ai4iacts / hexagdly Goto Github PK

I wonder if its possible to add a transposed convolution layer? My understanding is that it should not be very different to the convolution layers, but there might be some tricky details.

I'm wondering whether HexagDLy could be used for processing of spherical images by interpolating the weights of networks trained on perspective images.

In Orientation-Aware Semantic Segmentation on Icosahedron Spheres it is mentioned that "since our kernels operate on the tangent of the sphere, standard feature weights, pretrained on perspective data, can be directly transferred with only small need for weight refinement".

Same in Tangent Images for Mitigating Spherical Distortion
a similar idea is used "we show that we can transfer networks trained on perspective images to spherical data without fine-tuning"

Seeing Hexagonal Convolutional Neural Networks for
Hexagonal Grids, I would assume we can also sample, interpolate and realigned for efficient processing ?

Has there been experiments swapping nn.Conv2D with hexagdly.Conv2d with the "transfer" of convolution kernels weights from one euclidian to hexagonally sampled manifolds ?

colomuns -> columns?

Hello,

Any update regarding the optimization procedure. it should definitely attract more experience.

Hello there,

I think you are into something.
I am sure it should outperform every other kernels.

The only issue is the speed.
It would be interesting to have a highly low level implementation of this in Cuda or Cudnn, or even Tensor Comprehension.

Best.
T.C