Doc fix and enhancement for lstm_unit python wrapper.

[pkuyym]

Resolves #7156

[sidgoyal78]

Thanks for the PR. Looks good. Just a minor clarification: Since, we aren't dealing with the W_{xc} c_{t-1} (where x = i and o), the current implementation is a minor variant (with no peephole connections) of the more general LSTM right?

[pkuyym]

@sidgoyal78
Yes, you are right, implementation of lstm_unit mainly refers caffe2, and peephole is not supported.
https://github.com/caffe2/caffe2/blob/master/caffe2/operators/lstm_unit_op.h#L1

[sidgoyal78]

LGTM

[sidgoyal78]

Oops, as you might have seen already, i think the build failed.

[pkuyym]

Thanks. An random failure, will restart CI ASAP.

[sidgoyal78]

@pkuyym : Actually, it is not random. I think this line needs to be corrected:https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/test_layers.py?utf8=%E2%9C%93#L180 and L181
Since we want second dim of prev_hidden and prev_cell to be same, and they are not same here!

[sidgoyal78]

No worries, I think I should have also thought of that while reviewing. So no problems.

[sidgoyal78]

LGTM.