hello, all,
i try to use following code to test the time of matrix multiple and
slice-matrix multiple, some strange and confusing things happen for me:
please have a look.
import theano.tensor as T
import theano
import numpy
import numpy as np
from theano import function, config, shared, tensor, sandbox
import time
vlen = 10 * 30 * 768 # 10 x #cores x # threads per core
iters = 1000
nw = np.random.sample((512, 30000))
nw = shared(np.array(nw, dtype=theano.config.floatX))
def multiple(x):
return T.dot(x, nw)
def multiple2(x, i):
return T.dot(x, nw[:, i])
x = T.dmatrix('x')
result1 = multiple(x)
result2 = multiple2(x, 29998)
fn1 = theano.function(inputs=[x], outputs=result1)
fn2 = theano.function(inputs=[x], outputs=result2)
nx = np.random.sample((2560, 512))
s = time.time()
fn1(nx)
e = time.time()
print 'no slice: ', (e - s)
s = time.time()
fn2(nx)
e = time.time()
print 'slice 1: ', (e - s)
i = T.ivector('i')
fn3 = theano.function(inputs=[x, i], outputs=multiple2(x, i))
ni = [29998]
s = time.time()
fn3(nx, ni)
e = time.time()
print 'slice 2: ', (e - s)
i = T.iscalar('i')
fn3 = theano.function(inputs=[x, i], outputs=multiple2(x, i))
s = time.time()
fn3(nx, 29998)
e = time.time()
print 'slice 3: ', (e - s)
when i run this cpu, i got following result:
no slice: 46.1797990799
slice 1: 0.00203585624695
slice 2: 0.0838360786438
slice 3: 0.00367498397827
and when i run this gpu, i got:
no slice: 0.525362014771
slice 1: 0.0137898921967
slice 2: 0.00286388397217
slice 3: 0.0097770690918
for first row, when i do not use slice, multipling 30K column large matrix
is very slow, this is make sense.
however, i am confusing with the other three rows:
when i use slice, the way how i give parameter i to the function will
effect the speed of slice
multiple, which way should i choose to make the slice multiple to be the
fastest.
any reply will be appreciated 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.
