Scan does not have a C implementation, it does have a Cython
implementation called from Python, which is why it shows up as "Py".
That being said, if you do not need to backpropagate gradients through
that loop, you may simply compile a Theano function for 1 timestep,
without inputs or outputs (updating shared variables instead) and
calling f.fn(n_calls=100000), since _that_ loop would be in C.
On Thu, Oct 13, 2016, Jozef Vesely wrote:
>
> For the cumsum example profiler says 100% time is spent in
> theano.scan_module.scan_op.Scan of type Py
>
> Scan Op profiling ( scan_fn )
> ==================
> Message: None
> Time in 1 calls of the op (for a total of 100000 steps) 1.166187e+00s
>
> Total time spent in calling the VM 3.082674e-01s (26.434%)
> Total overhead (computing slices..) 8.579197e-01s (73.566%)
>
> 10000 calls of add inside the loop take total 0.076s (still 10x worse than
> pure C):
>
> 0.076s 7.62e-07s C 100000 1 Elemwise{add,no_inplace}
>
> However the real overhead seems to be managing the slices and calling the
> VM, no
> wonder if its done in python. I tried to force compilation into C using:
>
> mode=theano.compile.mode.Mode(linker='c')
>
> but I got:
> AttributeError: 'Scan' object has no attribute 'c_compile_args'
>
> There no C implementation of Scan operation?
>
>
>
>
> --
>
> ---
> 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 theano-users+unsubscr...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.
--
Pascal
--
---
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 theano-users+unsubscr...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
