Error on Neural Turing Machine Demo

[joschu]

Error on Neural Turing Machine Demo

[sbos]

I have the same problem running this demo and sometimes with my code.
It's strange that in my code I can overcome this by disabling simplifications.

Meanwhile, 7 tests fail with the same error.

[jramapuram]

What did you use for your CUDA build settings? I enabled CUDA & CUDNN and am running v7 of CUDA

[sbos]

I don't use CUDA at all and I'm not sure that this error is related to it.
I tried to debug my code which fails with the same error and find what causes it.

def py_numeric_apply(node, vals):
...
        #vals is np.array with dtype=float32
        out = alloc_output(node,vals)
        #out is np.array with dtype=int64 because node.typ.dtype is so
        callable.call(vals, out) 
        # this causes the error since np.ceil always returns floats (even if the input is integer)
        # which cannot be written to int64 array

If got it right according to UNARY_INFO the iceil operation should cast it's result to integer somehow, but it doesn't.

[sbos]

Changing

def _nu_iceil(x,out=None):
    if out is None:
        return np.ceil(x)
    else:
        np.ceil(x, out)

to

def _nu_iceil(x,out=None):
    if out is None:
        return np.ceil(x)
    else:
        out[...] = np.ceil(x)

helps me and allows to run the demo. My code however fails with the following error:

Python(87053,0x7fff71141000) malloc: *** mach_vm_map(size=140458053210112) failed (error code=3)
*** error: can't allocate region
*** set a breakpoint in malloc_error_break to debug
Segmentation fault: 11

[jramapuram]

This is now fixed with the commit from @sbos [ #46 ]