머신러닝스터디/2016/2016_07_09

머신러닝스터디/2016
머신러닝스터디/2016/목차

내용 ¶

Embedding에는 word index가 필요함.
- 초기에 Tokenizer로 word frequency를 input으로 썼는데 학습이 잘 안됨.
- http://keras.io/layers/embeddings/

tokenizer = Tokenizer(nb_words=1000)
X_train = tokenizer.sequences_to_matrix(X_train, mode="freq")

optimizer
- adamax 를 썼는데 accuracy가 50% 대에 머무름
  - tensorflow는 adamax를 제공하지 않음. keras 자체 구현됨(code).
적절한 batch size
- batch size가 너무 작으면(e.g. 32) 학습이 오래 걸린다.
- 반면 너무 크면 메모리를 많이 사용하게 된다.

[edit]

코드 ¶

import keras
import numpy as np
from keras.datasets import imdb
from keras.preprocessing.text import Tokenizer
from keras.models import Sequential
from keras.layers import Dense, Dropout, Embedding, LSTM

(X_train, y_train), (X_test, y_test) = imdb.load_data(nb_words=1000)

from keras.preprocessing.sequence import pad_sequences
X_train = pad_sequences(X_train, 1000)
X_test = pad_sequences(X_test, 1000)

model = Sequential()
model.add(Embedding(1000, 64, input_length=1000))
model.add(LSTM(output_dim=32, activation='sigmoid', inner_activation='hard_sigmoid'))
model.add(Dense(16, activation="relu"))
model.add(Dropout(0.5))
model.add(Dense(8, activation="relu"))
model.add(Dropout(0.5))
model.add(Dense(1, activation="sigmoid"))

model.compile(loss="binary_crossentropy", optimizer="adagrad", metrics=["accuracy"])

model.fit(X_train, y_train, batch_size=500, nb_epoch=100)
model.evaluate(X_test, y_test, batch_size=1000)
pred = model.predict(X_test, batch_size=20000)

print (pred[0], y_test[0])
print (pred[1], y_test[1])
print (pred[2], y_test[2])

[edit]

Padding ¶

pad_sequences은 배열의 길이가 다를 때 특정값을 채워넣어 길이를 맞춘다.

X_train = pad_sequences(X_train, 1000)

위의 코드는 X_train의 인풋 배열중 1000보다 길이가 짧은 배열에 0을 채워넣는다.
그러나 1000보다 더 긴 배열을 줄여주진 않는다.

(X_train, y_train), (X_test, y_test) = imdb.load_data(nb_words=1000)

다음과 같이 input length가 될 단어 인덱스 길이를 1000으로 제한해야 한다.

결과 예시

array([  0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0,   1,  20,  28, 716,  48, 495,  79,  27, 493,   8,   2,
         7,  50,   5,   2,   2,  10,   5, 852, 157,  11,   5,   2,   2,
        10,   5, 500,   2,   6,  33, 256,  41,   2,   7,  17,  23,  48,
         2,   2,  26, 269, 929,  18,   2,   7,   2,   2,   8, 105,   5,
         2, 182, 314,  38,  98, 103,   7,  36,   2, 246, 360,   7,  19,
       396,  17,  26, 269, 929,  18,   2, 493,   6, 116,   7, 105,   5,
       575, 182,  27,   5,   2,   2, 130,  62,  17,  24,  89,  17,  13,
       381,   2,   8,   2,   7,   5,   2,  38, 325,   7,  17,  23,  93,
         9, 156, 252,  19, 235,  20,  28,   5, 104,  76,   7,  17, 169,
        35,   2,  17,  23,   2,   7,  36,   2, 934,  56,   2,   6,  17,
       891, 214,  11,   5,   2,   6,  92,   6,  33, 256,  82,   7], dtype=int32)

nb_words로 배열의 최대 길이를 지정하지 않으면 Embedding 단계에서 out of index 에러가 난다.

IndexError: index 4414 is out of bounds for size 1000

[edit]

학습 실패 ¶

Using Theano backend.
Epoch 1/10
22500/22500 [==============================] - 115s - loss: 0.6932 - acc: 0.5014     
Epoch 2/10
22500/22500 [==============================] - 115s - loss: 0.6932 - acc: 0.5010     
Epoch 3/10
22500/22500 [==============================] - 114s - loss: 0.6932 - acc: 0.5014     
Epoch 4/10
22500/22500 [==============================] - 115s - loss: 0.6931 - acc: 0.5014     
Epoch 5/10
22500/22500 [==============================] - 115s - loss: 0.6931 - acc: 0.5014     
Epoch 6/10
22500/22500 [==============================] - 115s - loss: 0.6932 - acc: 0.5014     
Epoch 7/10
22500/22500 [==============================] - 114s - loss: 0.6931 - acc: 0.5014     
Epoch 8/10
22500/22500 [==============================] - 114s - loss: 0.6932 - acc: 0.5016     
Epoch 9/10
22500/22500 [==============================] - 115s - loss: 0.6932 - acc: 0.5014     
Epoch 10/10
22500/22500 [==============================] - 115s - loss: 0.6932 - acc: 0.5014

[edit]

학습 성공 ¶

Epoch 1/10
22500/22500 [==============================] - 14282s - loss: 0.6927 - acc: 0.5164     
Epoch 2/10
22500/22500 [==============================] - 10235s - loss: 0.6864 - acc: 0.5618     
Epoch 3/10
22500/22500 [==============================] - 3236s - loss: 0.6541 - acc: 0.6508     
Epoch 4/10
22500/22500 [==============================] - 3230s - loss: 0.5829 - acc: 0.7528     
Epoch 5/10
22500/22500 [==============================] - 3222s - loss: 0.5490 - acc: 0.7745     
Epoch 6/10
22500/22500 [==============================] - 3229s - loss: 0.5250 - acc: 0.7946     
Epoch 7/10
22500/22500 [==============================] - 3230s - loss: 0.5052 - acc: 0.8030     
Epoch 8/10
22300/22500 [============================>.] - ETA: 28s - loss: 0.4963 - acc: 0.8046

(사실 다음날 보니 프로세스가 죽어있어서 Epoch 8/10 이후의 결과는 없음... 학습 실패임)

[edit]

다음 시간에는 ¶

Coursera 동영상 week 7 보기

[edit]

머신러닝스터디/2016/2016_07_09

내용 ¶

코드 ¶

Padding ¶

학습 실패 ¶

학습 성공 ¶

다음 시간에는 ¶

더 보기 ¶