Theano: Indexing inside a compiled function (GPU)

Question

Since Theano allows to update the memory on the graphics card DRAM by simply defining which memory area has to be updated and how this should be done, I was wondering if the following is somehoe possible (imho it should be).

I have a 2x5 randomly initialized matrix which first column will be initialized with start-values. I would like to write a function that depends of the preceeding column and updates the next one based on arbitrary calulations.

I think this code explains it very well:

Note: This code is not working, it's just an illustration:

from theano import function, sandbox, shared
import theano.tensor as T
import numpy as np

reservoirSize = 2
samples       = 5

            # To initialize _mat first column
_vec      = np.asarray([1 for i in range(reservoirSize)], np.float32)

            # Random matrix, first column will be initialized correctly (_vec)
_mat      = np.asarray(np.random.rand(reservoirSize, samples), np.float32)
_mat[:,0] = _vec

print "Init:\n",_mat

_mat      = shared(_mat)

idx  = T.iscalar()
test = function([idx], updates= {
            # The indexing causes the problem here. Imho there should be
            # a way to do something like this:
            #    update: _mat[:, idx] = _max[:, idx-1] * 2
            _mat[:,idx]:sandbox.cuda.basic_ops.gpu_from_host(_mat[:,idx-1] * 2) 
            })

for i in range(1, samples):
    test(i)

print "Done:\n",_mat

My desired output would be:

Init:
[[ 1. 0.62166548  0.17463242  0.00858122  0.59709388]
 [ 1. 0.52690667  0.20800008  0.86816955  0.43518791]]
Done: 
[[ 1. 2. 4. 8. 16. ]
   1. 2. 4. 8. 16. ]]

but instead I get

Init:
[[ 1. 0.62166548  0.17463242  0.00858122  0.59709388]
 [ 1. 0.52690667  0.20800008  0.86816955  0.43518791]]
Traceback (most recent call last):
  File "/home/snooc/workspace/eclipse-python/Bachelorarbeit/theano/test.py", line 20, in <module>
    _mat[:,idx]:sandbox.cuda.basic_ops.gpu_from_host(_mat[:,idx-1] * 2) })
  File "/usr/lib/python2.7/site-packages/theano/compile/function.py", line 223, in function
    profile=profile)
  File "/usr/lib/python2.7/site-packages/theano/compile/pfunc.py", line 490, in pfunc
    no_default_updates=no_default_updates)
  File "/usr/lib/python2.7/site-packages/theano/compile/pfunc.py", line 194, in rebuild_collect_shared
    store_into)
TypeError: ('update target must be a SharedVariable', Subtensor{::, int32}.0)

Can anybody help me here?

Wow: This question is 9 minutes after asking already in the Top 4 of google results for "Theano indexing gpu" for me. O_o

score 1 · Accepted Answer · edited May 23 '17 at 10:32

Have a look at: How can I assign/update subset of tensor shared variable in Theano?

For your code this translates to:

from theano import function, sandbox, shared
import theano.tensor as T
import numpy as np

reservoirSize = 2
samples       = 5

# To initialize _mat first column
_vec      = np.asarray([1 for i in range(reservoirSize)], np.float32)

# Random matrix, first column will be initialized correctly (_vec)
_mat      = np.asarray(np.random.rand(reservoirSize, samples), np.float32)
_mat[:,0] = _vec

print "Init:\n",_mat

_mat      = shared(_mat)

idx  = T.iscalar()
test = function([idx], updates= {
            # -> instead of
            #_mat[:,idx]:sandbox.cuda.basic_ops.gpu_from_host(_mat[:,idx-1] * 2)
            # -> do this:
            _mat:T.set_subtensor(_mat[:,idx], _mat[:,idx-1]*2) 
            })

for i in range(1, samples):
    test(i)

print "Done:\n",_mat.get_value()  # use get_value() here to retrieve the data

Thank you! Btw I needed to know this for my bachelor thesis which is supervised by Prof. Legenstein. You may know him :) — Stefan Falk, Jun 20 '14 at 14:41

Theano: Indexing inside a compiled function (GPU)

1 Answers1