WordNet ImageNet/generate_wordnet_oodness.py

from nltk.corpus import LazyCorpusLoader
from nltk.corpus.reader import WordNetCorpusReader, CorpusReader
from nltk.corpus import wordnet as wn30, wordnet_ic
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import minmax_scale

wn31 = LazyCorpusLoader(
    "wordnet31",
    WordNetCorpusReader,
    LazyCorpusLoader("omw", CorpusReader, r".*/wn-data-.*\.tab", encoding="utf8"),
)

metrics = ['Path', 'Leacock-Chodorow', 'Wu-Palmer', 'Resnik', 'Jiang-Conrath', 'Lin']


def write_similarities():

    brown_ic = wordnet_ic.ic('ic-brown.dat')

    places_classes = []
    places_synsets = []

    with open('places365_labels.txt', "r") as f:

        for line in f:

            fields = line.split()
            _class = {
                'full_name': fields[0],
                'cleaned_name': fields[0].split('/')[2].replace("_", " "),
                'index': int(fields[1]),
                'synsets': [wn31.synset(f) for f in fields[2:]]
            }
            # print(fields[0], [(s, s.definition()) for s in synsets])
            places_classes.append(_class)
            places_synsets += _class['synsets']

    scores = []

    with open("imagenet_places_similarities.txt", "w") as o:
        o.write("in_synset;in_label;Path;Path_synset;Leacock-Chodorow;\
    LC_synset;Wu-Palmer;WP_synset;Resnik;Resnik_synset;Jiang-Conrath;JC_synset;Lin;Lin_synset\n")
        with open('LOC_synset_mapping.txt', "r") as f:

            for line in f:

                fields = line.split()
                synset_id = fields[0]

                pos = synset_id[0]
                offset = int(synset_id[1:])
                synset = wn30.synset_from_pos_and_offset(pos, offset)

                similarities = {
                    'Path':             [s.path_similarity(synset) for s in places_synsets],
                    'Leacock-Chodorow': [s.lch_similarity(synset) for s in places_synsets],
                    'Wu-Palmer':        [s.wup_similarity(synset) for s in places_synsets],
                    'Resnik':           [s.res_similarity(synset, brown_ic) for s in places_synsets],
                    'Jiang-Conrath':    [s.jcn_similarity(synset, brown_ic) for s in places_synsets],
                    'Lin':              [s.lin_similarity(synset, brown_ic) for s in places_synsets],
                }

                most_similar = {}
                for metric in similarities:
                    val = max(similarities[metric])
                    idx = similarities[metric].index(val)

                    most_similar[metric] = {
                        'value': val,
                        'synset': places_synsets[idx]
                    }

                scores.append((fields[1:], most_similar))

                sims = ";".join([";".join([str(most_similar[metric]['value']),
                                           most_similar[metric]['synset'].name()]) for metric in metrics])
                label = " ".join(fields[1:])
                output_line = ";".join([fields[0], label, sims])
                o.write(output_line + "\n")


def read_similarities():

    res = []

    with open('imagenet_places_similarities.txt', "r") as f:
        for line in f.readlines()[1:]:

            fields = line.split(";")
            obj = {}
            cnt = 2
            for m in metrics:
                obj[m] = {
                    'value':  float(fields[cnt]),
                    'synset': wn31.synset(fields[cnt + 1])
                }
                cnt += 2

            res.append((fields[1], obj))

    return res


def draw_histogram(data, labels, name):
    fig, axs = plt.subplots(len(labels), 1, figsize=(6.4, 10))

    for i, l in enumerate(labels):
        axs[i].set_title(l)
        axs[i].hist(data[:, i], bins=30)

    fig.show()
    fig.savefig(name)


def analysis(data):
    similarity_arr = np.array([[c[1][m]['value'] for m in metrics] for c in data])  # imagenet_classes x metrics

    draw_histogram(similarity_arr, metrics, "histogram")

    scaled_data = minmax_scale(similarity_arr, axis=0, feature_range=(0, 1))
    columns = metrics + ["Average", "Average (no ic)"]
    avg = np.average(scaled_data, axis=1).reshape(-1, 1)
    avg_no_ic = np.average(scaled_data[:, 0:3], axis=1).reshape(-1, 1)
    data_and_average = np.concatenate((scaled_data, avg, avg_no_ic), axis=1)

    draw_histogram(data_and_average, columns, "scaled_histogram")

    fig, ax = plt.subplots(figsize=(20, 180))
    im = ax.pcolormesh(data_and_average)
    fig.colorbar(im, ax=ax)
    ax.set_xticks(0.5 + np.arange(len(columns)))
    ax.set_yticks(0.5 + np.arange(len(data)))
    ax.set_xticklabels(labels=columns, rotation=90)
    ax.set_yticklabels(labels=[c[0].split(",")[0] for c in data])

    for y in range(len(data)):
        for x in range(len(metrics)):
            ax.text(x+0.1, y+0.2, data[y][1][metrics[x]]['synset'].name(), color="grey")

        ax.text(len(columns) - 1 + 0.1, y + 0.2, "%.2f" % data_and_average[y, -1], color="grey")
    fig.show()
    fig.savefig("heatmap")

    return data_and_average


def write_output(similarities):

    with open("in_wordnet_oodness.txt", "w") as f:
        for s in similarities:
            f.write("%.2f\n" % (1 - s))


def main():

    scores = read_similarities()
    res = analysis(scores)
    write_output(res[:, -1])


main()