# Ignore the new feature as it messes up PCA
data_dict = pickle.load(open("data/own_data_dict.pkl", "r"))
features_list = getallFeatures(data_dict)
data = featureFormat(data_dict, features_list, sort_keys = True)
# Scale features:
mins = np.min(data, axis=0)
maxs = np.max(data, axis=0)
data = (data-mins)/(maxs-mins)
labels, features = targetFeatureSplit(data)
features_train, features_test, labels_train, labels_test = \
stratifiedShuffleSplit(features, labels)
### Do some PCA
pca = PCA.doPCA(features_train, n = 4)
transformed_train = pca.transform(features_train)
# Do some hyperparam validation:
best_svc, svc_grid_scores = ClassifySVM.gridsearch(
transformed_train, labels_train
)
svmfit = ClassifySVM.train(transformed_train, labels_train, best_svc)
test_classifier(svmfit, data)
dump_classifier_and_data(svmfit, data_dict, features_list)
results = {}
for k in xrange(1, len(features_list)):
print k
# Get feature importance:
importance, selector = getFeatureImportance(features_train, labels_train, features_list, k)
### Next try with selector
selector_train = selector.transform(features_train)
# Do some hyperparam validation:
best_svc, svc_grid_scores = ClassifySVM.gridsearch(
selector_train, labels_train
)
svmfit = ClassifySVM.train(selector_train, labels_train, best_svc)
precision, recall, f1, f2 = test_classifier(svmfit, data)
results[k] = [precision, recall, f1, f2]
pprint.pprint(results)
pickle.dump(results, open('featureselection_orig.pkl', "w") )
pcaresults = {}
for k in xrange(1, 7):
print k
### Do some PCA
pca = PCA.doPCA(features_train, n = k)
transformed_train = pca.transform(features_train)
# Do some hyperparam validation:
best_svc, svc_grid_scores = ClassifySVM.gridsearch(
transformed_train, labels_train