皮尔逊相关作为损失函数

时间:2017-10-22 11:05:32

标签: optimization machine-learning tensorflow neural-network pearson-correlation

我正在训练TF前馈网络,我的目标是产生从0到1的预测,这些预测尽可能接近目标分数。单个训练实例由大约450个特征组成,数据集中大约有1500个示例。我在我的网络中使用4层,每层都有一个Relu激活,然后最后的'out'层有一个sigmoid激活。当我使用MSE作为损失函数时,我得到了不错的(但不是最优的)结果。我试图使用以下作为损失函数:

# Define loss and optimizer
#pearson correlation as loss function
length = 443

#apply regularization (l2)
Beta = 0.01
regularizer = tf.nn.l2_loss(weights['h1']) +   
tf.nn.l2_loss(weights['h2']) + tf.nn.l2_loss(weights['h3']) + 
tf.nn.l2_loss(weights['h4'])

#used to report correlation 
pearson = tf.contrib.metrics.streaming_pearson_correlation(intensity, 
Y, name="pearson")


#pearson corr. as loss?
# multiply by -1 to maximize correlation i.e. minimize negative 
correlation 
original_loss = -1 * length * tf.reduce_sum(tf.multiply(intensity, Y)) 
- (tf.reduce_sum(intensity) * tf.reduce_sum(Y))
divisor = tf.sqrt(
        (length * tf.reduce_sum(tf.square(intensity)) - 
tf.square(tf.reduce_sum(intensity)))) *\
        tf.sqrt(
        length * tf.reduce_sum(tf.square(Y)) - 
tf.square(tf.reduce_sum(Y)))

 loss_op = tf.truediv(original_loss, divisor)
 loss_op = tf.reduce_mean(loss_op + Beta * regularizer)
#Init optimizer
 optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, 
 epsilon = 1e-09)
 train_op = optimizer.minimize(loss_op) 

该想法是最小化负相关,即最大化正相关。然而,经过对超参数的大量实验,这仍然给我“纳”错误并报告'nan'Pearson相关性。关于为什么会这样的任何想法?

1 个答案:

答案 0 :(得分:0)

请注意,tf.contrib.metrics.streaming_pearson_correlation()返回一个元组(pearson_r, update_op),因此原则上您应该能够将update_op直接输入optimizer.minimize()