Ideal is transparent CPU/GPU Tensor library +fwd/bwd API over Tensor ops in scripting lang. No protos. No symbolics. i.e. Torch but Python
-
@karpathy@traboukos it's a tradeoff :) for me, automatic diff. and optimization more than make up for the less straightforward debugging
@sedielem @karpathy @traboukos & the auto diff. is a godsent. Allows you to think in symbolic terms (math) rather than in terms of data flow
