On Wed, Sep 21, 2016, Petar Palasek wrote:
> Hi,
>
> I am getting the following error when using images2neibs with neib_shape is
> given as variables and used in a function with scan.
>
> python shapebug.py
> Using gpu device 0: Quadro K4200 (CNMeM is disabled, cuDNN 5004)
> theano version 0.8.2
> WARNING (theano.tensor.opt): Failed to infer_shape from Op
> Images2Neibs{valid}.
> Input shapes: [(TensorConstant{1}, Shape_i{1}.0, Shape_i{2}.0,
> Shape_i{3}.0), (TensorConstant{2},), (TensorConstant{2},)]
> Exception encountered during infer_shape: <type 'exceptions.ValueError'>
> Exception message: length not known: <TensorType(int8, vector)> [id A]
>
> Traceback: Traceback (most recent call last):
> File "/usr/local/lib/python2.7/dist-packages/theano/tensor/opt.py", line
> 917, in get_node_infer_shape
> [self.shape_of[r] for r in node.inputs])
> File
> "/usr/local/lib/python2.7/dist-packages/theano/tensor/nnet/neighbours.py",
> line 258, in infer_shape
> c, d = node.inputs[1]
Here, node.inputs[1] is a vector, and Theano is not able to guarantee
that it has a length of exactly 2, which is the error.
It works for constants such as (2, 2) because then Theano can check that the
length is 2.
In this case, the code should probably be simply:
c = node.inputs[1][0]
d = node.inputs[1][1]
> File "/usr/local/lib/python2.7/dist-packages/theano/tensor/var.py", line
> 549, in __iter__
> for i in xrange(theano.tensor.basic.get_vector_length(self)):
> File "/usr/local/lib/python2.7/dist-packages/theano/tensor/basic.py",
> line 4300, in get_vector_length
> raise ValueError("length not known: %s" % msg)
> ValueError: length not known: <TensorType(int8, vector)> [id A]
>
>
> The code I am running is:
>
> import theano
> import theano.tensor as T
> from theano.tensor.nnet.neighbours import images2neibs
> import numpy as np
>
> def process_one_input_image(input_image, filter_h, filter_w):
> all_windows = images2neibs(input_image.dimshuffle('x', 0, 1, 2),
> neib_shape=(filter_h, filter_w), neib_step=(1, 1))
>
> # some other code
>
> return all_windows
>
> def process_all_input_images(input_images, filter_h, filter_w):
> s, _ = theano.scan(fn=process_one_input_image,
> sequences=[input_images], non_sequences=[filter_h, filter_w])
>
> return s
>
> print 'theano version', theano.__version__
>
> input_images = T.tensor4('input_images')
>
> num_input_images = 3
> image_h = 5
> image_w = 5
> image_ch = 2
>
> filter_h = 2
> filter_w = 2
>
> windows = process_all_input_images(input_images, filter_h, filter_w)
>
> f = theano.function([input_images], windows)
>
> images = np.arange(num_input_images * image_h * image_w *
> image_ch).reshape((num_input_images, image_ch, image_h,
> image_w)).astype('float32')
>
> print f(images)
>
> If I change neib_shape=(filter_h, filter_w) to e.g. neib_shape=(2, 2) the
> error goes away.
>
> Is this a bug or am I doing something wrong? How do I get around it?
>
> Best,
> Petar Palasek
>
> --
>
> ---
> 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.
