Where communities thrive

  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
Repo info
Priyam Mehta
Hello, everyone, I have a usage question, which I have posted on SO, here's the link: https://stackoverflow.com/questions/65652054/not-able-to-feed-the-combined-smote-randomundersampler-pipeline-into-the-main, can someone can help me with this. Thank you.
Priyam Mehta
TypeError: All intermediate steps of the chain should be estimators that implement fit and transform or fit_resample. 'Pipeline(steps=[('smote', SMOTE(n_jobs=-1, random_state=42)), ('under', RandomUnderSampler(random_state=42))])' implements both) Also, can someone explain what this error means? The Pipeline only exposes fit and fit_resample methods, since, transform is not being implemented, the first condition is not met and the second one about fit_resample is being met. Then, shouldn't this work? Thank you.
Guillaume Lemaitre
Can you post the entire traceback to check which transformer/resampler is raising the conditionn
oh I see
you smote_pipeline implement fit_resample and transform as well
Basically you can use an imbalanced learn pipeline within another pipeline (we did not think about it) because you have an ambuiguity
the pipeline does not know if it should call fit_resample or fit/transform
In your case, you should be able to solve this issue using a flat pipeline
Main_Pipeline = imb_Pipeline([
     ('feature_handler', FeatureTransformer(list(pearson_feature_vector.index))),
     ('smote', SMOTE()),
     ('random_under_sampler', RandomUnderSampler()),
     ('scaler', StandardScaler()),
     ('pca', PCA(n_components=0.99)),
     ('model', LogisticRegression(max_iter=1750)),
It should be the equivalent
Priyam Mehta

Please correct me if my understanding is lacking.
So, when I call fit to the Main_Pipeline , since smote_pipeline as a fit present, it is assumed that transform is also present, actually it doesn't, I tried to call transform and got an error:

AttributeError: 'RandomUnderSampler' object has no attribute 'transform'

Pipeline code:

Smote_Under_pipeline = imb_Pipeline([
    ('smote', SMOTE(random_state=rnd_state, n_jobs=-1)),
    ('under', RandomUnderSampler(random_state=rnd_state)),

, and accordingly because of assumption fit/transform and fit_resample both become available. This causes ambuiguity and the code blows up?

Guillaume Lemaitre
Yes the sampler does not implement transform but the pipeline does
and try to call the transform of the underlying estimator
you need to do hasattr(smote_pipeline, "transform")
and you will see that this is true
yes there is an ambuiguity because we don't know if you would like to call transform or fit_resample
Priyam Mehta
Cool , thanks!!
Soledad Galli
Would it be possible to extend the functionality of the BalancedBaggingClassifier and BalancedRandomForests to other sampling techniques (eg, SMOTE) by allowing the user to enter the over or under-sampling method as parameter instead of hard-coding RandomUnderSampler?
Guillaume Lemaitre
If I recall properly, they leverage sample_weight and therefore you would need to have a Sampler that store indices to build the sample_weight vector
The second consideration is computational performance
Random US/OS are not costly
adding sampler based on k-NN will not scale
and in practice, I am tending to think that RUS and ROS would be enough to alleviate the issue with an ensemble learner.
Soledad Galli
makes sense, thank you!
Hanchung Lee


I am getting error loading a trained imblearn.pipeline Pipeline saved by joblib. Getting this error message:

ModuleNotFoundError: No module named 'imblearn.over_sampling._smote.base'; 'imblearn.over_sampling._smote' is not a package

The trained pipeline was saved via joblib.dump(pipeline, 'filename.joblib'). Any tips as to where the saving and loading process went wrong?

Guillaume Lemaitre
make sure that the version installed is the same as the version used to pickle

Hello everyone.
I have an usage question about EasyEnsembleClassifier. I have a dataset which has 450.000 data inputs with 13 columns(12 features, 1 target). My dataset is imbalanced (1:50) so I decided to use EasyEnsembleClassifier. I realized that all the subsets are exactly same for all the estimators.
I found this issue which is similar to my problem: scikit-learn-contrib/imbalanced-learn#116
In theory classifier method should create subsets for each estimators. These subsets should have all minority class samples and select same number of samples from majority class. In my case I should have roughly 18000 samples in each subset (I have roughly 9000 samples in minority class). However when I use "estimatorssamples" method it seems like output arrays for my estimators are exactly same and all of them have size of complete training set(80% of my dataset). So I decided to make a test:
import numpy as np
from sklearn.datasets import make_classification
from imblearn.ensemble import EasyEnsembleClassifier

X, y = make_classification(n_classes=2, class_sep=2, weights=[0.3, 0.7],
n_informative=3, n_redundant=1, flip_y=0,
n_features=20, n_clusters_per_class=1,
n_samples=10, random_state=1)

clf = EasyEnsembleClassifier(n_estimators=5, n_jobs=-1)

clf.fit(X, y)

arr = clf.estimatorssamples

[array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]),
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]),
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]),
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]),
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])]

What am I doing wrong here? Obviously I am missing a point.

Guillaume Lemaitre
To check the sample used by each estimator, you should use
In [13]: for est in clf.estimators_:
    ...:     print(est[0].sample_indices_)
[4 6 7 0 9 2]
[4 6 7 5 8 3]
[4 6 7 1 2 5]
[4 6 7 3 1 5]
[4 6 7 3 5 2]
I am not sure what estimator_samples_ is reporting. It might be a bug then
This weird that we don't document it
oh I see, we should add it in the documentation
Guillaume Lemaitre
estimator_samples_ gives the samples dispatch to the first estimator that later on will undersample
This attribute exist because we inherit from the BaggingClassifier from scikit-learn

The code you provided works fine with my generated dataset but when I use it on my real dataset this is what I get:

clf = EasyEnsembleClassifier(n_estimators=5, n_jobs=-1, sampling_strategy = 1.0)
clf.fit(X_train, y_train)
for est in clf.estimators_:

[279507 240017  23859 ...  94249  87790 120830]
[277730  75855  70104 ... 341432 318980 130029]
[166614    207  72374 ...  93568  76905 142951]
[304630  28272 143132 ... 159062 264981  41332]
[ 35943 358917  68200 ... 121931 209190 284075]

Is this a normal result? I would expect first three indices in each row to be the same. I mean; all of the samples that belong to the minority class are being used in all subsets. I am not saying this is wrong. I am just asking if this is normal?

Code runs just fine now. Thanks for help @glemaitre
Guillaume Lemaitre
It is possible that they are randomized
Ok got it.
What could be the reason of same prediction results for my dataset which I mentioned above no matter what I choose for the parameter "n_estimators"? Choosing 1000 or 1 no differs.
I'm having a weird problem with random oversampler. Running two python scripts on what is near identical data from two different sources. Getting a value error: can convert string to float for my one text-based feature. The feature formatting and number of unique values are the same in both sets. In one script it works and in one script I get the error. This apparently was an issue years ago - imbalanced learn didn't support text of pandas dataframes. I believe that has now been fixed (evidently since one of my scripts works). Guidance on how to handle? As mentioned, I have confirmed that formatting and values on the problem feature are the same. Thanks
Guillaume Lemaitre
I cannot say without a code snippet or the head of the dataset
but normally the RandomOverSampler will get a dataset and will not be an issue to have non-numerical data inside
Thanks for your reply. I tend to agree that it is something in my data vs. something in the library since it runs properly in one script.
Here is another wildcard question - I'll check myself but if someone has the answer off the top of the head it will save me some work - when oversampling and the new observations are created, are they appended to the bottom/end of the dataframe/array or are they placed adjacent to the observation from which it was created?
There we copy the original dataset
and then append new samples for each class

I'm using a multiclass dataset (cic-ids-2017), the target column is categorical (more than 4 classes), I used {pd.get_dummies} for One Hot Encoding. The dataset is very imbalanced, and when I tried to oversampling it using SMOTE method, doesn't work, I also tried to include them into a pipeline, but the pipeline cannot support get_dummies, I replaced it by OneHotEncoder, unfortunately, still not working :

X = dataset.drop(['Label'],1)
y = dataset.Label
steps = [('onehot', OneHotEncoder(), ('smt', SMOTE())]
pipeline = Pipeline(steps=steps)
X, y = pipeline.fit_resample(X, y)
Is there any proposition ?

My correlation matrix does not changed after using SMOTE, what could be the cause ?
I'm using resting-state fMRI correlation matrices, which are 4D, and I want to use SMOTE+ENN but it only allows me to use 2D data... How can I adress this problem without losing information from my original data?