These are chat archives for thunder-project/thunder

25th
Feb 2015
Jeremy Freeman
@freeman-lab
Feb 25 2015 03:14
What do you mean by default? When loading Series data we set to float32
Davis Bennett
@d-v-b
Feb 25 2015 03:15
hmm every time I've checked my Series data have been float64
Jeremy Freeman
@freeman-lab
Feb 25 2015 03:15
If you convert to "float" that's float64, that one is inherited from numpy
Davis Bennett
@d-v-b
Feb 25 2015 03:16
maybe normalizing or detrending converts to float64?
Jeremy Freeman
@freeman-lab
Feb 25 2015 03:16
Ah, so many operations convert automatically
Yup
Including those too
*two
Davis Bennett
@d-v-b
Feb 25 2015 03:16
that might be undesirable, given that it doubles the data size..
I guess people can always add more nodes, but 2X can make the difference between a packable dataset and one that lives on platters
but dealing with this seems like a job for the user
Jeremy Freeman
@freeman-lab
Feb 25 2015 03:19
Agreed! Very tough balance here. If we cast everything to float64 it's great for type consistency and we never get overflow, but it's big. If we try to keep to float16, you can get some very unexpected results on various operations.
So yeah, we basically leave it to the user
With float32 a happy medium as the starting point