The MNIST_DCGAN learning is too slow

[milk-bottle-liyu]

Sir, as u said the avg time for an epoch is around 180s, while on my server, it shows:
[1/20] - ptime: 372.38, loss_d: 0.597, loss_g: 5.759

My environment is:
ubuntu 16.04+cuda8.0+cudnn 6+ pytorch 0.2 +Titan XP

I also set the worker_num for train data loader to 2, so it shouldn't be a problem of IO.
Do u have any idea of what's going wrong , Sir?

[znxlwm]

Hi,

I don't know exactly why learning time is different.

I ran my code again using these settings - 64 x 64 MNIST 60,000 images, 128 batch size, 'd' = 128 and got a similar result.

Check your data, batch size, and dimension parameter of network ( 'd' value in my network) etc.

[milk-bottle-liyu]

Dear Kang: My apologize, I didnt check my email in time. I finally found out that, someone else was also using GPU, so the GPU is busy, thats the reaso. PS: Im new with pytorch. I used caffe. I was wondering why PyTorch's Volatile GPU-Util is quite high as by using caffe I can only achieve around 70%. But when using PyTorch, it is quite easy to achieve 90%+. Do you have any idea with that? 2017-11-23 10:26 GMT+08:00 Hyeonwoo Kang <[email protected]>:

…

Hi, I don't know exactly why learning time is different. I ran my code again using these settings - 64 x 64 MNIST 60,000 images, 128 batch size, 'd' = 128 and got a similar result. Check your data, batch size, and dimension parameter of network ( 'd' value in my network) etc. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQ_3nuZJS1yU1KZGmaTKCzJca-iYsq3tks5s5NflgaJpZM4Qm8iJ> .