[Kaggle DAY15]Real or Not? NLP with Disaster Tweets!

Kaggle 15회차

오늘은 GRU모델을 활용해보기로 했습니다.

 

첫번째 제출

model = Sequential()
model.add(Embedding(vocab_size, 100))
model.add(GRU(100))
model.add(Dropout(0.5))
model.add(Dense(2, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['acc'])
history = model.fit(x_train, y_train, batch_size=32, epochs=3, validation_split=0.1)

결과

 

두번째 제출

model3 = Sequential()
model3.add(Embedding(vocab_size, 100))
model3.add(GRU(100))
model3.add(Dropout(0.5))
model3.add(Dense(2, activation='sigmoid'))
model3.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['acc'])
history3 = model3.fit(x_train, y_train, batch_size=32, epochs=3, validation_split=0.1)

결과

 

세번째 제출

model5 = Sequential()
model5.add(Embedding(vocab_size, 100))
model5.add(GRU(64))
model5.add(Dropout(0.5))
model5.add(Dense(2, activation='sigmoid'))
model5.compile(loss='binary_crossentropy', optimizer='adam', metrics=['acc'])
model5.summary()

결과

 

뭔가 lightGBM의 최적의 하이퍼파라미터를 찾는 방법 중의 하나인 GridSearchCV 같은 방법이 없을까하다가

그냥 한번 만들어서 사용해보기로 했습니다.

from keras import optimizers

params_list = {
    "batch_size":[8, 16, 32],
    "learning_rate":[0.001, 0.005, 0.01],
    "epochs":[1, 2, 3]
}

from tqdm import tqdm

def getBestParams(params_list, gru_hidden, embedding):
  count = 0
  histories = []

  my_accs = []
  my_batch = []
  my_lr = []
  my_epoch = []

  for i in tqdm(range(len(params_list['batch_size']))):
    for j in range(len(params_list['learning_rate'])):
      for k in range(len(params_list['epochs'])):
        batch_size = params_list['batch_size'][i]
        learning_rate = params_list['learning_rate'][j]
        epoch = params_list['epochs'][k]

        print(str(count) + "th train")
        print("batch_size : {} / lr : {} / epoch : {}".format(batch_size, learning_rate, epoch))

        adam = optimizers.Adam(lr=learning_rate, decay=0.1)

        accs = []
        for l in range(3):
          model = Sequential()
          model.add(Embedding(vocab_size, embedding))
          model.add(GRU(gru_hidden))
          model.add(Dropout(0.5))
          model.add(Dense(2, activation='sigmoid'))
          model.compile(loss='binary_crossentropy', optimizer=adam, metrics=['acc'])
          history = model.fit(x_train_new, y_train_new, batch_size=batch_size, epochs=epoch, validation_split=0.1)
          # model_name = "./day14_model/day14" + str(count)+"th_" + str(gru_hidden) + '_' + str(Embedding) +"model.h5"
          # model.save(model_name)
          acc = model.evaluate(x_valid_new, y_valid_new)[1]*100
          accs.append(acc)
        my_acc = np.array(accs).mean()
        print("{}th mean acc : {}%".format(i, my_acc))

        my_accs.append(my_acc)
        my_batch.append(batch_size)
        my_lr.append(learning_rate)
        my_epoch.append(epoch)


        # histories.append(history)
        
        # result_dict[count] = {"acc":my_acc, "batch_size":batch_size, "learning_rate":learning_rate, "epoch":epoch}
        # count = count + 1

  my_result_df = pd.DataFrame({"batch_size":my_batch, "learning_rate":my_lr, "epoch":my_epoch, "accuracy":my_accs})
  csv_name = "./day14_model/day14" + str(gru_hidden) + '_' + str(Embedding) + "" +"model.csv"
  my_result_df.to_csv()
    
  return my_result_df

파라미터를 입력해주면 모든 파라미터를 모델로 생성하여

세번씩 실행해보고 평가하여 나온 정확도를 평균내어 

결과를 DataFrame형식으로 return 해주도록 하였습니다.

 

여기서 평가를 위한 데이터가 따로 존재하지 않으므로 

from sklearn.datasets import load_iris # 샘플 데이터 로딩
from sklearn.model_selection import train_test_split

x_train_new, x_valid_new = train_test_split(x_train, test_size=0.1, shuffle=False, random_state=34)
y_train_new, y_valid_new = train_test_split(y_train, test_size=0.1, shuffle=False, random_state=34)

학습데이터에서 10%의 데이터를 추출하여 만들었습니다.

 

실제로 해보면

params_list4 = {
    "batch_size":[32],
    "learning_rate":[0.005, 0.01,  0.03, 0.05],
    "epochs":[1, 2, 3]
}

my_result_df_4 = getBestParams(params_list4, 128, 128)
my_result_df_4

다음과 같은 결과를 얻을 수 있었습니다.

 

여러 파라미터를 테스트한 결과를 하나로 합치고 정렬하여

가장 좋은 결과를 낸  조합의 파라미터를 사용하여 결과를 도출하고 제출해보았습니다.

my_result_df_8 = getBestParams(params_list8, 32, 100)
my_result_df_8

my_result_df_8_1 = my_result_df_8.sort_values(by=['accuracy'], axis=0, ascending=False)
my_result_df_8_1

 

 

네번째 제출

adam = optimizers.Adam(lr=0.05, decay=0.1)

model_1 = Sequential()
model_1.add(Embedding(vocab_size, 100))
model_1.add(GRU(32))
model_1.add(Dropout(0.5))
model_1.add(Dense(2, activation='sigmoid'))
model_1.compile(loss='binary_crossentropy', optimizer=adam, metrics=['acc'])
history = model_1.fit(x_train_new, y_train_new, batch_size=16, epochs=1, validation_data=(x_valid_new, y_valid_new))

결과

 

다섯번째 제출

adam1 = optimizers.Adam(lr=0.03, decay=0.1)

model_2 = Sequential()
model_2.add(Embedding(vocab_size, 100))
model_2.add(GRU(32))
model_2.add(Dropout(0.5))
model_2.add(Dense(2, activation='sigmoid'))
model_2.compile(loss='binary_crossentropy', optimizer=adam1, metrics=['acc'])
history = model_2.fit(x_train_new, y_train_new, batch_size=16, epochs=1, validation_data=(x_valid_new, y_valid_new))

결과