What happened?

Loss of precision is observed in several Heat functions when using the float64 data type. This is due to pre-conversion to float32 (the pytorch default floating data type) before converting the resulting tensor back to float64.

The following functions are affected:

• arange
• array
• linspace
• abs

Code snippet

>>> import heat as ht
>>> ht.arange(16777217.0, 16777218, 1, dtype=ht.float64)
DNDarray([16777216.], dtype=ht.float64, device=cpu:0, split=None)
>>> ht.array(16777217.0, dtype=ht.float64)
DNDarray(16777216., dtype=ht.float64, device=cpu:0, split=None)

Version

main (development branch)

Python version

3.10

PyTorch version

1.11