IMDbTextCategorizationDemo.py

import argparse
import logging
import numpy as np
import keras
from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import chi2
from keras.datasets import imdb
from sklearn.feature_extraction.text import CountVectorizer
from tmu.models.classification.vanilla_classifier import TMClassifier
from tmu.tools import BenchmarkTimer

_LOGGER = logging.getLogger(__name__)

if __name__ == "__main__":

    parser = argparse.ArgumentParser()
    parser.add_argument("--num_clauses", default=10000, type=int)
    parser.add_argument("--T", default=8000, type=int)
    parser.add_argument("--s", default=2.0, type=float)
    parser.add_argument("--device", default="GPU", type=str)
    parser.add_argument("--weighted_clauses", default=True, type=bool)
    parser.add_argument("--epochs", default=40, type=int)
    parser.add_argument("--clause_drop_p", default=0.75, type=float)
    parser.add_argument("--min_n", default=1, type=int)
    parser.add_argument("--max_n", default=2, type=int)
    parser.add_argument("--analyzer", default="word", type=str)
    parser.add_argument("--features", default=5000, type=int)
    parser.add_argument("--imdb-num-words", default=5000, type=int)
    parser.add_argument("--imdb-index-from", default=2, type=int)
    parser.add_argument("--max_included_literals", default=32, type=int)

    args = parser.parse_args()

    _LOGGER.info("Preparing dataset")
    train, test = keras.datasets.imdb.load_data(num_words=args.imdb_num_words, index_from=args.imdb_index_from)
    train_x, train_y = train
    test_x, test_y = test

    word_to_id = keras.datasets.imdb.get_word_index()
    word_to_id = {k: (v + args.imdb_index_from) for k, v in word_to_id.items()}
    word_to_id["<PAD>"] = 0
    word_to_id["<START>"] = 1
    word_to_id["<UNK>"] = 2
    _LOGGER.info("Preparing dataset.... Done!")

    _LOGGER.info("Producing bit representation...")

    id_to_word = {value: key for key, value in word_to_id.items()}

    training_documents = []
    for i in range(train_y.shape[0]):
        terms = []
        for word_id in train_x[i]:
            terms.append(id_to_word[word_id])

        training_documents.append(" ".join(terms))

    testing_documents = []
    for i in range(test_y.shape[0]):
        terms = []
        for word_id in test_x[i]:
            terms.append(id_to_word[word_id])

        testing_documents.append(" ".join(terms))

    vectorizer_X = CountVectorizer(
        ngram_range=(args.min_n, args.max_n),
        analyzer=args.analyzer,
        lowercase=True,
        binary=True
    )

    X_train = vectorizer_X.fit_transform(training_documents)
    Y_train = train_y.astype(np.uint32)

    X_test = vectorizer_X.transform(testing_documents)
    Y_test = test_y.astype(np.uint32)
    _LOGGER.info("Producing bit representation... Done!")

    _LOGGER.info("Selecting Features....")

    SKB = SelectKBest(chi2, k=args.features)
    SKB.fit(X_train, Y_train)

    selected_features = SKB.get_support(indices=True)
    X_train = SKB.transform(X_train).toarray()
    X_test = SKB.transform(X_test).toarray()

    _LOGGER.info("Selecting Features.... Done!")

    tm = TMClassifier(args.num_clauses, args.T, args.s, platform=args.device, weighted_clauses=args.weighted_clauses,
                      clause_drop_p=args.clause_drop_p)

    _LOGGER.info(f"Running {TMClassifier} for {args.epochs}")
    for epoch in range(args.epochs):
        benchmark1 = BenchmarkTimer(logger=_LOGGER, text="Training Time")
        with benchmark1:
            tm.fit(X_train, Y_train)

        benchmark2 = BenchmarkTimer(logger=_LOGGER, text="Testing Time")
        with benchmark2:
            Y_test_predicted, Y_test_scores = tm.predict(X_test, return_class_sums=True)
            result = 100 * (Y_test_predicted == Y_test).mean()

        _LOGGER.info(f"Epoch: {epoch + 1}, Accuracy: {result:.2f}, Training Time: {benchmark1.elapsed():.2f}s, "
                     f"Testing Time: {benchmark2.elapsed():.2f}s")

        np.savetxt("class_sums/IMDBAnalyzer_%d_%d_%d_%.2f_%d_%d_%d_%d_%s_%.2f_%d.txt" % (epoch+1, args.num_clauses, args.T, args.s, args.min_n, args.max_n, args.max_included_literals, args.weighted_clauses, args.analyzer, args.clause_drop_p, args.features), Y_test_scores, delimiter=',')