this pull request was from Kisuk, I am redirecting this request to development branch to have some tests.
Layer-wise FFT optimization takes quite long if the network being optimized is large. The optimized results are saved as fft=1 or fft=0 options in filter edges, but omitted from loading. They should be loaded from the saved instance, so that time-consuming FFT optimization need not be repeated.
this pull request was from Kisuk, I am redirecting this request to
developmentbranch to have some tests.Layer-wise FFT optimization takes quite long if the network being optimized is large. The optimized results are saved as fft=1 or fft=0 options in filter edges, but omitted from loading. They should be loaded from the saved instance, so that time-consuming FFT optimization need not be repeated.