TensorType.shape
The shape attribute of TensorType is more difficult to include that I originally thought, but I have experiment code that depends on this until Nov 6 2009.
It should be deleted from Theano afterwards.
Shape, stride and [constant] data attributes could be handled by partially-known type attributes, but this would require revising many places in theano where type equality is used when type compatibility is meant. Before embarking on a fork that develops this idea, do some basic estimates of how much we could benefit from having shape and stride information in the graph. How many more bugs would we catch, how much faster would our auto-generated code be?
It should be deleted from Theano afterwards.
Shape, stride and [constant] data attributes could be handled by partially-known type attributes, but this would require revising many places in theano where type equality is used when type compatibility is meant. Before embarking on a fork that develops this idea, do some basic estimates of how much we could benefit from having shape and stride information in the graph. How many more bugs would we catch, how much faster would our auto-generated code be?
Leave a comment