Reduction of accuracy

ND4J
1

Hi all,
I have an array namely “real”. it has a vector of double numbers. I am trying to create an INDArray based on my vector.
my code is below :

double real_d = (double) ArrBD.get(0);
INDArray real = Nd4j.create(real_d);

It works but unfortunately, it converts big numbers to infinity and small numbers to zero.
Here you can see the double array:

[0.0, 5.911259568298103E-306, -1.8401529637727614E-221, 1.7562463582928743E-268, 2.7206514809021945E-133, 4.583815262762733E90, 2.5637698710586 …

and here the data of INDArray:

[0.0,0.0,-0.0,0.0,0.0,Infinity,Infinity,-0.0,-0.0 …

I appreciate your help in advance

2

correct me if im wrong but your loosing precision? each number is only allowed so much accuracy because of its memory size? is this what you mean?

1 Like
3

This is likely due to not specifying the data type. We don’t implement a “double” etc any different way than what you’d see in numpy etc.

1 Like
4

Thank you. You are right. I understood what is going behind it. I read binary file in double format and Nd4j shows it in float format during debugging, but when I call this line

data.real.data().asDouble();

it returns correct values.