becomes available.
For hyper-parameter tuning in general, you may begin your search with random selections of the hyper-parameter space to help narrow down your search and then use a more narrow grid searching approach. The goal here is to find the global minimum of your model's loss with respect to your hyper-parameter space.
Also, be sure to use a cross-validation method and regularization to avoid over-fitting to your training data.
- Badges
- Science topic