I'm also interested in how to do that efficiently. Currently, when I want to quantize weights, I pull them from GPU using get_value(), quantize them in Python, and them import them back to GPU with set_value(). But, of course, this is very slow. For binary quantization, I can use Theano function:
Wb = T.cast(T.switch(W,1,-1), theano.config.floatX) Any suggestions? On Sunday, July 10, 2016 at 8:45:18 PM UTC-7, Kan Kawabata wrote: > > Hello, I am trying to study the effect of quantization error in the input > and was wondering if there are any theano function that allows me to round > a tensor value to its x-bit representation (e.g. round from float64 to > float16 representation but keep the tensor as float64 type). I'm not sure > what is the best way to go about this in theano and would appreciate any > insight. > > Thank you, > > Kan Kawabata > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
