Deeplearningで誤差を確認しながら分類問題を解かせる

やること

単純な分類問題。誤差の減り具合をプロットして、オーバーフィッテングしないように、バッチサイズも徐々に狭めていく。

各ステップ詳細

データをロード、多項式特徴量を追加

# load data
data_train = pd.read_csv('./training_data.csv')
data_tornm = pd.read_csv('./test_data.csv')
X, y = np.array(data_train.ix[:,:-1]), np.array(data_train.ix[:,-1])

# extend data
from sklearn.preprocessing import PolynomialFeatures
X_p = PolynomialFeatures(2).fit_transform(X)

モデルを定義

# create the model
model = Sequential()
model.add(Dense(128, input_dim=X_p.shape[1], activation='relu'))
model.add(Dense(16, activation='linear'))
model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['accuracy'])

今回は2クラス分類だから最終出力は1。

コスト関数の減少具合をプロット

class LossHistory(Callback):
    def on_train_begin(self, logs={}):
        self.losses = []
    def on_batch_end(self, batch, logs={}):
        self.losses.append(logs.get('loss'))

losshist = LossHistory()

定義したモデルの訓練

print(model.summary())
model.fit(X_train, y_train, nb_epoch=5, batch_size=500, verbose=2, callbacks=[losshist])
scores = model.evaluate(X_test, y_test, verbose=2)

実行すると

____________________________________________________________________________________________________
Layer (type)                     Output Shape          Param #     Connected to                     
====================================================================================================
dense_9 (Dense)                  (None, 128)           32512       dense_input_4[0][0]              
____________________________________________________________________________________________________
dense_10 (Dense)                 (None, 16)            2064        dense_9[0][0]                    
____________________________________________________________________________________________________
dense_11 (Dense)                 (None, 1)             17          dense_10[0][0]                   
====================================================================================================

のような出力がされる。モデルに含まれるパラメータを確認しながら、解きたい問題に対して複雑すぎ、単純すぎないかを確認する。

ループごとに精度改善を確かめながらバッチサイズを減らす

LOOP = 100
losses, rocaucs = [], []
for i in range(LOOP):
    # train
    batch_size = 200*(LOOP-i+1)/LOOP
    model.fit(X_train, y_train, nb_epoch=10, batch_size=batch_size, verbose=2, callbacks=[losshist])
    scores = model.evaluate(X_test, y_test, verbose=2)

    # roc auc
    pred = model.predict(X_test)
    rocaucs += [roc_auc_score(y_test, pred)]
    print 'logloss:', log_loss(y_test, pred), 'rocauc', roc_auc_score(y_test, pred)
    
    # losshist
    losses += list(losshist.losses)
    plt.figure(figsize=(10, 3))
    plt.subplot(131)
    plt.plot(losses)
    plt.title('score:'+str(scores[1])[:5]+' steps:'+str(i)\
              +'  batchsize:'+str(batch_size)[:5])
    
    plt.subplot(132)
    plt.plot(losshist.losses)
    
    plt.subplot(133)
    plt.plot(rocaucs)
    plt.show()

かなりごり押しだけどバッチサイズを200から徐々に小さくしながら

誤差を確認していく。 左から順に、これまでの全ステップでの損失関数の出力、最近10エポックでの損失関数の出力、ROCAUCの推移。この場合は完全にデタラメな予測をしてしまっているので(ROCAUC= 0.5近くのまま)前処理から考え直す必要がありそう。

コード

from keras.models import Sequential
from keras.layers import Dense
from keras.preprocessing import sequence
import tensorflow as tf
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt;plt.style.use('ggplot')

from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split
from keras.layers import Dropout
from keras.callbacks import Callback
from sklearn.metrics import log_loss
%matplotlib inline


data_train = pd.read_csv('./_training_data.csv')
data_test = pd.read_csv('./test_data.csv')
X, y = np.array(data_train.ix[:,:-1]), np.array(data_train.ix[:,-1])

# extend data
from sklearn.preprocessing import PolynomialFeatures
X_p = PolynomialFeatures(2).fit_transform(X)

# callback
class LossHistory(Callback):
    def on_train_begin(self, logs={}):
        self.losses = []
    def on_batch_end(self, batch, logs={}):
        self.losses.append(logs.get('loss'))

# split data
X_train, X_test, y_train, y_test = train_test_split(X_p, y, test_size=0.05, random_state=42)

# create the model
model = Sequential()
model.add(Dense(128, input_dim=X_p.shape[1], activation='relu'))
model.add(Dense(16, activation='linear'))
model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='rmsprop',
              loss='binary_crossentropy',
              metrics=['accuracy'])

print(model.summary())
losshist = LossHistory()
model.fit(X_train, y_train, nb_epoch=5, batch_size=500, verbose=2, callbacks=[losshist])
scores = model.evaluate(X_test, y_test, verbose=2)

LOOP = 100
EPOCH = 10

losses, rocaucs = [], []
for i in range(LOOP):
    # train
    # 少しずつバッチサイズを小さくしていく
    batch_size = 200*(LOOP-i+1)/LOOP
    model.fit(X_train, y_train, nb_epoch=EPOCH, batch_size=batch_size, verbose=2, callbacks=[losshist])
    scores = model.evaluate(X_test, y_test, verbose=2)

    # roc auc
    pred = model.predict(X_test)
    rocaucs += [roc_auc_score(y_test, pred)]
    print 'logloss:', log_loss(y_test, pred), 'rocauc', roc_auc_score(y_test, pred)
    
    # losshist
    losses += list(losshist.losses)
    plt.figure(figsize=(10, 3))
    plt.subplot(131)
    plt.plot(losses)
    
    plt.subplot(132)
    plt.plot(losshist.losses)
    
    plt.subplot(133)
    plt.plot(rocaucs)
    plt.show()