/** * Adds the diagonal of the Hessia of the expected log-likelihoood of * the specified data point to the Hessian diagonal h. */ void add_hessian_diag(const table<T>& phead, const uint_vector& tail, T w, table<T>& h) const { assert(phead.arity() + tail.size() == h.arity()); std::size_t index = h.offset().linear(tail, phead.arity()); for (std::size_t i = 0; i < phead.size(); ++i) { h[index + i] -= phead[i] * w / (f[index + i] * f[index + i]); } }
/** * Adds the gradient of the expected log-likelihood of the specified * data point to the gradient table g. * * \param phead the distribution over a leading set of indices of f * \param tail a fixed assignment to the remaining indices of f * \param w the weight of the data point */ void add_gradient(const table<T>& phead, const uint_vector& tail, T w, table<T>& g) const { assert(phead.arity() + tail.size() == g.arity()); std::size_t index = g.offset().linear(tail, phead.arity()); for (std::size_t i = 0; i < phead.size(); ++i) { g[index + i] += phead[i] * w; } }
/** * Restricts the given table to an assignment to all the dimensions * >= n, where n is the arity of this table. The parameter x_start is * the linear index of the first element in the input table that should * be copied. This table must be preallocated, and its first n * dimensions must match the first n dimensions of the input table. */ void restrict(const table& x, std::size_t x_start) { assert(arity() <= x.arity()); assert(std::equal(shape_.begin(), shape_.end(), x.shape_.begin())); assert(x_start + size() <= x.size()); std::copy(x.data() + x_start, x.data() + x_start + size(), data()); }