I am doing a time series analysis for the prediction of a hydroclimatic variable on a daily scale. in case of rainfall, my training loss is always much higher than the validation loss. I have tried different architectures but still no improvement. Can someone suggest me how can I tune the model using Keras?
When your training loss is more than the validation loss, it means your validation set is not a true representative of the training set. Try to use different validation sets and then see if you are getting the same performance.
The simplest way to tune the hyperparameter (considering you have the set of layers ready but need to tune epoch, neurons, dropout, etc) is to use sklearn Grid Search.
Make sure you have the computational resource for grid searching, it is a good way to start tuning.
Good luck
- Badges
- Science topic