C2W3 - Pandas vs Caviar

Course Q&A Deep Learning Specialization Improving Deep Neural Networks: Hyperparameter tun
1

Hi all,

I have a (maybe) silly question.

In the lecture about the two main approaches that are followed for tuning the hyperparameters, the final choice about yheir values is done referring to a key metric. For example, this metric could be the prediction error.

Here my question: must this metric be evaluated on the development set, right? In fact, in an earlier lecture it is mentioned that the dev set is used for choosing among different models (i.e. among different hyperparameters settings).

Am I right?

Thanks, ciao, Giacomo

2

Hi Giacomo,

You are correct. We split the whole dataset into 3 parts: training, validation (development) and test set. All model evaluations (different algorithms and different hyperparameter options) are evaluated on the development set.

Once we are satisfied by the performance of our model on the dev set, or have selected the best possible model, we finally pass in the test set through it to ensure that the model can actually generalize well and it didn’t overfit to the development set.

Hope this helps.

3

Thanks so much SomeshChatterjee, now everything’s clear!

Giacomo