search on cifar10 cost about 6 hours, using 1080Ti

Open hx621 opened this issue 4 years ago • 4 comments

May 13 '21 12:05 hx621

Hi, did you install PyTorch (and other dependencies) correctly? This implementation is slightly slower than the one used in the paper because of type checking, but the training requires at most 30 mins.

May 13 '21 12:05 moskomule

Hi, did you install PyTorch (and other dependencies) correctly? This implementation is slightly slower than the one used in the paper because of type checking, but the training requires at most 30 mins.

i use python 3.7.4, does it matter ?

May 13 '21 12:05 hx621

It is not supported.

May 13 '21 13:05 moskomule

@hx621 did you use the subset of the training set for training the sub-policies? With 4000 images, I trained 20 sub-policies with 4 operations for 200 epochs and it took me 36 minutes using a single GPU. I used Python 3.7, Pytorch 1.13, CUDA 11.3.1 and Homura.

Dec 16 '23 14:12 saigontrade88