training.py 文件源码

def treegauss_add_row(
        data_row,
        tree_grid,
        program,
        latent_row,
        vert_ss,
        edge_ss,
        feat_ss, ):
    # Sample latent state using dynamic programming.
    TODO('https://github.com/posterior/treecat/issues/26')

    # Update sufficient statistics.
    for v in range(latent_row.shape[0]):
        z = latent_row[v, :]
        vert_ss[v, :, :] += np.outer(z, z)
    for e in range(tree_grid.shape[1]):
        z1 = latent_row[tree_grid[1, e], :]
        z2 = latent_row[tree_grid[2, e], :]
        edge_ss[e, :, :] += np.outer(z1, z2)
    for v, x in enumerate(data_row):
        if np.isnan(x):
            continue
        z = latent_row[v, :]
        feat_ss[v] += 1
        feat_ss[v, 1] += x
        feat_ss[v, 2:] += x * z  # TODO Use central covariance.