CODEWITHSUNDEEP
tensorflow
Vedanshu
Vedanshu

Reputation: 2296

how to set MAPE cost function in tensorflow

I'm trying to implement MAPE(Mean Absolute Percentage Error) in tensorflow. I'm taking output from the RNN.

Here is my code :

loss = tf.reduce_mean(tf.abs(tf.divide(tf.subtract(outputs,y),y)))

But on running the session I'm getting MAPE as NaN. What am I doing wrong ?

Upvotes: 1

Views: 2285

Answers (2)

BigBadMe
BigBadMe

Reputation: 1842

I strongly suspect you are getting NaN because y is zero. To get around this you can add a very small epsilon value to y denominator to ensure it's never zero, but small enough not to skew your calculations:

loss = tf.reduce_mean(tf.abs(tf.divide(tf.subtract(outputs,y),(y + 1e-10))))

Upvotes: 3

Fivos Delemis
Fivos Delemis

Reputation: 1

You get nan because y includes a zero value. Therefore it is impossible tensorflow to produce the division as a result it outputs inf.

Upvotes: 0

Related Questions