search

mvoelk / keras_layers

Various Additional Keras Layers: SparseConv2D, PartialConv2D, GroupConv2D, DeformableConv2D, DepthwiseConv2D, MaxPoolingWithArgmax2D, MaxUnpooling2D, AddCoords2D, Blur2D, LayerNormalization, InstanceNormalization, WeightNormalization, EqualizedLearningRates
MIT License
36 stars 16 forks source link

SparseConv2D Speed? #3

Open danielmimimi opened 3 years ago

danielmimimi commented 3 years ago

Hey, thanks for the implementation - works well, also the example. My Question is, is there a speed gain vs the common Conv2d?

mvoelk commented 3 years ago

No, the aim is to apply convolution to incomplete input data.

danielmimimi commented 3 years ago

So is it rather slower of faster than the normal Conv2 operation in processing? It was not clear in those 2 papers you mentioned, maybe you had more insight coding it?

mvoelk commented 3 years ago

There are other approaches that aim that. In the literature, the term "sparse convolution" is used in different ways, which may be confusing. I have no experience with these approaches, sorry...