I am using Bayesian hyperparameter optimization for LSTM hyperparameters. This manages to find an optimum set of hyperparameters in much fewer iterations than a grid search.
My problem is that I have a large amount of training data ~14000 samples, each with their own time-series sequence since I am using an LSTM model. This large amount is still slowing the training process substantially.
- Is it possible to trim the training data and optimize hyperparameters on a subset?
- Will this skew the hyperparameter optimizer or will it still manage to find an optimum?
- Is there any research that has been done on this kind of idea?