Hello Folks!
To introduce myself first, my name is Mirko, and in my spare time I like to learn Deep Learning.
I’m using Laurence Moroney book.
While testing jupyter notebook example “transfer learning cats and dogs” I can not get satisfactory results of cats vs dogs recognition after feeding the model with real images from the net.
Eventhough I have relatively good training and validation accuracy:
225/225 [==============================] - 185s 761ms/step - loss: 0.2232 - acc: 0.9110 - val_loss: 0.0975 - val_acc: 0.9668
Epoch 2/3
225/225 [==============================] - 172s 765ms/step - loss: 0.1491 - acc: 0.9391 - val_loss: 0.0884 - val_acc: 0.9704
Epoch 3/3
225/225 [==============================] - 166s 738ms/step - loss: 0.1391 - acc: 0.9422 - val_loss: 0.0694 - val_acc: 0.9776
and here are the results, you can see by the name of the files what images are cats, what dogs, and the model keeps choosing dogs:
…
[[1.]]
[1.]
dogsmall4.jpg is a dog
[[1.]]
[1.]
dogsmall3.jpg is a dog
[[1.]]
[1.]
dogsmall2.jpg is a dog
[[1.]]
[1.]
dogsmall1.jpg is a dog
[[1.]]
[1.]
catsmall4.jpg is a dog
[[1.]]
[1.]
catsmall3.jpeg is a dog
[[1.]]
[1.]
catsmall2.jpg is a dog
[[1.]]
[1.]
catsmall1.jpg is a dog
I have tried to use smaller square pictures (as input is 150 x 150) but it doesnt get better.
What could be the issue?
Regards,
Mirko