Question about accuracy evaluation?

[happsky]

For compute_accuracies.py, I just want to make sure, is the
final result = total match found for synthesized / total images into consideration?

[kregmi]

It's the prediction accuracy between real and generated images. We use the pre-trained Alexnet model to obtain annotations for real images and class predictions for generated images. If the model predicts the real and synthesized image pair to be of same class, then its a hit otherwise its a miss.
final result = total hits/total images into consideration.

[happsky]

All evaluation metrics are too slow, do you have any ideas to speed them?

[kregmi]

Sadly they are slow. A possible way to make it faster is to compute features in batches, currently it is written to work on single image at a time.

[happsky]

I also noted that you did not use GPUs, is there any possibility to use GPU while evaluation.

[kregmi]

Not sure, that part of code was borrowed from some other people, I didn't spend time on making it quicker.

[happsky]

@kregmi for the result on the image, am I correct?
top-1, all = 1847/20826
top-1, >0.5 = 229/2470
top-5, all = 7194/20826
top-5, >0.5 = 883/2470