DNN Regressionの自分の実装がダメな件

Thu, 08 Sep 2016 21:00:46 +0919

TensorflowのDNN Regressionを使っていて便利なんだけどハイレベルなAPIばっかり使っていると自分でもっと別なの実装したくなった時に困るだろうなぁと、自分で実装してみたら…

$python boston.py
 RF: R2: 0.736624, RMSE:4.631010
SVM: R2: 0.515467, RMSE:6.281300
TFL: R2: 0.647992, RMSE:5.353827
 MY R2: -1.007581, RMSE:12.785702

精度悪すぎ… sofでも同じような質問を見つけたが同業だろう…ｗ

ログを取ってみたところ、どうも収束が悪いっぽい。

import numpy as np
import tensorflow as tf
from sklearn import cross_validation
from sklearn import datasets
from sklearn.svm import SVR
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_squared_error
from sklearn.metrics import r2_score
from math import sqrt
import random

tf.logging.set_verbosity(tf.logging.ERROR)

def test_svm(x_train, x_test, y_train, y_test):
    clf = SVR(kernel='linear').fit(x_train, y_train)
    y_pred = clf.predict(x_test)
    r2 = r2_score(y_test, y_pred)
    rmse = sqrt(mean_squared_error(y_test, y_pred))
    print('SVM: R2: {0:f}, RMSE:{1:f}'.format(r2, rmse))

def test_rf(x_train, x_test, y_train, y_test):
    rlf = RandomForestRegressor().fit(x_train, y_train)
    y_pred = rlf.predict(x_test)
    r2 = r2_score(y_test, y_pred)
    rmse = sqrt(mean_squared_error(y_test, y_pred))
    print(' RF: R2: {0:f}, RMSE:{1:f}'.format(r2, rmse))

def test_tf_learn_dnn(x_train, x_test, y_train, y_test, steps=10000, hidden=[20, 20]):
    feature_columns = [tf.contrib.layers.real_valued_column("", dimension=13)]
    tfl = tf.contrib.learn.DNNRegressor(hidden_units=hidden, feature_columns=feature_columns)
    tfl.fit(x=x_train, y=y_train, steps=steps)
    y_pred = tfl.predict(x_test)
    r2 = r2_score(y_test, y_pred)
    rmse = sqrt(mean_squared_error(y_test, y_pred))
    print('TFL: R2: {0:f}, RMSE:{1:f}'.format(r2, rmse))

def inference(x):
    hidden1 = 20
    hidden2 = 20

    with tf.name_scope("l1") as scope:
        w1 = tf.Variable(tf.truncated_normal([13, hidden1]), name="w1")
        b1 = tf.Variable(tf.zeros([hidden1]), name="b1")
        h1 = tf.nn.relu(tf.matmul(x, w1) + b1)
        #h1 = tf.nn.dropout(h1, 0.9)

    with tf.name_scope("l2") as scope:
        w2 = tf.Variable(tf.truncated_normal([hidden1, hidden2]), name="w1")
        b2 = tf.Variable(tf.zeros([hidden2]), name="b2")
        h2 = tf.nn.relu(tf.matmul(h1, w2) + b2)
        #h2 = tf.nn.dropout(h2, 0.9)

    with tf.name_scope("l3") as scope:
        w3 = tf.Variable(tf.truncated_normal([hidden2, 1]), name="w2")
        b3 = tf.Variable(tf.zeros([1]), name="b3")
        y = tf.matmul(h2, w3) + b3
    return y

def loss(model, y_):
    return tf.reduce_mean(tf.square(tf.sub(model, y_)))

def training(loss, rate):
    return tf.train.AdagradOptimizer(rate).minimize(loss)

def test_my_dnn(x_train, x_test, y_train, y_test, batch_size=32, epoch=10000, shuffle=True):
    max_size = x_train.shape[0]
    n = max_size - batch_size
    idx = list(range(x_train.shape[0]))
    x = tf.placeholder(tf.float32, shape=[None, 13])
    y_ = tf.placeholder(tf.float32, shape=[None])

    model = inference(x)
    loss_value = loss(model, y_)
    train_op = training(loss_value, 0.1)

    init = tf.initialize_all_variables()
    sess = tf.Session()
    sess.run(init)

    for e in range(epoch):
        if shuffle:
            random.shuffle(idx)
            x_train = x_train[idx]
            y_train = y_train[idx]
        for i in range(n / batch_size):
            batch = batch_size * i
            x_train_b = x_train[batch:batch + batch_size]
            y_train_b = y_train[batch:batch + batch_size]
            _, l = sess.run([train_op, loss_value], feed_dict={x: x_train_b, y_: y_train_b})
        #if e % 100 == 0:
        #    print e, l

    y_pred = sess.run(model, feed_dict={x: x_test})
    y_pred = y_pred.T[0]
    r2 = r2_score(y_test, y_pred)
    rmse = sqrt(mean_squared_error(y_test, y_pred))
    print(' MY R2: {0:f}, RMSE:{1:f}'.format(r2, rmse))

if __name__ == "__main__":
    boston = datasets.load_boston()
    x_train, x_test, y_train, y_test = cross_validation.train_test_split(
        boston.data, boston.target, test_size=0.2, random_state=0)

    test_rf(x_train, x_test, y_train, y_test)
    test_svm(x_train, x_test, y_train, y_test)
    test_tf_learn_dnn(x_train, x_test, y_train, y_test)
    test_my_dnn(x_train, x_test, y_train, y_test, shuffle=True, epoch=30000)

というわけでソースコードを読んでみたんだけど、わかりにくかったので途中までしか追えてなく、結局原因がつかめていない状態。

以下メモ

実際に使われているクラスはDNNLinearCombinedRegressorで、名前の由来がわからなかったけど、TensorFlow Wide & Deep Learning Tutorialを読めばわかる
fitで訓練するんだけど、_get_input_fnで入力、出力のplaceholderを返すinput_fnとバッチ用のデータセットを返すfeed_fnという関数が返される。
バッチは_get_input_fnでランダムシャッフルされているし、デフォルトのサイズは32
デフォルトのoptimizerはAdaGrad
clip-gradientsというオプションがあるのだがどこで使われているのかわからなかった

DNNをRandom Forest (RF)やSupport Vector Machine (SVM)と比較したい

Thu, 11 Aug 2016 09:07:30 +0919

TensorFlowのDNNチュートリアルだとトレーニングセットとテストセットをファイルから読みだすので、実用的にはちょっと面倒くさい。scikit-learnのよろしく分割してくれるメソッド使ったほうが楽でしょう。

またこScikit-learnとTensorFlowを組み合わせることでそれぞれのアルゴリズムの精度を比較することが簡単にできるので便利。

import tensorflow as tf
import numpy as np
from sklearn import datasets
from sklearn import svm
from sklearn.ensemble import RandomForestClassifier
from sklearn import cross_validation

iris = datasets.load_iris()
x_train, x_test, y_train, y_test = cross_validation.train_test_split(iris.data, iris.target, test_size=0.4, random_state=0)

classifier = tf.contrib.learn.DNNClassifier(hidden_units=[10, 20, 10], n_classes=3)
classifier.fit(x=x_train, y=y_train, steps=200)
dnn_accuracy_score = classifier.evaluate(x=x_test, y=y_test)["accuracy"]
print('DNN Accuracy: {0:f}'.format(dnn_accuracy_score))

clf = svm.SVC(kernel='linear').fit(x_train, y_train)
svm_accuracy_score = clf.score(x_test, y_test)
print('SVM Accuracy: {0:f}'.format(svm_accuracy_score))

rlf = RandomForestClassifier().fit(x_train, y_train)
rf_accuracy_score = rlf.score(x_test, y_test)
print('RF Accuracy: {0:f}'.format(rf_accuracy_score))

scikit-learn / Drkcore

DNN Regressionの自分の実装がダメな件

DNNをRandom Forest (RF)やSupport Vector Machine (SVM)と比較したい