Hi,
I'm currently trying to implement the method in the paper "
Compressing Neural Networks with the Hashing Trick
<https://arxiv.org/pdf/1504.04788>
"
The idea is very simple. I just get stuck how to aggregate the elements in
the 4D tensor quickly. Each element in the 4D tensor is mapped to a bucket
in a low-dimensional vector, and elements mapped to the same bucket needs
to be added up together. Currently, I just use a 'map' function to
implement this, and at each iteration, I add the element in the 4D tensor
in the corresponding bucket. But it is extremely slow.
'gp_value' is the 4D value of the tensor. idxW stores the bucket ID of the
low dimensional vector. 'grad_h_weight' is the vector.
def fn_reduce(cur_grad,index):
return {grad_h_weight: T.inc_subtensor(grad_h_weight[index],
cur_grad)}
_, updates = theano.map(fn_reduce,
sequences=[theano.shared(np.asarray(gp_value,dtype='float32'),borrow=True).flatten(),
idxW])
cal = theano.function([],updates=updates)
cal()
Any one can help?
Thanks very much.
--
---
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.
