search

facebookresearch / SparseConvNet

Submanifold sparse convolutional networks
https://github.com/facebookresearch/SparseConvNet
Other
2.04k stars 332 forks source link

About the parameters of InputLayer #228

Closed Hazeliii closed 2 years ago

Hazeliii commented 2 years ago

Hi Benjamin, thanks for your release. I couldn't figure out the parameters of the InputLayer, especially the _spatialsize. Would you please tell me what's the meaning of these parameters? Is it related to the output ? Thanks.

ge75buc commented 2 years ago

Hey Hazeliii,

could you figure out what spatial_size is referencing in the parameters of InputLayer? I am also not completely grasping the meaning.

Kind regards

yxKryptonite commented 2 years ago

Me neither. Have anyone of you figured out what spatialSize means? I'm really confused.

btgraham commented 2 years ago

A dense ConvNet operates on tensors with shape (B,C,H,W) where B is the batch size, C is the number of channels, and HxW is the spatial size of the input. The input_layer in https://github.com/facebookresearch/SparseConvNet/blob/main/README.md#hello-world takes the spatial size as a parameter, to know how big the input grid will be. The input spatial size is typically 2 or 3 dimensional. It can be very large is 230 x 230 x 2**30 if the input is very sparse,