How can I assign a class_weight in Keras in a simple way?

Question

Can anyone tell me what is the simplest way to apply class_weight in Keras when the dataset is unbalanced please?

I only have two classes in my target.

Thanks.

petezurich · Answer 1 · 2018-09-12T06:42:05.650

45

The class_weight parameter of the fit() function is a dictionary mapping classes to a weight value.

Lets say you have 500 samples of class 0 and 1500 samples of class 1 than you feed in class_weight = {0:3 , 1:1}. That gives class 0 three times the weight of class 1.

train_generator.classes gives you the proper class names for your weighting.

If you want to calculate this programmatically you can use scikit-learn´s sklearn.utils.compute_class_weight().

The function looks at the distribution of labels and produces weights to equally penalize under or over-represented classes in the training set.

See also this useful thread here: https://github.com/fchollet/keras/issues/1875

And this thread might also be of help: Is it possible to automatically infer the class_weight from flow_from_directory in Keras?

edited Sep 12 '18 at 06:42

answered Jun 23 '17 at 12:48

petezurich

9,280
9
43
57

2

Finally a clear example! Nowhere else does it seem to say whether the weighting scales up or down – Tom Walker Oct 22 '17 at 13:34
2

This should be the top answer - it answers what the question was asking (about Keras, not scikit-learn) and also provides docs & explanation. – MLavrentyev Jul 27 '19 at 17:28

Johnny Hsieh · Answer 2 · 2017-07-26T11:34:26.190

Using class_weight from sklearn kit.

Im also using this method to deal with the imbalance data

from sklearn.utils import class_weight
class_weight = class_weight.compute_class_weight('balanced'
                                               ,np.unique(Y_train)
                                               ,Y_train)

then model.fit

Classifier.fit(train_X,train_Y,batch_size = 100, epochs = 10
           ,validation_data= (test_X,test_Y),class_weight = class_weight )

score 6 · Answer 3 · answered Mar 07 '19 at 12:06

6

1- Define a dictionary with your labels and their associated weights

class_weight = {0: 0.1,
                1: 1.,
                2: 2.}

2- Feed the dictionary as a parameter:

model.fit(X_train, Y_train, batch_size = 100, epochs = 10, class_weight=class_weight)

answered Mar 07 '19 at 12:06

javac

2,819
1
20
22

classes are named from 0 or 1? – mahdi Jun 16 '20 at 17:18

score 4 · Answer 4 · answered Jun 23 '17 at 08:18

4

Are you asking about the right weighting to apply or how to do that in the code? The code is simple:

class_weights = {}
    for i in range(2):
        class_weights[i] = your_weight

and then you pass the argument class_weight=class_weights in model.fit.

The right weighting to use would be some sort of inverse frequency; you can also do a bit of trial and error.

answered Jun 23 '17 at 08:18

Michele Tonutti

4,298
1
21
22

Thanks, that's what I was looking for – Javi Jun 23 '17 at 10:27

score 2 · Answer 5 · answered Oct 07 '19 at 05:48

2

class weights takes a dictionary type.

from collections import Counter
itemCt = Counter(trainGen.classes)
maxCt = float(max(itemCt.values()))
cw = {clsID : maxCt/numImg for clsID, numImg in itemCt.items()}

answered Oct 07 '19 at 05:48

Allie

101
1
4

How can I assign a class_weight in Keras in a simple way?

5 Answers5

Linked