Way for proper Integer array indexing?

[HuaMuLanChina]

Question

In numpy:

>>> array = np.ones((5,2))
>>> sel = np.array([1,2,4])
>>> array[sel].shape
(3, 2)

what i've tried:

NDArray array = manager.ones(new Shape(5,2));
long[] d = {1,2,4};
NDArray sel = manager.create(d);
array.get(sel)//error

[lanking520]

There is a way, try to do following the javadoc: https://javadoc.io/doc/ai.djl/api/latest/ai/djl/ndarray/index/NDIndex.html

NDIndex index = new NDIndex().addIndices(0) // pick first dim
                                                      .addPickDim(sel);
array.get(index);

[HuaMuLanChina]

import ai.djl.ndarray.NDArray;
import ai.djl.ndarray.NDManager;
import ai.djl.ndarray.types.Shape;
import ai.djl.ndarray.index.NDIndex;

NDManager manager = NDManager.newBaseManager();
NDArray array = manager.arange(10).reshape(-1,2);
long[] d = {1,2,4};
NDArray sel = manager.create(d);
NDIndex index = new NDIndex().addIndices(0).addPickDim(sel);
array.get(index);

can't pass block runner.

jdk.jshell.EvalException: get() currently supports all, fixed, and slices indices
	at ai.djl.ndarray.index.NDArrayIndexer.get(NDArrayIndexer.java:75)
	at ai.djl.ndarray.NDArray.get(NDArray.java:498)
	at .(#10:1)

[zachgk]

We don't support this right now.

For clarification, there are two types of indexing with numeric indices: pick and gather. We have support for pick, but need to add gather to support this.

[ManuLasker]

@HuaMuLanChina Hello, what you can do and it functions for me with the pytorch engine is:

NDManager manager = NDManager.newBaseManager();
NDArray array = manager.arange(10).reshape(-1,2);
long[] d = {1,2,4};
NDArray sel = manager.create(d).repeat(2).reshape(-1, 2);
NDIndex index = new NDIndex().addPickDim(sel);
array.get(index);

The results are:

ND: (3, 2) cpu() int32
[[ 2,  3],
 [ 4,  5],
 [ 8,  9],
]

The same results for numpy python:

array = np.arange(10).reshape(-1,2)
sel = np.array([1,2,4])
array[sel]
array([[2, 3],
       [4, 5],
       [8, 9]])

Hope this can help you :D

[HuaMuLanChina]

'repeat' works .