I have the following snippet from
http://stackoverflow.com/questions/31733166/overlapping-iteration-over-theano-tensor:
*import theano*
*import theano.tensor as T*
*# Input variables*
*x = T.vector("x")*
*slice_size = T.iscalar("slice_size")*
*def step(idx, vect, length):*
* # From the idx of the start of the slice, the vector and the length of*
* # the slice, obtain the desired slice.*
* my_slice = vect[idx:idx + length]*
* # Do something with the slice here. I don't know what you want to do*
* # to I'll just return the slice itself.*
* output = my_slice*
* return output*
*# Make a vector containing the start idx of every slice*
*slice_start_indices = T.arange(x.shape[0] - slice_size + 1)*
*out, updates = theano.scan(fn=step,*
* sequences=[slice_start_indices],*
* non_sequences=[x, slice_size])*
*fct = theano.function([x, slice_size], out)*
*print fct(range(17), 5)*
And it fails with:
ValueError: ('The following error happened while compiling the node',
for{gpu,scan_fn}(Elemwise{Composite{((i0 + i1) - i2)}}.0,
Elemwise{add,no_inplace}.0, Subtensor{int64:int64:int8}.0,
Elemwise{Composite{((i0 + i1) - i2)}}.0, GpuFromHost.0), '\n', 'numpy.dtype
has the wrong size, try recompiling')
--
---
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.
