search

chengyangfu / caffe

Caffe: a fast open framework for deep learning.
http://caffe.berkeleyvision.org/
Other
169 stars 47 forks source link

513*513train need 22GB memory ? #12

Closed teleger closed 6 years ago

teleger commented 6 years ago

i need. low memory .. so ... config file. set. or add pruneing file of model.. thanks.. from Chinese.

chengyangfu commented 6 years ago

Yes, that is a problem I faced during training. I am really lucky to access some m40/p40 machines which contain 24 GB. I would suggest to use smaller batch size, freeze the batch normalization layers in Resnet, and don't use the batch normalization layers in the newly added layers.