cvb0.py

import numpy.random.mtrand
import numpy as np

class LdaCvb0:
    def __init__(self, word_indexes, word_counts, n_topics, alpha=0.1, beta=0.01):
        self.word_counts = word_counts
        self.word_indexes = word_indexes
        self.n_words = np.max(np.max(word_indexes)) + 1
        self.n_topics = n_topics
        self.gamma_dik = []
        self.mean_ndk = np.zeros([word_indexes.shape[0], n_topics])
        self.var_ndk = np.zeros([word_indexes.shape[0], n_topics])
        self.alpha = alpha
        self.beta = beta
        self.mean_nkv = np.zeros([n_topics, self.n_words])
        self.var_nkv = np.zeros([n_topics, self.n_words])

        for d, (doc_w, doc_c) in enumerate(zip(word_indexes,word_counts)):
            gamma_ik = []
            for i, (word, count) in enumerate(zip(doc_w,doc_c)):
                gamma_k = numpy.random.mtrand.dirichlet([alpha] * n_topics)
                gamma_ik.append(gamma_k)


            self.gamma_dik.append(gamma_ik)
            self.count_update(i, d, word, count)

        print self.gamma_dik
    def count_update(self, d, i, v, scale):

        for k in range(self.n_topics):
            mc = scale * self.gamma_dik[i][d][k]
            vc = scale * self.gamma_dik[i][d][k] * (1. - self.gamma_dik[i][d][k])
            self.mean_ndk[d][k] += mc
            self.var_ndk[d][k] += vc

            self.mean_nkv[k][v] += mc
            self.var_nkv[k][v] += vc

    def infer(self, word_numbers, word_counts, n_topics):

        for d, (doc_w, doc_c) in enumerate(zip(word_indexes, word_counts)):
            for i, (word, count) in enumerate(zip(doc_w, doc_c)):
                self.count_update(i, d, word, -count)
                for k in range(self.n_topics):
                    new_gamma = (self.mean_nkv[k][word] + self.beta) / (self.mean_nkv[k].sum() + self.beta * self.n_words)
                    new_gamma *= self.mean_ndk[d][k] + self.alpha
                    self.gamma[d][i][k]



        n_doc = len(word_indexes)
        for doc_no in range(n_doc):
            word_no = word_numbers[doc_no]
            word_conter = word_counts[doc_no]
            """
            for word, word_count in zip(word_no,word_conter):
                gamma[]
            """

docs_w = np.array([[1,2],
          [1,2],
          [3,4],
          [3,4],
          [0],
          [0]])
docs_c = np.array([[2,1],
          [4,1],
          [3,1],
          [4,2],
          [5],
          [4]])

ctx = LdaCvb0(docs_w, docs_c, 3)
print ctx