Suppose I have one input 1D tensor, I want to get indices for unique elements in 1D tensor.
input 1D tensor
[ 1 3 0 0 0 3 5 6 8 9 12 2 5 7 0 11 6 7 0 0]
expected output
Values: [1, 3, 0, 5, 6, 8, 9, 12, 2, 7, 11]
indices: [0, 1, 2, 6, 7, 8, 9, 10, 11, 13, 15]
Here is my strategy now.
input = [ 1, 3, 0, 0, 0, 3, 5, 6, 8, 9, 12, 2, 5, 7, 0, 11, 6, 7, 0, 0,]
unique_value_in_input, _ = tf.unique(input) # [1 3 0 5 6 8 9 12 2 7 11]
number_of_unique_value = tf.shape(unique_value_in_input)[0] #11
y = tf.reshape(y, (number_of_unique_value, 1)) #[[1], [3], [0], [5], [6], [8], [9], ..]
input_matrix = tf.tile(input, [number_of_unique_value]) # repeat the tensor for tf.equal()
input_matrix = tf.reshape(input, [number_of_unique_value,-1])
cols = tf.where(tf.equal(input_matrix, y))[:,-1] #[[ 0 0] [ 1 1] [ 1 5] [ 2 6] [ 2 12] ...]
Since I will have repeat value in tf.where() step, which means I have duplicated True in result. Is there any function I can use in this issue?
You should be able to do the following and get the desired output. We do the following. For each value in unique values, you get a boolean tensor and get the maximum index (i.e only the first maximum index) through tf.argmax.
import tensorflow as tf
input = tf.constant([ 1, 3, 0, 0, 0, 3, 5, 6, 8, 9, 12, 2, 5, 7, 0, 11, 6, 7, 0, 0,], tf.int64)
unique_vals, _ = tf.unique(input)
res = tf.map_fn(
lambda x: tf.argmax(tf.cast(tf.equal(input, x), tf.int64)),
unique_vals)
with tf.Session() as sess:
print(sess.run(res))