Sparse Convolution Implementation about minkowskiengine

[chrischoy]

So let's go over the generalized sparse convolution quickly https://stanfordvl.github.io/MinkowskiEngine/generalized_sparse_conv.html#id2

For all latex code define inside $$, use http://mathb.in/ to render them online.

The input coordinates and output coordinates can be defined arbitrarily as $\mathcal{C}^\text{in}$ and $\mathcal{C}^\text{out}$. There's no consistent definition in here that a sparse convolution has to give out the same result as the regular convolution. From my experience, if you expand the coordinate space so that the sparse convolution gives the same result as the standard discrete dense convolution (regular conv2d or conv3d), the computation increases rapidly while it gives no performance increase. So I would not go that route unless you have a specific reason for that.

1. By default, it uses the same input coordinates for output $\mathcal{C}^\text{out} = \mathcal{C}^\text{in}$.

2. As I mentioned, I would not go that route. If you need, I would simply use .dense() function to convert the sparse tensor to a dense one and do convolution after you decrease the resolution of the space significantly. However, if you absolutely need that, I can make a cleaner API function.

[johnny8376]

Thanks for your quick reply.
I was using the terminology of sparse conv and submanifold conv in Submanifold Sparse Convolutional Networks. I just want to make sure I can use your library to reproduce other network.
I appreciate your generalized framework. The current API looks quit flexible to me. There's no need to include such expensive operator here.

[lshiwjx]

@chrischoy Hi, Thanks for your code.
I feel that the MinkowskiConvolutionTranspose(generate_new_coords=True) is equivalent to the Dense Convolution regardless of the negative coordinates of the output. Am I correct?

[chrischoy]

No it is not equivalent. The input and output are both sparse tensors. How could it be equivalent?

[lshiwjx]

@chrischoy Thanks for your reply.

I mean that in your implement, the input coordinates and output coordinates are the same. But for regular convolution, the out_coords_key includes in_coords_key's neighborhood.

As you have mentioned, in your experiments, expanding the output coordinate space will harm the performance, but I want to try this in my tasks where the input is very sparse.

I found that it seems the MinkowskiConvolutionTranspose(generate_new_coords=True) can achieve this function, i.e., the regular convolution, because the output coordinates of the transpose convolution are expanded. Am I correct?