An attempt to combine dense, depthwise and groupwise conv through DenseConvDims
#146
Conversation
|
Cool @arhik! To make it easier for you, I just merged #142 so that you can more easily rebase. Here are my thoughts:
|
…interface `DenseConvDims`
| @eval begin | ||
| function $(Symbol("∇conv_filter$(backend)"))( | ||
| x::AbstractArray{xT,N}, dy::AbstractArray{yT,N}, | ||
| cdims::ConvDims; kwargs...) where {xT, yT, N} | ||
| dw = similar(dy, kernel_size(cdims)..., channels_in(cdims), | ||
| dw = similar(dy, kernel_size(cdims)..., div(channels_in(cdims),group_count(cdims)), |
There was a problem hiding this comment.
It would be helpful to have a succinct description of what to do when groupcount is not equal to 1 or input_channels(). E.g. what does groupcount == 7 look like? If you can succinctly describe what should happen, I can help you adjust the implementations of the direct and im2col implementations.
@staticfloat. This is only change. Only weights dimensions shrink as in here by groupcount value in third axis. if groupcount == 1 Nothing changes. When groupcount == 2 (lets say), then one group of weights operate only on the half of the input channels. and produce only one output channel. These weights groups have to be occupied across all the channels and we will have to use new group of weights (occupying all input channels in blocks) until their output matches output number of channels.
There was a problem hiding this comment.
When groupcount == 7; we should make sure input channels can be divided exactly into 7 groups. Then we should check if output channels are multiple of groupcount(Since one group can only produce one output).
There was a problem hiding this comment.
We should then distribute these 7 groups to operate on different input channels in blocks of div(input_channels(), 7)
…and removing seperate implementations of Depthwise and Groupwise.
arhik
force-pushed
the
common_convdims
branch
from
23f0976 to
14da5b5
Compare
|
Can this PR be moved to https://github.com/FluxML/NNlibCUDA.jl, or has it diverged too much to be workable there? |
With #JuliaGPU/CuArrays.jl#521 (comment) and
JuliaGPU/CuArrays.jl#523 (comment)
We can define depthwise and groupwise convolutions using DenseConvDims (Naming will become confusing then).
directandim2colbackend codes.@staticfloat Please feel to comment on this. Open to any changes or suggestions. I will do my consider your changes in #142.