How to find wrong prediction cases in test set (CNNs using Keras)

Question

I'm using MNIST example with 60000 training image and 10000 testing image. How do I find which of the 10000 testing image that has an incorrect classification/prediction?

S.Mohsen sh · Accepted Answer · 2017-10-05T18:33:35.437

24

Simply use model.predict_classes() and compare the output with true labes. i.e:

incorrects = np.nonzero(model.predict_class(X_test).reshape((-1,)) != y_test)

to get indices of incorrect predictions

edited Oct 05 '17 at 18:33

answered Sep 03 '16 at 06:29

S.Mohsen sh

2,028
3
21
32

How can I generate a sub-folder with the wrong classified cases inside each class folder?? – Joaquín Fernández Feb 25 '18 at 20:02
if x, y are numpy check out the [answer](https://stackoverflow.com/a/28440249/13717851) – Sayan Dey Aug 23 '20 at 13:12

SnigA · Answer 2 · 2020-10-29T14:52:32.773

2

Editing as was not clear earlier

To identify the image files that are wrongly classified, you can use:

fnames = test_generator.filenames ## fnames is all the filenames/samples used in testing
errors = np.where(y_pred != test_generator.classes)[0] ## misclassifications done on the test data where y_pred is the predicted values
for i in errors:
    print(fnames[i])

edited Oct 29 '20 at 14:52

answered Apr 09 '20 at 09:11

SnigA

147
10

What are all the variables? – fdermishin Oct 18 '20 at 17:12
How can I print the actual class and predicted class as well along with the image name? – Hardik Raval Apr 23 '21 at 06:38
when I print fnames, it printed same file name multiple times. Any update? – Hello-experts Mar 08 '22 at 07:55

How to find wrong prediction cases in test set (CNNs using Keras)

2 Answers2