挑戦！ word2vecで自然言語処理（Keras＋TensorFlow使用）：ディープラーニング習得、次の一歩（2/2 ページ）

class Prediction :
  def __init__(self, input_dim, output_dim):
    self.input_dim = input_dim
    self.output_dim = output_dim
            
  def create_model(self):
    model = Sequential()
    model.add(Embedding(self.input_dim, self.output_dim, input_length=1, embeddings_initializer=uniform(seed=20170719)))
    model.add(Flatten())
    model.add(Dense(self.input_dim, use_bias=False, kernel_initializer=glorot_uniform(seed=20170719)))
    model.add(Activation("softmax"))
    model.compile(loss="categorical_crossentropy", optimizer="RMSprop", metrics=['categorical_accuracy'])
    print('#2')
    return model
  
  # 学習
  def train(self, x_train, t_train,batch_size, epochs, maxlen, emb_param) :
    early_stopping = EarlyStopping(monitor='categorical_accuracy', patience=1, verbose=1)
    print ('#1', t_train.shape)
    model = self.create_model()
    #model.load_weights(emb_param)    # 埋め込みパラメーターセット。ファイルをロードして学習を再開したいときに有効にする
    print ('#3')
    model.fit(x_train, t_train, batch_size=batch_size, epochs=epochs, verbose=1,
          shuffle=True, callbacks=[early_stopping], validation_split=0.0)
    return model

vec_dim = 100
epochs = 10
batch_size = 200
input_dim = len(words)
output_dim = vec_dim

emb_param = 'param_skip_gram_2_1.hdf5'    # 学習済みパラメーターファイル名
prediction = Prediction(input_dim, output_dim)
row = t_train.shape[0]

t_one_hot = np.zeros((row, input_dim), dtype='int8')    # ラベルデータをN-hot化
for i in range(0, row) :
  for j in range(0, maxlen*2):
    t_one_hot[i, t_train[i,j]] = 1
  
x_train = x_train.reshape(row,1)
model = prediction.train(x_train, t_one_hot, batch_size, epochs, maxlen, emb_param)

model.save_weights(emb_param)           # 学習済みパラメーターセーブ

param_lstm = model.get_weights()
param = param_lstm[0]
word0 = '一'
word1 = '１'
word2 = '２'
vec0 = param[word_indices[word0],:]
vec1 = param[word_indices[word1],:]
vec2 = param[word_indices[word2],:]

vec = vec0 - vec1 + vec2
vec_norm = math.sqrt(np.dot(vec, vec))

w_list = [word_indices[word0], word_indices[word1], word_indices[word2]]
dist = -1.0
m = 0
for j in range(0, 5) :
  dist = -1.0
  m = 0
  for i in range(0, len(words)) :
    if i not in w_list :
      dist0 = np.dot(vec, param[i,:])
      dist0 = dist0 / vec_norm / math.sqrt(np.dot(param[i,:], param[i,:]))
      if dist < dist0 :
        dist = dist0
        m = i
  print('第' + str(j+1) + '候補:')
  print('コサイン類似度=', dist, ' ', m, ' ', indices_word[m])
  w_list.append(m)