Obsolete/erroneous code in convolutional_mlp.py at DeepLearningTutorials? - neural-network

Question

This code contains the following tidbit:
from theano.tensor.nnet import conv2d
...
# convolve input feature maps with filters
conv_out = conv2d(
input=input,
filters=self.W,
filter_shape=filter_shape,
input_shape=image_shape
)
which raises an exception due to 'input_shape' not being found, despite being mentioned in the documentation where it says that:
"image_shape ... – Deprecated alias for input_shape"
Looking at conv.py both locally and in the source I found:
def conv2d(input, filters, image_shape=None, filter_shape=None,
border_mode='valid', subsample=(1, 1), **kargs):
Needless to say, there is no trace of input_shape.
If one modifies the code above as follows
# convolve input feature maps with filters
conv_out = conv2d(
input=input,
filters=self.W,
filter_shape=filter_shape,
image_shape=image_shape
)
, the exception disappears and the code runs fine.
What am I missing? If image_shape is deprecated, how come it works while input_shape does not?
Is the theano version at the repository obsolete?
PS: I would have liked to ask directly the folks at http://deeplearning.net, but I could not find how.

Answer 1

Are you sure you have the latest version installed?
conv.py contains the deprecated implementation of conv2d. The new implementation can be found in __init__.py
Make sure you are using the import statement
from theano.tensor.nnet import conv2d
and not
from theano.tensor.nnet.conv import conv2d
since the second one is going to import the deprecated implementation