Model selection with MEmPaMaL

Example with Scikit-learn + MEmPaMaL + Soma-Workflow

In this example, we take the classical iris dataset.



In [1]:

    
from mempamal.datasets import iris
X, y = iris.get_data()



In [2]:

    
from sklearn.linear_model.logistic import LogisticRegression
from sklearn.preprocessing.data import StandardScaler
from sklearn.cross_validation import StratifiedKFold
from sklearn.pipeline import Pipeline
from sklearn.metrics import f1_score

The pipeline will contains:

scaling of the data: centering and scaling wrt. the standard deviation
logistic regression with default parameters

The goodness of fit is estimated with:

a 5-folds (stratified) cross-validation
the score function is a F1 score

The model selection is performed with:

a 5-folds (stratified) cross-validation
the score function is a F1 score



In [3]:

    
s1 = StandardScaler(with_mean=True, with_std=True)
s2 = LogisticRegression()
p = [("scaler", s1), ("logit", s2)]
est = Pipeline(p)



In [4]:

    
from mempamal.configuration import JSONify_estimator, JSONify_cv, build_dataset
from mempamal.examples.parameters_grid import make_log_grid
from mempamal.workflow import create_wf, save_wf

We jsonify the estimator and the cross-validation configuration:



In [5]:

    
method_conf = JSONify_estimator(est, out="./est.json")
cv_conf = JSONify_cv(StratifiedKFold, cv_kwargs={"n_folds": 5},
                     score_func=f1_score,
                     inner_cv=StratifiedKFold,
                     inner_cv_kwargs={"n_folds": 5},
                     inner_score_func=f1_score,
                     stratified=True,
                     out="./cv.json")

We build the dataset in the current directory. It's create a dataset.joblib file. Then we create the workflow in our internal format (create_wf). With verbose=True, it prints the commands on stdout. And finally, we output the workflow (save_wf) in the soma-workflow format and write it to workflow.json (need soma-workflow).



In [6]:

    
grid = make_log_grid(X, y)
dataset = build_dataset(X, y, method_conf, cv_conf, ".", grid=grid)
wfi = create_wf(dataset['folds'], cv_conf, method_conf, ".",
               verbose=True)
wf = save_wf(wfi, "./workflow.json", mode="soma-workflow")









    



python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_0_0.pkl 0 --inner 0
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_0_1.pkl 0 --inner 1
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_0_2.pkl 0 --inner 2
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_0_3.pkl 0 --inner 3
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_0_4.pkl 0 --inner 4

python mempamal/scripts/inner_reducer.py ./cv.json ./est.json ./dataset.joblib ./red_res_0.pkl ./map_res_0_{inner}.pkl 0

python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_1_0.pkl 1 --inner 0
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_1_1.pkl 1 --inner 1
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_1_2.pkl 1 --inner 2
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_1_3.pkl 1 --inner 3
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_1_4.pkl 1 --inner 4

python mempamal/scripts/inner_reducer.py ./cv.json ./est.json ./dataset.joblib ./red_res_1.pkl ./map_res_1_{inner}.pkl 1

python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_2_0.pkl 2 --inner 0
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_2_1.pkl 2 --inner 1
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_2_2.pkl 2 --inner 2
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_2_3.pkl 2 --inner 3
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_2_4.pkl 2 --inner 4

python mempamal/scripts/inner_reducer.py ./cv.json ./est.json ./dataset.joblib ./red_res_2.pkl ./map_res_2_{inner}.pkl 2

python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_3_0.pkl 3 --inner 0
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_3_1.pkl 3 --inner 1
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_3_2.pkl 3 --inner 2
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_3_3.pkl 3 --inner 3
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_3_4.pkl 3 --inner 4

python mempamal/scripts/inner_reducer.py ./cv.json ./est.json ./dataset.joblib ./red_res_3.pkl ./map_res_3_{inner}.pkl 3

python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_4_0.pkl 4 --inner 0
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_4_1.pkl 4 --inner 1
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_4_2.pkl 4 --inner 2
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_4_3.pkl 4 --inner 3
python mempamal/scripts/mapper.py ./cv.json ./est.json ./dataset.joblib ./map_res_4_4.pkl 4 --inner 4

python mempamal/scripts/inner_reducer.py ./cv.json ./est.json ./dataset.joblib ./red_res_4.pkl ./map_res_4_{inner}.pkl 4

python mempamal/scripts/outer_reducer.py ./final_res.pkl ./red_res_{outer}.pkl

Now, we create a WorkflowController and we submit the workflow. We wait for workflow completion then we read the final results.



In [7]:

    
from soma_workflow.client import WorkflowController

import time
import json
import sklearn.externals.joblib as joblib

controller = WorkflowController()
wf_id = controller.submit_workflow(workflow=wf, name="second example")

while controller.workflow_status(wf_id) != 'workflow_done':
    time.sleep(2)
print(joblib.load('./final_res.pkl'))









    



light mode
{'std': 0.025080367485459092, 'raw': array([ 0.93333333,  1.        ,  0.96658312,  0.96658312,  0.93265993]), 'median': 0.96658312447786132, 'mean': 0.9598319029897977}