Not utilizing all resources

[kinhunt]

Cool project. I have looking for this for months.

I tried to train (retrain.sh) 6000 images with 20 labels. During the training process, I found that

1. the project is not using much of the CPUs nor GPUs.
2. training precision raised from 73% to 94% in just 2 steps.
3. training precision stabilized in 94% during the following 5000 steps.
4. labels with most images always have high scores in prediction

Am I doing something wrong? thank you.

[BartyzalRadek]

Hi,

1. I just modified code from Google's tutorial about retraining so I am afraid I can't help you with that.
   I didn't test it on GPU because the retraining is pretty fast...

2-3. That is definitely weird, I just ran retraining on 5 classes - each having around 400 images and the training accuracy looks like this:

4. Its always a good idea to have ~same number of training examples in each class - I have 400-500 images per class and the network works correctly as far as I can tell.

Try cloning the repo now, I fixed some issues, maybe you were using the old version?
Or have incorrectly generated label_files for your images?

[harelfar2]

@kinhunt
I'm experiencing the exact same problems- not utilizing the GPU and too good accuracy.
Did you to figure it out?

@BartyzalRadek
Don't you find it a little suspicious that the accuracy raises so quickly?
And thank you for this project, it really helped me.