I want to use sklearn.metrics.confusion_matrix(y_true, y_pred) to create a confusion matrix for a keras model.
After training a model I can use predict_generator(generator) to get predictions for a test dataset, which gives me y_pred. How can I get the corresponding true labels, y_true from a data generator?
generator.classes will give you observed values in sparse format. You probably need it in dense (i.e., one-hot encoded format). You could get that with:
import pandas as pd
pd.get_dummies(pd.Series(generator.classes)).to_dense()
NOTE though: you must set the generator's shuffle attribute to False before generating the predictions and fetching the observed classes, otherwise your predictions and observations will not line up!
After creating a data generator, either your own or the built in ImageDataGenerator, use your trained model to make predictions:
true_labels = data_generator.classes
predictions = model.predict_generator(data_generator)
sklearn's confusion matrix expects a 1-d array of labels, so you have to convert your predictions using np.argmax()
y_true = true_labels
y_pred = np.array([np.argmax(x) for x in predictions])
Then you can use those variables directly in the confusion_matrix function
cm = sklearn.metrics.confusion_matrix(y_true, y_pred)
And you can plot it using the example plot_confusion_matrix() function found here:
https://scikit-learn.org/stable/auto_examples/model_selection/plot_confusion_matrix.html
