Time spending on training

[knazeri]

Time spending on training

[knazeri]

@cmyyy We have trained the model mostly on Nvidia Titan XP and Titan V GPUs, however, GTX 1080s also work pretty fine. Smaller datasets, like celebA and Paris StreetView converge in 2-3 days, large datasets like places2 take more than two weeks to converge!
Once you have a pre-trained model, you do not need to train from scratch, you can load places2 weights and fine tune the training on your own dataset.

[cmyyy]

What if i want to make some modifications to the model ? In this case , i have to train from scratch, right?

[knazeri]

If you modify the network in any way, you need to train it from scratch.

[cmyyy]

You mentioned ' Each image is shown 10 times in total' in visual turing tests,was it possible that in one test, an image appeared more than one time? Are the participants college students or people on AMT?

[knazeri]

We use AMT and made sure that an image is only shown only once to each participant. The size of the validation set for Places2 dataset is roughly 36,000 images, we randomly selected 300 images without replacement which is ok since we are following the 10% rule! Then we take samples of n=100 images and show them to testers; The sampling distribution of the sample mean is also roughly normal (n > 30) and the statistics are shown using 95% confidence interval!

[cmyyy]

Thanks for your helpful answer!