Fix ConvTranspose symmetric non-constant padding (#2463)

The ConvTranspose was not able to handle symmetric non-constant
padding (ie, `pad=(1, 0)` for 2D ConvTranspose). Constant padding (ie
`pad=1` for 2D ConvTranspose) and assymetric non-constant padding (ie,
`pad=(1, 0, 2, 3)`) worked correctly. This commit fixes symmetric
non-constant padding and adds unit tests to ensure it produces the same
output size as an equivalent fully expanded padding.

Fixes #2424

ext/FluxAMDGPUExt/conv.jl

@@ -5,8 +5,16 @@ function Flux.conv_dims(c::Conv, x::T) where T <: ROCArray
 end
 
 function Flux.conv_transpose_dims(c::ConvTranspose, x::T) where T <: ROCArray
+    # Calculate combined pad in each dimension
+    nd = ndims(x) - 2
+    if length(c.pad) == nd
+        # Handle symmetric non-constant padding
+        combined_pad = ntuple(i -> 2 * c.pad[i], nd)
+    else
+        combined_pad = ntuple(i -> c.pad[2i-1] + c.pad[2i], nd)
+    end
+
     # Calculate size of "input", from ∇conv_data()'s perspective...
-    combined_pad = (c.pad[1:2:end] .+ c.pad[2:2:end])
     I = (size(x)[1:end - 2] .- 1) .* c.stride .+ 1 .+
         (size(c.weight)[1:end - 2] .- 1) .* c.dilation .- combined_pad .+ c.outpad
     C_in = size(c.weight)[end - 1] * c.groups

src/layers/conv.jl

@@ -321,12 +321,21 @@ end
 @layer ConvTranspose
 
 function conv_transpose_dims(c::ConvTranspose, x::AbstractArray)
+  # Calculate combined pad in each dimension
+  nd = ndims(x) - 2
+  if length(c.pad) == nd
+    # Handle symmetric non-constant padding
+    combined_pad = ntuple(i -> 2 * c.pad[i], nd)
+  else
+    combined_pad = ntuple(i -> c.pad[2i-1] + c.pad[2i], nd)
+  end
+
   # Calculate size of "input", from ∇conv_data()'s perspective...
   calc_dim(xsz, wsz, stride, dilation, pad, outpad) = (xsz - 1) * stride + 1 + (wsz - 1) * dilation - pad + outpad
-  combined_pad = ntuple(i -> c.pad[2i-1] + c.pad[2i], length(c.pad) ÷ 2)
   I = map(calc_dim, size(x)[1:end-2], size(c.weight)[1:end-2], c.stride, c.dilation, combined_pad, c.outpad)
   C_in = size(c.weight)[end-1] * c.groups
   batch_size = size(x)[end]
+
   # Create DenseConvDims() that looks like the corresponding conv()
   w_size = size(c.weight)
   return DenseConvDims((I..., C_in, batch_size), w_size;

test/ext_amdgpu/basic.jl

@@ -46,6 +46,21 @@ end
     end
 end
 
+@testset "Convolution with symmetric non-constant padding" begin
+    for conv_type in (Conv, ConvTranspose), nd in 1:3
+        kernel = tuple(fill(2, nd)...)
+        x = rand(Float32, fill(10, nd)..., 3, 5) |> gpu
+
+        pad = ntuple(i -> i, nd)
+        m = conv_type(kernel, 3 => 4, pad=pad) |> f32 |> gpu
+
+        expanded_pad = ntuple(i -> pad[(i - 1) ÷ 2 + 1], 2 * nd)
+        m_expanded = conv_type(kernel, 3 => 4, pad=expanded_pad) |> f32 |> gpu
+
+        @test size(m(x)) == size(m_expanded(x))
+    end
+end
+
 @testset "ConvTranspose output padding" begin
     x = randn(Float32, 10, 11, 3, 2)
     m = ConvTranspose((3, 5), 3=>6, stride=3, outpad=(1, 0))

test/layers/conv.jl

@@ -257,6 +257,19 @@ end
   end
 end
 
+@testset "$ltype $(nd)D symmetric non-constant padding" for ltype in (Conv, ConvTranspose, DepthwiseConv, CrossCor), nd in (1, 2, 3)
+  kernel = ntuple(Returns(3), nd)
+  data = ones(Float32, (kernel .+ 5)..., 1,1)
+
+  pad = ntuple(i -> i, nd)
+  l = ltype(kernel, 1=>1, pad=pad)
+
+  expanded_pad = ntuple(i -> pad[(i - 1) ÷ 2 + 1], 2 * nd)
+  l_expanded = ltype(kernel, 1=>1, pad=expanded_pad)
+
+  @test size(l(data)) == size(l_expanded(data))
+end
+
 @testset "$ltype SamePad kernelsize $k" for ltype in (Conv, ConvTranspose, DepthwiseConv, CrossCor), k in ( (1,), (2,), (3,), (4,5), (6,7,8))
   data = ones(Float32, (k .+ 3)..., 1,1)
   l = ltype(k, 1=>1, pad=SamePad())