Implement Unshared Convolution

Check shapes of spatial transform theta parameters

Add data type configuration to algorithms hash for cuDNN convolutions caching system

Errors fixed. Suggestions implemented.

Minor changes

Minor changes

Gradient calculation added.

Test for forward added. Errors fixed

Grad wrt weights done. Grad wrt inputs incomplete

Grad inp

Fix typo

Tests and bug fix for Grad Inp

Modified flops to raise error

Mostly convdim agnostic. Cleaner code for gradInput

Some corr changes

MiChecks for convdim=2 added. Some more misc changes

Unshared code moved into one func

Re-added unshared flag to get_conv_output_shape

Simpler grad inputs. Unshared removeded from get_conv_output_shape.

C code changes in corr.py

wdim bug fix

opt and abstract_conv changes

CPU code for fwd and gradWeights. Added tests. Some errors

gemv increment fixed. Values for fwd still don't match

Forward perfect. Gradweights inverts regions; to be corrected.

Added grad inputs and tests but allclose error

Python gradInputs simplified

Grad input fixed

gradweights flipping problem solved

Weight dimension order changed. C cache version updated.

Docstring changes

if unshared is True -> if unshared. Specific error messages for unshared in C code.

Unshared tests integrated with AbstractConv. Subsampling errors fixed. Allclose errors with optimiser enabled

Kern flip in optimiser fixed. Still some errors

Errors fixed

GPU corr_gemm code (untested)

Unnecessary changes rolled back

More GPU code but gemm error 11

Fixed mistakes caused while copying from CPU

Errors fixed

Fixed error with .data for gpuarray

GPU tests

Suggestions implemented for error messages

Jenkins errors fixed

Commits squashed

Small errors fixed. Tests need to be rewritten

Tests moved to separate class. Mistakes fixed

Tests sped up

Suggestions implemented. Tests modified

with different data type configurations.

runtime algorithms and different data type configurations.

Add more tests for meta-optimizer

Fix Theano function overhead increase due to gh-6275. 

Fix Theano function overhead increase due to gh-6275. Shared variable was not triggering the input alias check and that was right.

Support NumPy 1.13

Restrict cuDNN reduction to contiguous inputs