Training, validation, and test

In the previous lesson we introduce the validation data set.

In addition we said we want to divide the initial dataset into three parts.

Training validation end and the topic of this lesson after we have trimmed the model and validated it

it is time to measure its predictive power words.

Actually this is done by running the model on a new data.

It hasn’t seen before.

That’s equivalent to applying the model in real life.

So our data is trained and validated.

We have the final version of the machine learning blackbox right.

We are ready to test it with the test data set.

The accuracy of the prediction we get from this test is the accuracy we would expect the model to have

if we deploy in real life.

So the test data set is the last step we take.

Let’s summarize first.

You get a data set then you split it into three parts.

Here’s where you would probably like to ask a question.

Do we split the data evenly.

How do practitioners approach this.

Well there is no set rule but splits like 80 percent training 10 percent validation and 10 percent test

or 70 2010 are commonly used.

Obviously the data set where we treat the model should be considerably larger than the other two.

You want to devote as much data as possible to the training of the model while having enough samples

to validate and test on.

Third we train the model using the training data set and the training data set.

Only forth.

Every now and then we validate the model by running it for the validation data set.

That’s where you would probably like to ask a second question what does every now and then mean.

Glad you asked.

Usually we validate the data set for every epoch every time we adjust all weights and calculate the

training loss revalidate if the treating loss and the validation loss go hand-in-hand we carry on training

the model Ilive the validation loss is increasing.

We are overfitting.

So we should stop.

Fifth and final step.

You test the model with the test data set the accuracy you obtain at this stage is the accuracy of your

machine learning algorithm.

All right.

These topics are getting more interesting aren’t they.

Keep up the good work and I’ll see you in our next video.