`ys` dtype in _integrate_ode

[ReHoss]

This line imposes the data type torch.float32 by default. Consequently, integration with torch.float64 can't be performed because of floating point precision mix.

https://github.com/thibmonsel/torchdde/blob/b28e4ef90e2c40bcfa5fc58b5b39ddca91f79164/torchdde/integrate.py#L386

[thibmonsel]

Do you have a MWE ?

[thibmonsel]

This torch.empty doesn't impose any type since dtype=None here (https://pytorch.org/docs/stable/generated/torch.empty.html).

If you want to use float64, please set :

torch.set_default_dtype(torch.float64)

[ReHoss]

The code won't work here if the data type of ys is different from the data type of ts or y0. I came across this flaw.

I don't think the right solution is to force the user to set the tensor global type. Indeed, for instance, he could manipulate float32 data in some part of the code and float64 in another part.

In my opinion, a data type uniformity check should be done on the input arguments such as ts and y0 (also t0 and t1). Then, in the scope of this function, the data type of ys should match the data type of the input.

https://github.com/thibmonsel/torchdde/blob/b28e4ef90e2c40bcfa5fc58b5b39ddca91f79164/torchdde/integrate.py#L386

[thibmonsel]

This is intended y0 defines ys's type. Pytorch default type is float32 so you have to specify and define all your tensors accordingly

[ReHoss]

This is intended y0 defines ys's type. Pytorch default type is float32 so you have to specify and define all your tensors accordingly

I don't see where in the code it defines ys's data type. To me, it defines only the device.

[thibmonsel]

Your suggestion is to add dtype=y0.dtype ?

[ReHoss]

Your suggestion is to add dtype=y0.dtype ?

From above:

• Check data type uniformity across all input arguments at the beginning to increase informativity in case of precision mix error
• Setting ys datatype accordingly, possibly with dtype=y0.dtype

The associated unit test should capture some dtype error if the input data types are not uniform (e.g. y0 as float32 and ts as float64). It should also pass without setting the global data type with torch.set_default_dtype.

[thibmonsel]

Latest commit (ie https://github.com/thibmonsel/torchdde/commit/a659f8c62e9274a86e6bdbad722e4f2e3666f265) should fix the issue. If so please close the issue. If not please provide a MWE.

For information :

>>> import torch 
>>> a = torch.arange(10, dtype=torch.float64)
>>> b = 2 * torch.arange(10, dtype=torch.float32)
>>> c = a * b 
>>> c.dtype
torch.float64

[thibmonsel]

Closing this issue since there's no more activity.