need to unify the shape of loss operators' output

[QiJune]

We have implement cos_sim_operator. And we use a reduce operation sum to get (x1 * x1 + x2 * x2 + ...). The reduce operations in Eigen will reduce dimensions, please refer to https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md#reduction-operations.
And this means we will get a column vector(batch_size) instead of a matrix(batch_size * 1).

Now, cos_sim operator return a matrix(batch_size * 1), and it can pass unittest in ubuntu/gcc, but can not pass in mac/clang. In clang, it will cause a error.

Assertion failed: (dimensions_match(m_leftImpl.dimensions(), m_rightImpl.dimensions())), function evalSubExprsIfNeeded

And it's obvious:

auto x = EigenMatrix<T>::From(*input_x, new_dims);
auto x_norm = EigenMatrix<T>::From(*input_x_norm);
x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({1})).sqrt();

The result of sum is a rank 1 vector, and x_norm is a rank 2 matrix.

So, I suggest that the output of loss operator is a rank 1 vector, it's more reasonable and can pass unittest in clang build version.

[qingqing01]

Usually the cost is a matrix with shape [batch_size x 1]. Maybe we can use the vector with rank 1 to represent x_norm.

[QiJune]

Ok, since we can not tell a rank 1 vector is a row vector or column vector, we usually use (batch_size, 1) matrix to represent column vector. I will change x_norm to EigenVector to pass clang development environment.