training.clj - Githubissues

hellonico / fruits-classification

Using Artifical Intelligence to perform classification of fruits

1 stars 0 forks source link

training.clj #3

Open rafariva opened 6 years ago

rafariva commented 6 years ago

when run this, get severals errors, could you help me, why? i am trying on windows, and on ubuntu... diferent errores on each one (how to fix on any or both?), i have no clue. i am new in clojure. thanks

on windows:

on ubuntu:

hellonico commented 6 years ago

https://github.com/originrose/cortex/issues/184

I fixed the dependencies and the change of code. But I am getting very bad accuracy at this stage:

Classification accuracy: 0.1829470198675497 Classification accuracy: 0.1771523178807947 Classification accuracy: 0.1697019867549669

rafariva commented 6 years ago

Maybe is because the preparation.clj, reported in the other issue (training folder and test folder has same images)

hellonico commented 6 years ago

reversely, it would get a higher classification accuracy at the training stage. (since the training images and the validation images would be the same).

rafariva commented 6 years ago

You are right. Never the less, both folder shouldnt have the same, right?

hellonico commented 6 years ago

It's definitely better for training if they are not the same, but it's not an absolute requirement.

What happens is:

with the training images, you tell the network an apple is an apple
with the test image, you ask the network what it thinks the image is. You know it's an apple, but the only way the network can guess it is by going through its internal layers.

So if the test image is an image the network has already seen, it's a bit like cheating, resulting in a better, but faked, accuracy, as opposed to an image the network has never seen.

rafariva commented 6 years ago

Ohh ok. Gateyou. So, if it is "cheating" why so low accuaracy? Any idea?