Change StratifiedShuffleSplit to ttrain_test_split

[pronojitsaha]

Addresses #99

[rhiever]

Hmmm... it's saying that no changes were made. Did you update from master and overwrite your changes?

[pronojitsaha]

In hurry, forgot to commit! Should be good now.

[rhiever]

Thanks! Did you verify that it returns the same splits? If not, we'll have to verify that before merging.

[rhiever]

Just checked and, by default, train_test_split doesn't stratify the data by class. You have to pass the stratify option and a list of the class labels, e.g.,

X_train, X_test, y_train, y_test = train_test_split(input_data.drop('class', axis=1).values, 
                                                    input_data['class'].values,
                                                    train_size=0.75, test_size=0.25,
                                                    random_state=RANDOM_STATE,
                                                    stratify=input_data['class'].values)

Please make that change and let's see if that passes on Travis-CI. If it does, we'll merge away!

[pronojitsaha]

Ok..sure.

[pronojitsaha]

I have checked the splits, and the split ratio is same for both train_test_split and StratifiedShuffleSplit, but the split indices are somewhat different which is expected.

[rhiever]

If you set the random_state to the same thing in your tests, they should come out the same. That's what I verified on my end yesterday.

[rhiever]

This doesn't look right. The call needs to look something like:

(training_features, testing_features,
training_labels, testing_labels) = train_test_split(input_data.drop('class', axis=1).values, 
                                                    input_data['class'].values,
                                                    train_size=0.75, test_size=0.25,
                                                    random_state=RANDOM_STATE,
                                                    stratify=input_data['class'].values)

Have you tested this?

[pronojitsaha]

Yes, I have tested it on IRIS and MNIST dataset. Works the same. We can also do it the way you have pointed out, but using training_testing_data.index to get training_indices, testing_indices is in line with rest of our code format.

[pronojitsaha]

Ok, got that. Thanks.

[rhiever]

Understood. Alright, looks good to merge. Thanks again! :-)