Efficient way to perform grid search for reduced set of cross-validated data using sklearn

user_10537500 0
Viewed 64

I am using grid search to find the optimal parameters for 2 models. I have to build one model with entire dataset and another model with reduced dataset (required to keep the folds same for both the models). So in case of second model, a list of datapoints will be omitted/ removed from the same fold which has been used for first model (one with entire dataset). Following is my code:

rkf = RepeatedKFold(n_splits=2, n_repeats=5, random_state=24)
rkf_new_indices = []
for train_idx, test_idx in rkf.split(x):
    Model1x_train, Model1x_test = x[train_idx], x[test_idx]
    Model1y_train, Model1y_test = y[train_idx], y[test_idx]
    temp_list1 = train_idx.copy()
    temp_list2 = test_idx.copy()
    Model2trn_idx = remove_datapoints(temp_list1, out_list)
    Model2tst_idx = remove_datapoints(temp_list2, out_list)
    Model2train_idx = list(Model2trn_idx)
    Model2test_idx = list(Model2tst_idx)
    rkf_new_indices = np.append(Model2train_idx, Model2test_idx)
param_grid = [{'C': [1, 10, 100, 1000], 'kernel': ['linear']}, {'C': [1, 10, 100, 1000], 'gamma': [0.001, 0.0001], 'kernel': ['rbf']},]
svr_model = SVR()
# define search for model with entire dataset
BASE_SVR = GridSearchCV(svr_model, param_grid, scoring='neg_mean_absolute_error', n_jobs=-1, cv=rkf, return_train_score=True)
BASE_SVR_grid_results = BASE_SVR.fit(x, y)

# define search for model with reduced dataset
New_SVR = GridSearchCV(svr_model, param_grid, scoring='neg_mean_absolute_error', n_jobs=-1, cv=rkf_new_indices, return_train_score=True)
#         ^^^^^^^^^^^^ raises TypeError
New_SVR_grid_results = New_SVR.fit(x, y)

For second GridSearch (line #19), I am getting error:

for train, test in self.cv:

> TypeError: cannot unpack non-iterable numpy.int32 object

What am I doing wrong here with cv=rkf_new_indices and how can I fix this?

cross-validation hyperparameters data-science machine-learning python
1 Answers

When you run below segment, the output of the splits are

rkf_new_indices = []
for train_idx, test_idx in rkf.split([8,8,8,8,8,8,8,8,8]):
    print(train_idx, test_idx)
    rkf_new_indices = np.append(train_idx, test_idx)

[0 1 2 3] [4 5 6 7 8]
[4 5 6 7 8] [0 1 2 3]
[2 3 4 7] [0 1 5 6 8]
[0 1 5 6 8] [2 3 4 7]
[1 3 7 8] [0 2 4 5 6]
[0 2 4 5 6] [1 3 7 8]
[1 4 7 8] [0 2 3 5 6]
[0 2 3 5 6] [1 4 7 8]
[1 2 6 7] [0 3 4 5 8]
[0 3 4 5 8] [1 2 6 7]

However, rkf_new_indices = np.append(train_idx, test_idx) is only getting the last instance:

array([0, 3, 4, 5, 8, 1, 2, 6, 7])

You could try rkf_new_indices.append((train_idx, test_idx)) to get all of them in pairs:

[(array([0, 1, 2, 3]), array([4, 5, 6, 7, 8])),
 (array([4, 5, 6, 7, 8]), array([0, 1, 2, 3])),
 (array([2, 3, 4, 7]), array([0, 1, 5, 6, 8])),
 (array([0, 1, 5, 6, 8]), array([2, 3, 4, 7])),
 (array([1, 3, 7, 8]), array([0, 2, 4, 5, 6])),
 (array([0, 2, 4, 5, 6]), array([1, 3, 7, 8])),
 (array([1, 4, 7, 8]), array([0, 2, 3, 5, 6])),
 (array([0, 2, 3, 5, 6]), array([1, 4, 7, 8])),
 (array([1, 2, 6, 7]), array([0, 3, 4, 5, 8])),
 (array([0, 3, 4, 5, 8]), array([1, 2, 6, 7]))]
user_123706870
Related