Shape is not the same for the first row

ND4J
1 
 INDArray ones = Nd4j.ones(1, 8);
  System.out.println(ones.shapeInfoToString());
  ones = ones.getRow(0);
  System.out.println(ones.shapeInfoToString());

  ones = Nd4j.ones(2, 8);
  System.out.println(ones.shapeInfoToString());
  ones = ones.getRow(0);
  System.out.println(ones.shapeInfoToString());

Above code print below lines, why the second line is not the same as the 4th line?
Rank: 2, DataType: FLOAT, Offset: 0, Order: c, Shape: [1,8], Stride: [1,1]
Rank: 2, DataType: FLOAT, Offset: 0, Order: c, Shape: [1,8], Stride: [1,1]
Rank: 2, DataType: FLOAT, Offset: 0, Order: c, Shape: [2,8], Stride: [8,1]
Rank: 1, DataType: FLOAT, Offset: 0, Order: c, Shape: [8], Stride: [1]

2

@SidneyLann that’s likely due to it just returning the result as is. It’s already a vector. Nothing is really executed unless there’s a row to actually get. Feel free to file an issue and we can do a proper reshape call there.

3

The rank in the second line should be 1 but not 2, right?
In my app, indArray.getRow(0); return the whole INDArray when the first shape is 1, and return correct Row when the first shape is bigger then 1.

That means I can get the first row when an INDArray has multiple rows, but can’t get the first row when an INDArray has ONLY 1 row.

4

@SidneyLann why would you need to do that though? A 1 x 8 is already a row vector and by definition the result.
We generally treat 1 x 8 and 8 as the same.

If you want just reshape it for now. I already mentioned you could file an issue if you see it as a bug and I ccan take a look.

5

for (int j = 0; j < crawlIdsSize; j++) {
tmp = outArray.getRow(j);
…

The behavior must be the same, must always get [8] but not [1,8] when crawlIdsSize is 1.

6

@SidneyLann just detect when it’s length 1 for now and file an issue then. I’m not going to fix this on the spot from a forum post. I’ll do it with a quick reproducer and a corresponding issue I can reference on the tracker. I already mentioned the likely issue and a workaround a few times now.

I’m not going to do much more here.

7

Ok, will fire an issue.

8 
  INDArray ones = Nd4j.ones(1, 8);
  System.out.println(ones.shapeInfoToString());
  ones = ones.get(point(0),all());
  System.out.println(ones.shapeInfoToString());
  ones = Nd4j.ones(2, 8);
  System.out.println(ones.shapeInfoToString());
  ones = ones.getRow(0);
  System.out.println(ones.shapeInfoToString());

Rank: 2, DataType: FLOAT, Offset: 0, Order: c, Shape: [1,8], Stride: [1,1]
Rank: 1, DataType: FLOAT, Offset: 0, Order: f, Shape: [8], Stride: [1]
Rank: 2, DataType: FLOAT, Offset: 0, Order: c, Shape: [2,8], Stride: [8,1]
Rank: 1, DataType: FLOAT, Offset: 0, Order: c, Shape: [8], Stride: [1]

Why the order of the second line is f, how to get c by .get(point(0),all())?

9

@SidneyLann c vs f shouldn’t affect you here either. I would consider any of that as a distraction. As long as the results are correct it doesn’t matter.

C vs f mainly affects matrix multiplies and a few other specific ops.

C is the default ordering though. Why it returns certain orders I’d have to take a closer look at the indexing implementation again.

10

Ok. call tmp.setOrder(‘c’) if necessary.