CODEWITHSUNDEEP
pythontensorflow
Sergan89
Sergan89

Reputation: 75

Multi-hot labels encoding

I'm new to Tensorflow. I have a image dataset with several labels for one image. As far as I understand, I need to use tf.losses.sigmoid_cross_entropy(). I tried to apply tf.one_hot to labels but when I try to pass them into loss function I get error, shapes incompatible. How can I fix this?

Upvotes: 1

Views: 1408

Answers (1)

Sharky
Sharky

Reputation: 4533

You're right about tf.losses.sigmoid_cross_entropy. All you need to do is wrap tf.one_hot with tf.reduce_max to reduce dimensionality like this. 

tf.reduce_max(tf.one_hot(labels, num_classes, dtype=tf.int32), axis=0)

That should return tensor of shape (num_classes,), exactly what is needed for your loss function.

Upvotes: 4

Related Questions