the problem is that additional training of a pretrained model on a relatively simple vision-type model (not a complex model like resnet, but it has some convolutions), is converging on CPU but not GPU — validation does not converge, anyway.
is there an anaconda or pip package without cudnn to try without a rebuild? i don’t think rebuild is an option. On Thu, Jul 11, 2019 at 10:22 AM kellen sunderland < [email protected]> wrote: > Having runtime loadable / plugable operators might help with this. > > On Thu, Jul 11, 2019 at 10:20 AM kellen sunderland < > [email protected]> wrote: > > > Once it's compiled the forward / backward, etc kernel implementations are > > hard coded to use cuDNN. In theory we could support raw CUDA in addition > > to cuDNN but the additional CUDA kernel code would bloat the binary (it > > targets several GPU types). > > > > On Thu, Jul 11, 2019 at 9:36 AM Chris Olivier <[email protected]> > > wrote: > > > >> Is there an environment variable or some other way to not use CUDNN in > the > >> anaconda distribution of mxnet? > >> > > >
