search

NVIDIA / MinkowskiEngine

Minkowski Engine is an auto-diff neural network library for high-dimensional sparse tensors
https://nvidia.github.io/MinkowskiEngine
Other
2.37k stars 344 forks source link

Resnet #259

Open satabios opened 3 years ago

satabios commented 3 years ago

  1. It has been mentioned in the paper that or the first layer, instead of a 7×7 2D convolution, we use a 5×5×5×1 generalized sparse convolution. However, for the rest of the networks, we follow the original network architecture. why was a 5x5x5x1 convolution chosen?

  2. I'm trying to experiment with ResneXt on MinknowskiEngine, any guidance? I just need to introduce groups in the make_layer functions but is there any efficient way of doing so?

jac99 commented 3 years ago

From my experience, size of the first convolutional filter affects the performance. 5x5x5 filter has 125 elements whereas 7x7x7 almost 3 times more. I experimented with modified version of MinkowskiResnet architecture and when using 1x1x1, 3x3x3 or 5x5x5 first conv layers the performance was similar. For 7x7x7 it was 50% slower and for 9x9x9 over 100% slower (compared to 5x5x5) image