Tune multiple models simultaneously with GridSearchCV

You can tune 2+ models using the same grid search! Here's how:
1. Create multiple parameter dictionaries
2. Specify the model within each dictionary
3. Put the dictionaries in a list

👉 New tips every TUESDAY and THURSDAY! 👈

=== WANT TO GET BETTER AT MACHINE LEARNING? ===

3) LET'S CONNECT!

[Lumcoin]

OMG, a classic "YouTuber explains how to add 1+1" BUT NO ONE ELSE SAYS HOW TO ADD 1+1!!! Thanks a bunch, you most likely saved me two hours of frustrating trial and error.

[hemanthkotagiri]

I was literally on my hunt to see if I can train 4 different models with different parameters all at once and here you upload this. Perfect timing! Thanks, man! Love your content!

[wayfaring.stranger]

You're a good boy; this has streamlined my PhD's research.

[gauravmalik]

Amazing video, there are very few videos on these such unique topics on YouTube.
Had one doubt, didn't understood the placeholder part at 2:15.

[amarsharma]

Thanks for saving our time, i used to do loops

[AhmedThahir]

Amazing video!

Does this work for RandomizedSearchCV as well?

[marinapachecovillaschi]

Nice tip!!
As I'm solving a multi-output classification problem I'm using MultiOutputClassifier() for that and I think that's what's messing my code up when trying to run this solution. It looks something like this:

pipeline = Pipeline([
('vect', CountVectorizer(tokenizer=tokenize)),
('tfidf', TfidfTransformer()),
('classifier',
])

# parameter dict for logistic regression
params_lr = {
'vect__decode_error' : ['strict', 'ignore', 'replace'],
'tfidf__norm' : ['l1', 'l2'],
: ['l1', 'l2'],
'classifier__estimator__C' : [0.1, 1, 10],
'classifier__estimator' : [lr_clf]
}

# other dicts for different models

# list of parameters dicts
parameters = [params_lr, params_svc, params_rf]

cv = GridSearchCV(pipeline, param_grid = parameters, n_jobs=-1)
cv.fit(X_train, y_train)



Any tips on this? I think the gridsearch doesn't understand the MultiOutputClassifier.
Thanks in advance!!

[bakingbaker]

I don't doubt this works, but it seems a bit odd to specify an instantiated classifier in the pipeline only to override it with another instantiated classifier. Is there a way to make the classifier in the pipeline a generic placeholder?