0

I tried the keras tutorial that I found here...

https://github.com/eijaz1/Building-a-CNN-in-Keras-Tutorial/blob/master/cnn_tutorial.ipynb

Everything worked fine till line 10. But I am not able to predict correctly. I get the results like this...

model.predict(X_test[:4])

array([[0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0.]], dtype=float32)

The expected results as per the tutorial are:

array([[1.6117248e-09, 8.6684462e-16, 6.8095707e-10, 1.5486043e-08,
        6.2878847e-14, 1.2934288e-15, 1.1453808e-16, 9.9999928e-01,
        1.0626109e-08, 6.9729606e-07],
       [1.3555871e-07, 2.6465393e-06, 9.9999511e-01, 2.0351818e-08,
        1.9796262e-11, 1.6996018e-12, 2.1163373e-06, 1.2008194e-17,
        4.8792381e-10, 2.6086671e-13],
       [6.7238901e-08, 9.9785548e-01, 1.9031411e-04, 3.9194603e-08,
        1.2894072e-04, 1.5791730e-06, 1.2754040e-06, 4.1349044e-09,
        1.8221687e-03, 5.5910935e-08],
       [9.9999356e-01, 1.6909821e-12, 8.2496926e-10, 1.7359107e-11,
        1.7359230e-12, 1.8865266e-13, 6.4659162e-06, 2.3738855e-11,
        1.1319052e-08, 2.6948474e-08]], dtype=float32)

I am using keras version 2.2.2 if that matters.

Update:

While training the model, I am getting pretty low accuracy compared to that tutorial. 

model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=3)

Train on 60000 samples, validate on 10000 samples
Epoch 1/3
60000/60000 [==============================] - 75s 1ms/step - loss: 14.3141 - acc: 0.1118 - val_loss: 14.2677 - val_acc: 0.1148
Epoch 2/3
60000/60000 [==============================] - 74s 1ms/step - loss: 14.3741 - acc: 0.1082 - val_loss: 14.4692 - val_acc: 0.1023
Epoch 3/3
60000/60000 [==============================] - 134s 2ms/step - loss: 14.3691 - acc: 0.1085 - val_loss: 14.3483 - val_acc: 0.1098

How do I improve the accuracy? I am using exactly same code as shown in the tutorial.

Here is the output that I get even if I use the exactly same code:

https://github.com/shantanuo/Building-a-CNN-in-Keras-Tutorial/blob/master/cnn_tutorial_mismatch.ipynb

Maven Carvalho
  • 319
  • 1
  • 5
  • 14
shantanuo
  • 31,689
  • 78
  • 245
  • 403
  • The tutorial shows a black and white image for plt.imshow(X_train[0]) while I get color image showing digit 5. Has something changed for mnist.load_data() function of keras? – shantanuo Oct 17 '18 at 10:50
  • I tried both notebooks and I cannot reproduce your result. Did you try multiple runs to check its not an issues of just random weight initialization? – Dr. Snoopy Oct 17 '18 at 11:26
  • After running it I do see the image with color as well with fairly low accuracy, ~0.16, 0.18, 0.19. – Hatt Oct 17 '18 at 13:28
  • FWIW, I ran this again without changing anything and Acc went to .5, and a third time, changing only plt.imshow(X_train[0]) to plt.imshow(X_train[0],cmap='binary') (which I don't think actually did anything wrt the rest of the model) and the accuracy went to .98 and reproduced the results. – Hatt Oct 17 '18 at 17:50

0 Answers0