def parameterized_models():
for model_name, d in MODEL_PARAMS.iteritems():
Model = d['model']
params = d['parameters']
for dictionary in grid_search(params):
model = Model(**dictionary)
yield model_name, model
features_to_withhold = np.random.choice(np.arange(data.shape[1]), 2)
y = data[:, features_to_withhold]
data_limited = np.delete(data, features_to_withhold, axis=1)
slc = FullSLCluster(**param_dict)
wcv = WCVScore(slc)
wcv_score, _ = wcv.score(data_limited)
assignments = slc.fit_predict(data_limited)
final_scores = np.apply_along_axis(
lambda col: mean_cluster_variances(assignments, col), 0, y)
final_score = np.mean(final_scores)
print "done testing ", param_dict
return wcv_score, final_score, features_to_withhold
if __name__ == '__main__':
params_list = grid_search(SLC_PARAMS)
datasets = [
load_boston().data,
load_diabetes().data,
load_breast_cancer().data
]
outputs = []
for data in datasets:
outputs.append(
Parallel(n_jobs=-1)(delayed(test_params)(param_dict, data)
for param_dict in params_list))
with open('big_tune_params_by_model.pkl', 'w') as f:
pickle.dump(outputs, f)
with open('big_tune_params.pkl', 'w') as f:
pickle.dump(params_list, f)
initial_variances = [[
output.append((dataset, params[min_wcv]))
return output
def prepare_tuned_models(select_params_output):
params = []
for row in select_params_output:
param = row[1]
param['dataset'] = row[0]
param['model'] = FullSLCluster
params.append(param)
return params
if __name__ == '__main__':
models_to_search = grid_search(SLC_PARAMS)
# fit_model_output = Parallel(n_jobs=-1)(delayed(fit_model)(model) for model in models_to_search)
fit_model_output = [fit_model(model) for model in models_to_search]
with open('big_ole_grid_search_new_datasets.pkl', 'w') as f:
pickle.dump(fit_model_output, f)
params_per_dataset = select_params_per_dataset_per_k(fit_model_output)
params_to_retry = prepare_tuned_models(params_per_dataset)
# final_models = Parallel(n_jobs=-1)(delayed(fit_model)(model) for model in params_to_retry)
final_models = [fit_model(model) for model in params_to_retry]
with open('final_models_from_grid_search_new_datasets.pkl', 'w') as f:
pickle.dump(final_models, f)
models_to_search = grid_search(KMEANS_PARAMS)
# fit_model_output = Parallel(n_jobs=-1)(delayed(fit_model)(model) for model in models_to_search)
fit_model_output = [fit_model(model) for model in models_to_search]
with open('big_ole_grid_search_new_datasets_kmeans.pkl', 'w') as f: