Packed Sequences about meshgpt-pytorch HOT 3 CLOSED

I'm not the maintainer of this repository, but I'm curious what you mean by packed sequences.

from meshgpt-pytorch.

Typically when dealing with graphs or other heterogeneous data structures (like meshes) we represent things as a packed sequence, which is just flattened over the batches. You basically store an index table and flatten the sequence length into (batch * seq_len, features). And only form pads when you do vertex dependent stuff that needs to be aligned.

from meshgpt-pytorch.

@Algomancer hey Adam! only because i've never used it

this is actually my first time using graph convs

i have another project where i will need to use graph convs, so maybe i will give packed sequences a try there and once it is working and if it is more elegant (in handling the offsets for the face edges) port it over here. kind of "if it aint broke don't fix it" mood at the moment

from meshgpt-pytorch.