# \`groups\` argument in Flux.Conv : a misunderstanding

**URL:** <https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641>\
**Category:** Machine Learning\
**Tags:** flux\
**Created:** [December 26, 2021, 11:17am UTC](https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641 "2021-12-26T11:17:57Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![sim](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/sim/32/22207_2.png) [@sim](https://discourse.julialang.org/u/sim)\
**Post date:** [December 26, 2021, 11:17am UTC](https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641/1 "2021-12-26T11:17:57Z")

</div>

Hello,

I want to define a convolutional layer in Flux.jl using the `groups` option. Ideally, I want to use `Conv((1,1), 3=>3, groups=3)`. For reference, the [Flux.Conv documentation](https://fluxml.ai/Flux.jl/stable/models/layers/#Flux.Conv) defines this argument as

> - Keyword `groups` is expected to be an `Int` . It specifies the number of groups to divide a convolution into.

As I understand, `Conv((1,1), 3=>3, groups=3)` should define a 1 \times 1 convolution taking as input 3 channels, outputting 3 channels, but the convolution should be layer-wise. In other words, this should be equivalent to three convolutions applied independently to each channel, in particular the number of parameters should be 3 \times (1+1)=6 since a 1=\>1 convolution has just 2 parameters.

```julia
julia> T = Conv((1,1), 3=>3, groups=3)
Conv((1, 1), 1 => 3) # 6 parameters

```

The number of parameters fits, but the in/out channels seems bizarre. And then, when I try to apply T to an array with 3 channels, I get `ERROR: AssertionError: DimensionMismatch("Data input channel count (3 vs. 3)")` (see below for the full stacktrace).

For the sake of testing, I tried to apply T to an array with 1 channel, and I also get an `ERROR: DimensionMismatch("Input channels must match! (1 vs. 1)")`.

Can someone explain how the `groups` argument works in Flux.Conv ?

Thanks !

* * *

Stacktrace with 3 channels (this should work but does not) :

```julia
julia> T(rand(10, 10, 3, 1))
ERROR: AssertionError: DimensionMismatch("Data input channel count (3 vs. 3)")
Stacktrace:
  [1] check_dims(x::NTuple{5, Int64}, w::NTuple{5, Int64}, y::NTuple{5, Int64}, cdims::DenseConvDims{3, (1, 1, 1), 3, 3, 3, (1, 1, 1), (0, 0, 0, 0, 0, 0), (1, 1, 1), false})
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/dim_helpers/DenseConvDims.jl:73
  [2] conv_direct!(y::Array{Float64, 5}, x::Array{Float64, 5}, w::Array{Float32, 5}, cdims::DenseConvDims{3, (1, 1, 1), 3, 3, 3, (1, 1, 1), (0, 0, 0, 0, 0, 0), (1, 1, 1), false}; alpha::Float64, beta::Bool)
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/impl/conv_direct.jl:51
  [3] conv_direct!
    @ ~/.julia/packages/NNlib/P9BhZ/src/impl/conv_direct.jl:51 [inlined]
  [4] conv!(y::Array{Float64, 5}, in1::Array{Float64, 5}, in2::Array{Float32, 5}, cdims::DenseConvDims{3, (1, 1, 1), 3, 3, 3, (1, 1, 1), (0, 0, 0, 0, 0, 0), (1, 1, 1), false}; kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/conv.jl:293
  [5] conv!(y::Array{Float64, 5}, in1::Array{Float64, 5}, in2::Array{Float32, 5}, cdims::DenseConvDims{3, (1, 1, 1), 3, 3, 3, (1, 1, 1), (0, 0, 0, 0, 0, 0), (1, 1, 1), false})
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/conv.jl:291
  [6] conv!(y::Array{Float64, 4}, x::Array{Float64, 4}, w::Array{Float32, 4}, cdims::DenseConvDims{2, (1, 1), 3, 3, 3, (1, 1), (0, 0, 0, 0), (1, 1), false}; kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/conv.jl:151
  [7] conv!
    @ ~/.julia/packages/NNlib/P9BhZ/src/conv.jl:151 [inlined]
  [8] conv(x::Array{Float64, 4}, w::Array{Float32, 4}, cdims::DenseConvDims{2, (1, 1), 3, 3, 3, (1, 1), (0, 0, 0, 0), (1, 1), false}; kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/conv.jl:91
  [9] conv(x::Array{Float64, 4}, w::Array{Float32, 4}, cdims::DenseConvDims{2, (1, 1), 3, 3, 3, (1, 1), (0, 0, 0, 0), (1, 1), false})
    @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/conv.jl:89
 [10] (::Conv{2, 4, typeof(identity), Array{Float32, 4}, Vector{Float32}})(x::Array{Float64, 4})
    @ Flux ~/.julia/packages/Flux/ZnXxS/src/layers/conv.jl:163
 [11] top-level scope
    @ REPL[48]:1
` errors: 

```

Stacktrace with 1 channel (this should not work and does not work):

```julia
julia> T(rand(10,10,1,1))
ERROR: DimensionMismatch("Input channels must match! (1 vs. 1)")
Stacktrace:
 [1] DenseConvDims(x_size::NTuple{4, Int64}, w_size::NTuple{4, Int64}; stride::Tuple{Int64, Int64}, padding::NTuple{4, Int64}, dilation::Tuple{Int64, Int64}, flipkernel::Bool, groups::Int64)
   @ NNlib ~/.julia/packages/NNlib/P9BhZ/src/dim_helpers/DenseConvDims.jl:30
 [2] #DenseConvDims#7
   @ ~/.julia/packages/NNlib/P9BhZ/src/dim_helpers/DenseConvDims.jl:60 [inlined]
 [3] (::Conv{2, 4, typeof(identity), Array{Float32, 4}, Vector{Float32}})(x::Array{Float64, 4})
   @ Flux ~/.julia/packages/Flux/ZnXxS/src/layers/conv.jl:162
 [4] top-level scope
   @ REPL[49]:1

```

---

<div class="post-metadata">

**Author:** ![ToucheSir](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/touchesir/32/14411_2.png) [@ToucheSir](https://discourse.julialang.org/u/ToucheSir)\
**Post date:** [December 26, 2021, 10:40pm UTC](https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641/2 "2021-12-26T22:40:44Z")

</div>

It seems like you may have missed the preceding warning:

```julia-repl
julia> c(rand(10, 10, 3, 1))
┌ Warning: Slow fallback implementation invoked for conv! You probably don't want this; check your datatypes.
│ yT = Float64
│ T1 = Float64
│ T2 = Float32
└ @ NNlib ~/.julia/packages/NNlib/tvMmZ/src/conv.jl:291

```

Using Float32, everything works as expected:

```julia-repl
julia> c(rand(Float32, 10, 10, 3, 1)) |> size
(10, 10, 3, 1)

```

That said, the assert being triggered from the fallback definition does look fishy to me. Will report an issue on NNlib for this.

---

<div class="post-metadata">

**Author:** ![sim](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/sim/32/22207_2.png) [@sim](https://discourse.julialang.org/u/sim)\
**Post date:** [December 26, 2021, 11:48pm UTC](https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641/3 "2021-12-26T23:48:31Z")

</div>

Hi, you are actually right : I didn’t pay attention to the `Slow fallback` warning, and things work as intented with Float32. But the

> ERROR: AssertionError: DimensionMismatch(“Data input channel count (3 vs. 3)”)

and the stacktrace are super misleading !

Thank you.

---

<div class="post-metadata">

**Author:** ![ToucheSir](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/touchesir/32/14411_2.png) [@ToucheSir](https://discourse.julialang.org/u/ToucheSir)\
**Post date:** [December 27, 2021, 3:56am UTC](https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641/4 "2021-12-27T03:56:34Z")

</div>

The stacktrace is super misleading because it shouldn’t be there 😛. The fallback should still work without erroring, so this is a bug. If you wouldn’t mind filing an issue on NNlib, we can track a fix for this.

---

<div class="post-metadata">

**Author:** ![sim](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/sim/32/22207_2.png) [@sim](https://discourse.julialang.org/u/sim)\
**Post date:** [December 27, 2021, 8:50am UTC](https://discourse.julialang.org/t/groups-argument-in-flux-conv-a-misunderstanding/73641/5 "2021-12-27T08:50:23Z")

</div>

I opened the issue [here](https://github.com/FluxML/NNlib.jl/issues/369).  
Thank you for your help !