I'm training a dataset and then testing it on some other dataset.
To improve performance, I wanted to fine-tune my parameters with a 5-fold cross validation.
However, I think I'm not writing the correct code as when I try to fit the model to my testing set, it says it hasn't fit it yet. I though the cross-validation part fitted the model? Or maybe I have to extract it?
Here's my code:
svm = SVC(kernel='rbf', probability=True, random_state=42)
accuracies = cross_val_score(svm, data_train, lbs_train, cv=5)
pred_test = svm.predict(data_test)
accuracy = accuracy_score(lbs_test, pred_test)
That is correct, the cross_validate_score doesn't return a fitted model. In your example, you have cv=5 which means that the model was fit 5 times. So, which of those do you want? The last?
The function cross_val_score is a simpler version of the sklearn.model_selection.cross_validate. Which doesn't only return the scores, but more information.
So you can do something like this:
from sklearn.model_selection import cross_validate
svm = SVC(kernel='rbf', probability=True, random_state=42)
cv_results = cross_validate(svm, data_train, lbs_train, cv=5, return_estimator=True)
# cv_results is a dict with the following keys:
# 'test_score' which is what cross_val_score returns
# 'train_score'
# 'fit_time'
# 'score_time'
# 'estimator' which is a tuple of size cv and only if return_estimator=True
accuracies = cv_results['test_score'] # what you had before
svms = cv_results['estimator']
print(len(svms)) # 5
svm = svms[-1] # the last fitted svm, or pick any that you want
pred_test = svm.predict(data_test)
accuracy = accuracy_score(lbs_test, pred_test)
Note, here you need to pick one of the 5 fitted SVMs. Ideally, you would use cross-validation for testing the performance of your model. So, you don't need to do it again at the end. Then, you would fit your model one more time, but this time with ALL the data which would be the model you will actually use in production.
Another note, you mentioned that you want this to fine tune the parameters of your model. Perhaps you should look at hyper-parameter optimization. For example: https://datascience.stackexchange.com/a/36087/54395 here you will see how to use cross-validation and define a parameter search space.
Related
I have a time-dependent data set, where I (as an example) am trying to do some hyperparameter tuning on a Lasso regression.
For that I use sklearn's TimeSeriesSplit instead of regular Kfold CV, i.e. something like this:
tscv = TimeSeriesSplit(n_splits=5)
model = GridSearchCV(
estimator=pipeline,
param_distributions= {"estimator__alpha": np.linspace(0.05, 1, 50)},
scoring="neg_mean_absolute_percentage_error",
n_jobs=-1,
cv=tscv,
return_train_score=True,
max_iters=10,
early_stopping=True,
)
model.fit(X_train, y_train)
With this I get a model, which I can then use for predictions etc. The idea behind that cross validation is based on this:
However, my issue is that I would actually like to have the predictions from all the test sets from all cv's. And I have no idea how to get that out of the model ?
If I try the cv_results_ I get the score (from the scoring parameter) for each split and each hyperparameter. But I don't seem to be able to find the prediction values for each value in each test split. And I actually need that for some backtesting. I don't think it would be "fair" to use the final model to predict the previous values. I would imagine there would be some kind of overfitting in that case.
So yeah, is there any way for me to extract the predicted values for each split ?
You can have custom scoring functions in GridSearchCV.With that you can predict outputs with the estimator given to the GridSearchCV in that particular fold.
from the documentation scoring parameter is
Strategy to evaluate the performance of the cross-validated model on the test set.
from sklearn.metrics import mean_absolute_percentage_error
def custom_scorer(clf, X, y):
y_pred = clf.predict(X)
# save y_pred somewhere
return -mean_absolute_percentage_error(y, y_pred)
model = GridSearchCV(estimator=pipeline,
scoring=custom_scorer)
The input X and y in the above code came from the test set. clf is the given pipeline to the estimator parameter.
Obviously your estimator should implement the predict method (should be a valid model in scikit-learn). You can add other scorings to the custom one to avoid non-sense scores from the custom function.
I would like to pre-train a model and then train it with another model.
I have model Decision Tree Classifer and then I would like to train it further with model LGBM Classifier. Is there a possibility to do this in scikit learn?
I have already read this post about it https://datascience.stackexchange.com/questions/28512/train-new-data-to-pre-trained-model.. In the post it says
As per the official documentation, calling fit() more than once will
overwrite what was learned by any previous fit()
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=1)
# Train Decision Tree Classifer
clf = DecisionTreeClassifier()
clf = clf.fit(X_train,y_train)
lgbm = lgb.LGBMClassifier()
lgbm = lgbm.fit(X_train,y_train)
#Predict the response for test dataset
y_pred = lgbm.predict(X_test)
Perhaps you are looking for stacked classifiers.
In this approach, the predictions of earlier models are available as features for later models.
Look into StackingClassifiers.
Adapted from the documentation:
from sklearn.ensemble import StackingClassifier
estimators = [
('dtc_model', DecisionTreeClassifier()),
]
clf = StackingClassifier(
estimators=estimators,
final_estimator=LGBMClassifier()
)
Unfortunately this is not possible at present. According to the doc at https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMClassifier.html?highlight=init_model, you can continue training the model if the model is from lightgbm.
I did try this setup with:
# dtc
dtc_model = DecisionTreeClassifier()
dtc_model = dtc_model.fit(X_train, y_train)
# save
dtc_fn = 'dtc.pickle.db'
pickle.dump(dtc_model, open(dtc_fn, 'wb'))
# lgbm
lgbm_model = LGBMClassifier()
lgbm_model.fit(X_train_2, y_train_2, init_model=dtc_fn)
And I get:
LightGBMError: Unknown model format or submodel type in model file dtc.pickle.db
As #Ferdy explained in his post, there is no simple way to perform this operation and it is understandable.
Scikit-learn DecisionTreeClassifier takes only numerical features and cannot handle nan values whereas LGBMClassifier can handle those.
By looking at the decision function of scikit-learn you can see that all it can perform is splits based on feature <= threshold.
On the contrary LGBM can perform the following:
feature is na
feature <= threshold
feature in categories
Splits in decision tree are selected at each step as they best splits the set of items. They try to minimize the node impurity (giny) or entropy.
The risk of further training a DecisionTreeClassifier is that you are not sure that splits performed in the original tree are the best, since you have new splits capabilities with LGBM that might/should lead in better performance.
I would recommend you to retrain the model with LGBMClassifier only as it might be possible that splits will be different from the original scikit-learn Tree.
Exploring some classification models in Scikit learn I noticed that the scores I got for log loss and for ROC AUC were consistently lower while performing cross validation than while fitting and predicting on the whole training set (done to check for overfitting), thing that did not make sense to me.
Specifically, using cross_validate I set the scorings as ['neg_log_loss', 'roc_auc'] and while performing manual fitting and prediction on the training set I used the metric functions log_loss' and roc_auc_score.
To try to figure out what was happening, i wrote a code to perform the cross validation manually in order to be able to call the metric functions manually on the various folds and compare the results with the ones from cross_validate. As you can see below, I got different results even like this!
from sklearn.model_selection import StratifiedKFold
kf = KFold(n_splits=3, random_state=42, shuffle=True)
log_reg = LogisticRegression(max_iter=1000)
for train_index, test_index in kf.split(dataset, dataset_labels):
X_train, X_test = dataset[train_index], dataset[test_index]
y_train, y_test = dataset_labels_np[train_index], dataset_labels_np[test_index]
log_reg.fit(X_train, y_train)
pr = log_reg.predict(X_test)
ll = log_loss(y_test, pr)
print(ll)
from sklearn.model_selection import cross_val_score
cv_ll = cross_val_score(log_reg, dataset_prepared_stand, dataset_labels, scoring='neg_log_loss',
cv=KFold(n_splits=3, random_state=42, shuffle=True))
print(abs(cv_ll))
Outputs:
4.795481869275026
4.560119170517534
5.589818973403791
[0.409817 0.32309 0.398375]
The output running the same code for ROC AUC are:
0.8609669592272686
0.8678563239907938
0.8367147503682851
[0.925635 0.94032 0.910885]
To be sure to have written the code right, I also tried the code using 'accuracy' as scoring for cross validation and accuracy_score as metric function and the results are instead consistent:
0.8611584327086882
0.8679727427597955
0.838160136286201
[0.861158 0.867973 0.83816 ]
Can someone explain me why the results in the case of the log loss and the ROC AUC are different? Thanks!
Log-loss and auROC both need probability predictions, not the hard class predictions. So change
pr = log_reg.predict(X_test)
to
pr = log_reg.predict_proba(X_test)[:, 1]
(the subscripting is to grab the probabilities for the positive class, and assumes you're doing binary classification).
I have a very unbalanced dataset (5000 positive, 300000 negative). I am using sklearn RandomForestClassifier to try and predict the probability of the positive class. I have data for multiple years and one of the features I've engineered is the class in the previous year, so I am withholding the last year of the dataset to test on in addition to my test set from within the years I'm training on.
Here is what I've tried (and the result):
Upsampling with SMOTE and SMOTEENN (weird score distributions, see first pic, predicted probabilities for positive and negative class are both the same, i.e., the model predicts a very low probability for most of the positive class)
Downsampling to a balanced dataset (recall is ~0.80 for the test set, but 0.07 for the out-of-year test set from sheer number of total negatives in the unbalanced out of year test set, see second pic)
Leave it unbalanced (weird scoring distribution again, precision goes up to ~0.60 and recall falls to 0.05 and 0.10 for test and out-of-year test set)
XGBoost (slightly better recall on the out-of-year test set, 0.11)
What should I try next? I'd like to optimize for F1, as both false positives and false negatives are equally bad in my case. I would like to incorporate k-fold cross validation and have read I should do this before upsampling, a) should I do this/is it likely to help and b) how can I incorporate this into a pipeline similar to this:
from imblearn.pipeline import make_pipeline, Pipeline
clf_rf = RandomForestClassifier(n_estimators=25, random_state=1)
smote_enn = SMOTEENN(smote = sm)
kf = StratifiedKFold(n_splits=5)
pipeline = make_pipeline(??)
pipeline.fit(X_train, ytrain)
ypred = pipeline.predict(Xtest)
ypredooy = pipeline.predict(Xtestooy)
Upsampling with SMOTE and SMOTEENN : I am far from being an expert with those but by upsampling your dataset you might amplify existing noise which induce overfitting. This could explain the fact that your algorithm cannot correctly classify, thus giving the results in the first graph.
I found a little bit more info here and maybe how to improve your results:
https://sci2s.ugr.es/sites/default/files/ficherosPublicaciones/1773_ver14_ASOC_SMOTE_FRPS.pdf
When you downsample you seem to encounter the same overfitting problem as I understand it (at least for the target result of the previous year). It is hard to deduce the reason behind it without a view on the data though.
Your overfitting problem might come from the number of features you use that could add unnecessary noise. You might try to reduce the number of features you use and gradually increase it (using a RFE model). More info here:
https://machinelearningmastery.com/feature-selection-in-python-with-scikit-learn/
For the models you used, you mention Random Forest and XGBoost, but you did not mention having used simpler model. You could try simpler model and focus on you data engineering.
If you have not try it yet, maybe you could:
Downsample your data
Normalize all your data with a StandardScaler
Test "brute force" tuning of simple models such as Naive Bayes and Logistic Regression
# Define steps of the pipeline
steps = [('scaler', StandardScaler()),
('log_reg', LogisticRegression())]
pipeline = Pipeline(steps)
# Specify the hyperparameters
parameters = {'C':[1, 10, 100],
'penalty':['l1', 'l2']}
# Create train and test sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33,
random_state=42)
# Instantiate a GridSearchCV object: cv
cv = GridSearchCV(pipeline, param_grid=parameters)
# Fit to the training set
cv.fit(X_train, y_train)
Anyway, for your example the pipeline could be (I made it with Logistic Regression but you can change it with another ML algorithm and change the parameters grid consequently):
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import GridSearchCV, StratifiedKFold, cross_val_score
from imblearn.combine import SMOTEENN
from imblearn.over_sampling import SMOTE
from imblearn.pipeline import Pipeline
param_grid = {'C': [1, 10, 100]}
clf = LogisticRegression(solver='lbfgs', multi_class = 'auto')
sme = SMOTEENN(smote = SMOTE(k_neighbors = 2), random_state=42)
grid = GridSearchCV(estimator=clf, param_grid = param_grid, score = "f1")
pipeline = Pipeline([('scale', StandardScaler()),
('SMOTEENN', sme),
('grid', grid)])
cv = StratifiedKFold(n_splits = 4, random_state=42)
score = cross_val_score(pipeline, X, y, cv=cv)
I hope this may help you.
(edit: I added score = "f1" in the GridSearchCV)
I'm using RandomizedSearchCV to get the best parameters with a 10-fold cross-validation and 100 iterations. This works well. But now I would like to also get the probabilities of each predicted test data point (like predict_proba) from the best performing model.
How can this be done?
I see two options. First, perhaps it is possible to get these probabilities directly from the RandomizedSearchCV or second, getting the best parameters from RandomizedSearchCV and then doing again a 10-fold cross-validation (with the same seed so that I get the same splits) with this best parameters.
Edit: Is the following code correct to get the probabilities of the best performing model? X is the training data and y are the labels and model is my RandomizedSearchCV containing a Pipeline with imputing missing values, standardization and SVM.
cv_outer = StratifiedKFold(n_splits=10, shuffle=True, random_state=0)
y_prob = np.empty([y.size, nrClasses]) * np.nan
best_model = model.fit(X, y).best_estimator_
for train, test in cv_outer.split(X, y):
probas_ = best_model.fit(X[train], y[train]).predict_proba(X[test])
y_prob[test] = probas_
If I understood it right, you would like to get the individual scores of every sample in your test split for the case with the highest CV score. If that is the case, you have to use one of those CV generators which give you control over split indices, such as those here: http://scikit-learn.org/stable/tutorial/statistical_inference/model_selection.html#cross-validation-generators
If you want to calculate scores of a new test sample with the best performing model, the predict_proba() function of RandomizedSearchCV would suffice, given that your underlying model supports it.
Example:
import numpy
skf = StratifiedKFold(n_splits=10, random_state=0, shuffle=True)
scores = cross_val_score(svc, X, y, cv=skf, n_jobs=-1)
max_score_split = numpy.argmax(scores)
Now that you know that your best model happens at max_score_split, you can get that split yourself and fit your model with it.
train_indices, test_indices = k_fold.split(X)[max_score_split]
X_train = X[train_indices]
y_train = y[train_indices]
X_test = X[test_indices]
y_test = y[test_indices]
model.fit(X_train, y_train) # this is your model object that should have been created before
And finally get your predictions by:
model.predict_proba(X_test)
I haven't tested the code myself but should work with minor modifications.
You need to look in cv_results_ this will give you the scores, and mean scores for all of your folds, along with a mean, fitting time etc...
If you want to predict_proba() for each of the iterations, the way to do this would be to loop through the params given in cv_results_, re-fit the model for each of then, then predict the probabilities, as the individual models are not cached anywhere, as far as I know.
best_params_ will give you the best fit parameters, for if you want to train a model just using the best parameters next time.
See cv_results_ in the information page http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html