New models required for low precision?

[eplondke]

I think in order to use the low precision library, I get gemmlowp, compile with LOWP=1, and then run with -a 7. All that seems to work great, and I get different performance on x86.

But using the demo teradeep network, the answers are all crazy (most results are .8 or higher, lots over .99, and the #1 pick for my picture of a panda is not panda, but "candles"). Do I need to convert the network to a fixed point format to use the low precision library?

[mvitez]

8 bits precision is too low for the networks we currently have. After a few stages, the loss in precision will be so high that the results will diverge quickly. We currently don't have any network that will work in this mode, so this feature is experimental.