How much time the training costs? about hed HOT 2 CLOSED

Hi Zizhao,

We updated our code base to the newer version of caffe and use mini-batch size 1 for full resolution images, in this way we are able to boost our performance to 0.790 ODS. If you resize the images to 400x400 as in the paper and set the mini-batch size to 10, then it will be much faster. From my experiment log 7 hours will be enough for 10,000 iterations in the previous setting.

from hed.

Hi Saining,
Thanks for your answer.
So if I understand correctly, you mean you use image size 400x400 with a mini-batch size of 10 will be much faster than use full size and mini-batch size of 1?
Why? Large Mini-batch should slow down the time of one iteration, right?

I am still do not understand why my iteration is so slow, with the data you provided and the parameters you set in the code (mini-batch is 1), the one iteration time is 1 minutes. So totally 10,000 mins are needed. I am using Tesla K40c. Do you have any idea about this?

By the way, could you show me your loss plot? It is interesting to find the loss is high and vibrate but the results are already visually good after a few thousands of iterations.

Thank you so much for your answers.

from hed.