I don't think there is a numerically correct value, seeing as everyone will have different test sizes for their training sets.

mnguyen wrote:

I finished implementing everything and according to the python plot, it looks like it has a proper convergence.  But I don't know if it's "numerically correct".  What values is everyone getting for your initial cost functions before running gradient_descent?  Mine is 866.58, which seems off and doesn't look like it's predicting anything even though the graph is converging.

If it's converging, then you did it right. It might be there isn't a strong relationship between the features and the y value. You can try doing a few predictions and seeing how close it is. Next week there will be an exercise comparing your implementations with sci-kit learn's values. 

I think rkmx52 is right, we should split sets because we want to avoid overfitting.

As long as we're using test_size = 0.33, the returning training set size (not the content) should be equal.

rkmx52 wrote:

msargent wrote:

The training set should be the same size, if you are using the same milage data.

Are you saying that we shouldn't split the data into training/test sets? I thought that was encouraged.

FYI, I am referring to this, from sklearn.cross.validation: 

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42)

Got it. 

msargent wrote:

We split the data into training and test sets when we try to report how accurate our model is. We aren't trying to do that in this assignment: I just asked that you code the algorithm and do a learning curve. Use the whole data set for training. That way we can compare results.