how to calculate the average of elements above a certain threshold in tensorflow

Question

I want to calculate the mean of the elements that are greater than a fixed threshold (here 0.8).

Task in numpy :

X = np.array([[0.11,0.99,0.70]])
print(np.nanmean(X[X>0.8]))
Out : 0.99

What is the equivalent in tensorflow without converting the tensor c to a numpy array?

Example:

c = tf.constant([[0.11,0.99,0.70]])
tf.reduce_mean(tf.where(tf.greater(c,(tf.constant(0.8, dtype=tf.float32)))))

the output is equal to 0!

Output : <tf.Tensor: shape=(), dtype=int64, numpy=0>

Answer 1

You don't need tf.greater and tf.where.

c = tf.constant([[0.11,0.99,0.70]])
# Or : tf.reduce_mean(c[c > 0.8])
tf.reduce_mean(c[c > tf.constant(0.8, dtype=tf.float32)])

You can use tf.boolean_mask as an alternative:

c = tf.constant([[0.11,0.99,0.70]])
mask = c > 0.8
tf.reduce_mean(tf.boolean_mask(tensor, mask))

Output : <tf.Tensor: shape=(), dtype=float32, numpy=0.99>

Answer 2

You could also use a boolean_mask instead of tensor indexing:

c = tf.constant([[0.11,0.99,0.70]])
tf.reduce_mean(tf.boolean_mask(c, c > 0.8))

<tf.Tensor: shape=(), dtype=float32, numpy=0.99>