I have a dataset that I have split in train, validation and test subsets. I want to evaluate several CNN architectures and hyperparameters so I have trained several models with different hyperparameter combinations and evaluated them on the validation set.
For the final decission which CNN architecture performs best, I want to take the best hyperparameter combination for each architecture determined in the previous step and perform k-fold Cross Validation on the combined training and validation dataset for each architecture. I get the average accuracy as an estimate how good the final model performs and how robust wrt. data variability the model is.
Finally, I train a model with the train and validation dataset and evaluate it on the test set.
Does my approach make sense? I cannot use k-fold Cross validation for hyperparameter search as it is computationally very expensive.