make EquilibriumOptimizer use a dict to map ops to their nodes rather than running everything on everything.

list() the iterator for local_optimizers

replacing the test id (module path and fct name) with the first line
of the docstring, making it hard to identify what version of the test
failed.

This is mostly for tests that are shared between tensor and cuda/gpuarray.

Fix crash in python 3 recently inserted.

Remove Rebroadcast in the input of Join.

Add it only when needed in the grad.

Fix other crash in hessian of stack()

mod derivative

Fixes in downsample.py for python 3

Few changes to mod derivative

Add c code to rebroadcast.

This also fix a memory leak when Theano gc is disabled.

Convert string literals for Python3

Add missing include in the new gpuaelemwise op.

It was causing compilation crash in the DLT mlp

This one worked on Linux, but not on Windows, for some reason.

In RepeatOp, fix detection of numpy-supported dtypes

GpuShape

c_code() for GpuCAReduce (in gpuarray)

Add tests for DownsampleFactorMaxGradGrad

Added second derivative to max downsampling function