I'm attempting to write a differentiable, parameterized image rotation
layer in Theano with decent performance - as result I'm doing some slightly
unusual things in terms of indexing in order to get decent performance.
When I try to optimize the resulting theano graph using fast_run I get a
series of errors that don't appear when I use the straight python
implementation:
ERROR (theano.gof.opt): SeqOptimizer apply
<theano.scan_module.scan_opt.PushOutNonSeqScan object at 0x7f542b2db750>
ERROR (theano.gof.opt): Traceback:
ERROR (theano.gof.opt): Traceback (most recent call last):
File "/usr/local/lib/python2.7/dist-packages/theano/gof/opt.py", line
230, in apply
sub_prof = optimizer.optimize(fgraph)
File "/usr/local/lib/python2.7/dist-packages/theano/gof/opt.py", line 89,
in optimize
ret = self.apply(fgraph, *args, **kwargs)
File
"/usr/local/lib/python2.7/dist-packages/theano/scan_module/scan_opt.py",
line 228, in apply
self.process_node(fgraph, node)
File
"/usr/local/lib/python2.7/dist-packages/theano/scan_module/scan_opt.py",
line 313, in process_node
**dict(return_list=True))[0].owner
File "/usr/local/lib/python2.7/dist-packages/theano/gof/op.py", line 611,
in __call__
node = self.make_node(*inputs, **kwargs)
File "/usr/local/lib/python2.7/dist-packages/theano/tensor/subtensor.py",
line 2133, in make_node
index = tuple(map(as_index_variable, index))
File "/usr/local/lib/python2.7/dist-packages/theano/tensor/subtensor.py",
line 2079, in as_index_variable
idx = theano.tensor.as_tensor_variable(idx)
File "/usr/local/lib/python2.7/dist-packages/theano/tensor/basic.py",
line 167, in as_tensor_variable
"Variable type field must be a TensorType.", x, x.type)
AsTensorError: ('Variable type field must be a TensorType.', NoneConst,
<theano.tensor.type_other.NoneTypeT object at 0x7f542dbcab10>)
A snippet of relevant code:
W=2
> coordsP = T.add(pts.astype('float32'),self.D.reshape((-1,2)))
> contributors = getNeighborhood(coordsP,W)
> diffs = contributors - coordsP.reshape((-1,1,2))
> kernVals = kern(diffs,W)
> contributorsLin = contributorsToLinInd(contributors)
> # Loop over neighborhood
> result = T.zeros(input.shape)
> for i in range((2*W)**2):
> result = result +
> (input.astype('float32')[:,contributorsLin[:,i].astype('int16')])*(kernVals[:,i].T)
But basically there's a finite kernel that for each output pixel I want to
evaluate the sum of the kernel's values times the input pixels values over
the support of the kernel (neighborhood).
For performance I'm trying to explicitly write the loop over the input
pixels in the support of the kernel (rather than using a dynamically built
sparse matrix). My concern is that this is somewhat "weird" and I might be
hitting an edge case in the Theano optimization pipeline. I should add that
my goal is for this function to be thrice differentiable in self.D -
analytically this is the case, as my kernel has a triple zero at the bounds
of its support, but I need to make sure not to NaN anything along the way.
If anybody has any suggestions for debugging this I'd greatly appreciate it!
Thanks,
Michael
--
---
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.
