Pure python implementation of MultiScaleDeformableAttention

[bwittmann]

I would have a question regarding the cuda implementation of MultiScaleDeformableAttention. Do you also have a pure python version of MultiScaleDeformableAttention available that does not rely on c++ code?

Thanks in advance!

[kimshun09]

I found it here.

Deformable-DETR/models/ops/functions/ms_deform_attn_func.py

Line 41 in 11169a6

def ms_deform_attn_core_pytorch(value, value_spatial_shapes, sampling_locations, attention_weights):

But you should use it for debug and test only.
See #9 for more.

[bwittmann]

Hi @kimshun09,

Thank you very much for your reply! Do you by any chance know by how much the cuda implementation is faster than the python version? Is speed and memory efficiency the only reason?

BR Bastian