Pre_compute spatial indices for convolutions, etc .. possible ?

[tchaton]

Dear All,

In the case where the model is known and the data aren't going to be augmented,
will it be possible to pre_compute the spatial locations for the convolution, etc.. , store them with their associated data, and reload them directly during training time in order to speed up the process ?
If yes, do you have any examples ?

Best,
Thomas Chaton

[chrischoy]

It is possible, but it is not implemented. This requires serialization and I'm planning to implement it in the future releases.

[tchaton]

Hey @chrischoy, Sounds great ! What is the timeline for it ? And do you requiere any help ?

[chrischoy]

I'm working on other stuffs, so I am aiming for mid March for the release. For that, I need to implement serialization and deserialization of 1. coords key, 2. coords map, 3. coords manager, 4. and pybind11 porting all these. If you can contribute to any part of the serialization, that would be great!