in data-science ~ read.

Adding callback to a sklearn GridSearch

GridSearch in scikit-learn is just awesome - no doubt about that. But what is unfortunate is the fact that it only shows one metric in the results and you couldn't store any intermediate information or do some actions during the search (such as save every model, or compute additional metrics than just the one).

One (and only?) variant how to achieve that is to create your own custom scoring function, which would do something more than just scoring...

Here is an example and how far I have got.

import pickle  
import time  
import json

def log_metrics_and_params(results, model_savepath):  
    # log results and save path
    to_write = {}
    to_write['results'] = results
    to_write['model_savepath'] = model_savepath'%s', json.dumps(to_write))

def save_model(clf):  
    # save model with timestamp
    timestring = "".join(str(time.time()).split("."))
    model_savepath = 'model_' + timestring + '.pk'
    with open(model_savepath, 'wb') as ofile:
        pickle.dump(clf, ofile)
    return model_savepath

def get_train_metrics():  
    # currently impossible
    # X_train and y_train are in higher scopes

def get_val_metrics(y_pred, y_true):  
    return get_metrics(y_pred, y_true)

def get_metrics(y_pred, y_true):  
    # compute more than just one metrics

    chosen_metrics = {
        'conf_mat': metrics.confusion_matrix,
        'accuracy': metrics.accuracy_score,
        'auc' : metrics.roc_auc_score,

    results = {}
    for metric_name, metric_func in chosen_metrics.items():
            inter_res = metric_func(y_pred, y_true)
        except Exception as ex:
            inter_res = None
            log.parent.error("Couldn't evaluate %s because of %s", metric_name, ex)
        results[metric_name] = inter_res

    results['conf_mat'] = results['conf_mat'].tolist()

    return results

def _my_scorer(clf, X_val, y_true_val):  
    # do all the work and return some of the metrics

    y_pred_val = clf.predict(X_val)

    results = get_val_metrics(y_pred_val, y_true_val)
    model_savepath = save_model(clf)
    log_metrics_and_params(results, model_savepath)
    return results['accuracy']

and then just call gridsearch:

gs = GridSearchCV(clf, tuned_params, scoring=_my_scorer), y)  

Unfortunately, I wasn't able to get X_train and y_train in the scorer scope, so I can't get metrics for that :-( .