search

idiap / attention-sampling

This Python package enables the training and inference of deep learning models for very large data, such as megapixel images, using attention-sampling
Other
98 stars 18 forks source link

Batch size for all the experiments in the papaer #8

Closed timqqt closed 5 years ago

timqqt commented 5 years ago

Hey, I just want to repeat your work. I saw that your paper implied that the batch size for all experiments should be 1. However, I find that if I set the batch size to 1, I cannot get the same error as your experiments (about 10 times difference). But if I set the batch size to 32, I got decent result.

I would very appreciate your help to explain details in your experiments!

angeloskath commented 5 years ago

Hi, batch size of 1 is used to report memory requirements for all methods. The larger the batch size the better, since you want to minimize the variance of the gradients.

In the example scripts you can see the default batch size is 32 for the speed limits and 128 for the mnist experiment.

Let me know if I can help with anything more.

Cheers, Angelos

angeloskath commented 5 years ago

I am closing the issue but feel free to reopen it (or another one) if needed.

Angelos