TensorFlow: AttributeError: 'Tensor' object has no attribute 'log10'

Question

I have been using tensorflow with python3 support. There is line in my code which throws me an error.

The error here :

return -10. * np.log10(K.mean(K.square(y_pred - y_true)))

AttributeError: 'Tensor' object has no attribute 'log10'

I am using Keras (already installed) I have in my code.......from keras import backend as K — Surajit Saikia, Jan 24 '17 at 11:43
what does K.mean returns? it seems it returns a tensor object and not a nummpy array — Amitay Nachmani, Jan 24 '17 at 12:09
Yes, you are correct...actually i changed it to Theano backend and it started working :) — Surajit Saikia, Jan 25 '17 at 13:30
I just ran into exactly the same issue, could you elaborate on your solution please? — tohoho, Jan 25 '17 at 16:25
I figured out what you meant, changing the backend setting on the Keras config file to use Theano rather than TensorFlow. That did solve the issue (though it caused new, unrelated problems) — tohoho, Jan 25 '17 at 16:59
Can you show a bit more code? In particular, it's not obvious what `np` is and why it would be a `tf.Tensor` object. (In most example code I've seen, `np` is the alias for the `numpy` module.) — mrry, Jan 27 '17 at 00:01

score 4 · Answer 1 · answered Feb 10 '17 at 22:35

4

you can do like this:

return 10.0 * K.log(1.0 / (K.mean(K.square(y_pred - y_true)))) / K.log(10.0)

answered Feb 10 '17 at 22:35

Ansle liu

2

This works great for me, but I think you meant `-10`. Can you elaborate on why the prior code doesn't work in Tensorflow? I'm assuming it has something to do with numpy not knowing how to handle the tensorflow result. – Chris Foster Mar 11 '17 at 04:43

1 Answers1