Otherwise, if they are symbolic variables, they could be
used during shape inference, which could introduce dependencies
on variables outside of the graph.

Force older conda version on travis

Make BatchedDot don't try to use c code if there is no blas.

More hints for failed AbstractConv2d optimization

flake8 sandbox/*.py

fix flake8 'd3viz/tests/*py'

extension of #3983

Fix blocksparse

Fix doc build

Update NEWS_DEV.txt

Assert to never reintroduce Apply node in fgraph

C implementation of sample without replacement OP

Add debugmode implementation for abstractconv

Add lifter for IfElse.

Try to detect bug in nvcc.