If you build a model that correspond to many RNN at once, then yes, this should be faster.
But your model won't be just an RNN, but a group of RNN. Fred On Wed, Jul 13, 2016 at 3:03 PM, Vishal Ahuja <[email protected]> wrote: > A friend of mine has designed an RNN for a certain problem using Theano. > It has 14 input nodes, 2 hidden layers with 7 nodes each, and finally an > output node. We have around 30,000 such RNNs that need to be trained. I am > a software engineer with very little exposure to Machine Learning. What I > need to do is to speed up the training process of these RNNs. > > Looking at the problem from a CS perspective, I don't think that anything > can be done to speed up the training of a single RNN. Running such a small > RNN on a GPU makes no sense. Instead, we can achieve speed up by batching > the RNNs, say 1000 at a time, and sending them to the GPU. The nature of > the problem is SIMD - each RNN is identical, but it has to train on a > different data set. > > Can someone please explain how this could be done using Theano? > > -- > > --- > 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. > -- --- 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.
