Tensorflow索引

Question

I would like to choose a tensor by range of rows with all columns.

Something like:

x[10:20,:]

Rows 10 to 20 with all columns.

I have tried to use:

tf.gather_nd

What is the way to do that?

Answer 1

Tensorflow supports numpy style indexing: x[10:20,:] .

An example:

 x = tf.Variable(tf.truncated_normal(shape=(100, 100)))
 y = x[10:20,]

 sess = tf.InteractiveSession()
 tf.global_variables_initializer().run()
 y.eval().shape
 #output
 #(10, 100)

Answer 2

I believe you are looking for tf.slice

So assuming you have a 2D tensor as your example seems to indicated, to get 10:20 of the first dimension you would do: tf.slice(x, begin = [10,0], size = [10, x.get_shape().as_list()[1]])

Note: begin is 0 indexed and and size is 1 indexed. So the size I put will give you dimension 1 of length 10 from starting/begin point and all of dimension 2.