8ea593ba135faf00d9ba39f393682172366eee9b,conceptnet5/vectors/retrofit.py,,retrofit,#Any#Any#Any#Any#Any#,53

Before Change

    // Subtract the mean so that vectors don"t just clump around common
    // hypernyms
    orig_vecs -= orig_vecs.mean(0)

    // Delete the frame we built, we won"t need its indices again until the end
    del retroframe

    vecs = orig_vecs
    for iteration in range(iterations):
        if verbosity >= 1:
            print("Retrofitting: Iteration %s of %s" % (iteration+1, iterations))

        vecs = sparse_csr.dot(vecs)
        vecs -= vecs.mean(0)

        // use sklearn"s normalize, because it normalizes in place and
        // leaves zero-rows at 0
        normalize(vecs, norm="l2", copy=False)

After Change

    // Divide up the labels by what language they"re in -- we"ll use this to
    // subtract the mean of each language, reducing clumping by language and
    // improving multilingual alignment.
    rows_by_language = defaultdict(list)
    for i, label in enumerate(row_labels):
        lang = label.split("/")[2]
        rows_by_language[lang].append(i)
    all_languages = sorted(rows_by_language)
    row_groups = [rows_by_language[lang] for lang in all_languages]

    // Subtract the mean so that vectors don"t just clump around common
    // hypernyms
    for row_group in row_groups:
        orig_vecs[row_group] -= orig_vecs[row_group].mean(0)

    // Delete the frame we built, we won"t need its indices again until the end
    del retroframe

    vecs = orig_vecs
    for iteration in range(iterations):

Instances