Mdrnn Specialization for MDLSTMs does not give correct gradients

[bayerj]

However, an MDRNN on a sequence of shape (1, 1) works correct. Thus, the problem seems to be with the recurrent connections.

[ RUN ] TestGradient.MdlstmMdrnn New Parametrized instance. Numeric/Analytical:0.0127396 <-> 0.0127396 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:0.969693 <-> 0.969693 Numeric/Analytical:0.0127314 <-> 0.0127314 New Parametrized instance. Numeric/Analytical:0.0101248 <-> 0.0101248 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:0.770659 <-> 0.770659 Numeric/Analytical:0.0101182 <-> 0.0101182 New Parametrized instance. Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 New Parametrized instance. Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 Numeric/Analytical:-0 <-> 0 [ OK ] TestGradient.MdlstmMdrnn

[bayerj]

Resolved in http://github.com/bayerj/arac/commit/64c4670a3717d6a8fe7366d9ffee4220fa022871