word2vecリターンズ！ 品詞分類による精度改善：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, part_dim):
    self.input_dim = input_dim
    self.output_dim = output_dim
    self.part_dim = part_dim

  def create_model(self):
    word_input = Input(shape=(1,), dtype='int32', name='word_input')
    x_w = Embedding(output_dim=self.output_dim, input_dim=self.input_dim, input_length=1,
                    embeddings_initializer=uniform(seed=20170719))(word_input)
    part_input = Input(shape=(1,), dtype='int32', name='part_input')
    x_p = Embedding(output_dim=self.output_dim, input_dim=self.part_dim, input_length=1,
                    embeddings_initializer=uniform(seed=20170719))(part_input)
    x = keras.layers.add([x_w , x_p])
    x = Flatten()(x)
    word_output = Dense(self.input_dim, use_bias=False,activation='softmax', name='word_output',
                        kernel_initializer=glorot_uniform(seed=20170719))(x)
    part_output = Dense(self.part_dim, use_bias=False,activation='softmax', name='part_output',
                        kernel_initializer=glorot_uniform(seed=20170719))(x)
    model = Model(inputs=[word_input, part_input], outputs=[word_output, part_output])
    model.compile(optimizer='RMSprop',
        loss = {'word_output': 'categorical_crossentropy', 'part_output': 'categorical_crossentropy'},
        loss_weights = {'word_output': 1., 'part_output': 0.2})

    return model

  # 学習
  def train(self, x_train, t_train,x_p_train,t_p_train,batch_size,epochs, emb_param) :
    early_stopping = EarlyStopping(monitor='loss', patience=1, verbose=1)
    model = self.create_model()
    #model.load_weights(emb_param)               # 埋め込みパラメーターセット
    model.fit({'word_input': x_train, 'part_input': x_p_train},
              {'word_output': t_train, 'part_output': t_p_train},
                batch_size=batch_size, epochs=epochs,verbose=1,
                shuffle=True, callbacks=[early_stopping], validation_split=0.0)

    return model

vec_dim = 400 
epochs = 1
batch_size = 200
input_dim = len(words)
output_dim = vec_dim
part_dim = len(parts)

prediction = Prediction(input_dim, output_dim, part_dim)

emb_param = 'param_skip_gram_' + str(maxlen) + '_' + str(vec_dim) + '.hdf5'    # パラメーターファイル名
print(emb_param)

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
  if i % 50000 == 0 :
    print(i)
print(t_one_hot.shape)

t_p_one_hot = np.zeros((row, part_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_p_one_hot[i, t_p_train[i,j]] += 1 
  if i % 50000 == 0 :
    print(i)
print(t_p_one_hot.shape)

model = prediction.train(x_train, t_one_hot, x_p_train, t_p_one_hot, batch_size, epochs, emb_param)

model.save_weights(emb_param)                    # 学習済みパラメーターを保存

x_eval = x_train[:40000,:]
x_p_eval = x_p_train[:40000,:]
t_eval = t_one_hot[:40000,:]

y_pred = model.predict([x_eval, x_p_eval], batch_size=batch_size, verbose=1)   # 評価結果

n_true = 0
n_pred = 0
for i in range(0, x_eval.shape[0]) :
  label = set(t_one_hot[i,:].nonzero()[0])
  n_true += len(label)
  argmax_pred = y_pred[0][i,:].argsort()[::-1][0:len(label)]

  hit = set(label) & set(argmax_pred)

  n_pred += len(hit)
  del label,argmax_pred
  del hit

print()
print('正解率 =', n_pred * 1.0 / n_true )

param_list = model.get_weights()
print(len(param_list))
for i in range(0, len(param_list)):
  print(param_list[i].shape)

param = param_list[0]

word_m0 = '明智'
word_m1 = '小五郎'

word0 = '清太郎'
word1 = '清一郎'
word2 = '勇'
word3 = '武'
word4 = '廣介'
word5 = '源三郎'
word6 = '三郎'
word7 = '早苗'
word8 = '潤一'
word9 = '葉子'
word10 = '庄兵衛'

vec0 = param[word_indices[word_m0],:]
vec1 = param[word_indices[word_m1],:]
vec2 = param[word_indices[word2],:]

for i in range(0,11) :
  vec2 = param[word_indices[eval('word'+str(i))],:]
  vec = vec0 - vec1 + vec2
  vec_norm = math.sqrt(np.dot(vec, vec))

  w_list = [word_indices[word_m0], word_indices[word_m1], word_indices[eval('word'+str(i))]]
  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('コサイン類似度=', dist, ' ', m, ' ', indices_word[m])

param = param_list[0]

word_m0 = '明智'
word_m1 = '探偵'

word0 = '博士'
word1 = '判事'
word2 = '刑事'
word3 = '夫人'
word4 = '警部'

vec0 = param[word_indices[word_m0],:]
vec1 = param[word_indices[word_m1],:]
vec2 = param[word_indices[word2],:]

for i in range(0,5) :
  vec2 = param[word_indices[eval('word'+str(i))],:]
  vec = vec0 - vec1 + vec2
  vec_norm = math.sqrt(np.dot(vec,vec))

  w_list = [word_indices[word_m0], word_indices[word_m1], word_indices[eval('word'+str(i))]]
  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('コサイン類似度=', dist,' ', m, ' ', indices_word[m])