utils.py

'''Common functions
'''

import itertools
import numpy as np
import nltk


def softmax(x):
    maximum = np.max(x)
    norm = x - maximum
    numer = np.exp(norm)
    denom = np.sum(numer)
    return numer / denom


def softmax_batch(x):
    maximum = np.max(x, axis=1, keepdims=True)
    norm = x - maximum
    numer = np.exp(norm)
    denom = np.sum(numer, axis=1, keepdims=True)
    return numer / denom


def softmax_derivative(y):
    l = y.shape[0]
    der = np.outer(y, y)
    der -= y * np.eye(l)
    return -der


def softmax_derivative_batch(y):
    l = y.shape[1]
    der = y.T.dot(y)
    col_sum = np.sum(y, axis=1, keepdims=True)
    der -= col_sum * np.eye(l)
    return -der


def cross_entropy_one_hot_derivative(y, n):
    dEdy = np.copy(y)
    dEdy[n] -= 1
    return dEdy


def cross_entropy_one_hot_derivative_batch(y, n):
    dEdy = np.copy(y)
    indice = np.arange(n.shape[0])
    dEdy[indice, n] -= 1
    return dEdy


def softmax_cross_entropy_batch(y, n):
    l = y.shape[1]
    dEdy = cross_entropy_one_hot_derivative_batch(y, n)
    mul = dEdy * y
    der = y.T.dot(mul)
    col_sum = np.sum(mul, axis=1, keepdims=True)
    der -= col_sum * np.eye(l)
    return -der


def matrix_multi_derivative(w, x, dydx=True, x_is_multiplier=True):
    if w.ndim != 2 or x.ndim != 1:
        raise Exception('dimension error')

    m = w.shape[0]
    n = w.shape[1]

    if x_is_multiplier:
        '''y = np.dot(w,x)  m by n dot n by 1
        if multiplicand == True, return dy/dw
        else return dy/dx
        '''
        if dydx:
            # dy/dx
            return w.T
        else:
            # dy/dw
            tmpx = np.tile(x, n).reshape(m * n, 1)
            shape = np.eye(m).repeat(n, axis=0)
            return shape * tmpx
    else:
        '''y = np.dot(x,w) 1 by m dot m by n
        '''
        if dydx:
            return w
        else:
            shape = np.tile(np.eye(n), m).T
            tmpx = x.repeat(n).reshape(m * n, 1)
            return shape * tmpx


def one_hot_multiple(voca_size, targets):
    return np.eye(voca_size)[targets]


def one_hot(voca_size, target):
    one_hot = np.zeros(voca_size)
    one_hot[target] = 1
    return one_hot.reshape(1, voca_size)


def cos_sim(embedding):
    mul = embedding.dot(embedding.T)
    mag = np.diag(mul)
    inv_mag = 1 / mag
    inv_mag[np.isinf(inv_mag)] = 0
    inv_mag = np.sqrt(inv_mag)
    sim = mul * inv_mag
    return sim.T * inv_mag


corpus_file = 'data/corpus.txt'


def read_corpus(corpus_file=corpus_file):
    start = 'unkstart'
    end = 'unkend'

    with open(corpus_file) as f:
        text = f.read()
        tokens = nltk.sent_tokenize(text)
        sentences = ["%s %s %s" % (start, x.lower(), end) for x in tokens]

    print('%d sentences' % len(sentences))
    return sentences


def words(sentences):
    tokenized_sentences = [nltk.word_tokenize(sent) for sent in sentences]
    words = list(itertools.chain(*tokenized_sentences))
    print("%d words" % len(words))
    return words


def word_index(words):
    i2w = list(set(words))  # index to word
    w2i = dict([(word, i) for i, word in enumerate(i2w)])  # word to index
    return w2i, i2w


def train_data(words, w2i, i2w):
    '''window size = 1
    '''
    length = len(words)
    data = []
    for i in range(1, length - 1):
        left = [w2i[words[i]], w2i[words[i + 1]]]
        right = [w2i[words[i]], w2i[words[i - 1]]]
        data.append(left)
        data.append(right)

    data = np.array(data)

    train_x = data[:, 0]
    train_y = data[:, 1]
    return train_x, train_y