Merge pull request #125 from theabhirath/resmlp

Implementation of ResMLP and gMLP (with improvements to MLPMixer and `PatchEmbedding`)

Author: darsnack

src/Metalhead.jl

@@ -43,13 +43,13 @@ export  AlexNet,
         DenseNet, DenseNet121, DenseNet161, DenseNet169, DenseNet201,
         ResNeXt,
         MobileNetv2, MobileNetv3,
-        MLPMixer,
+        MLPMixer, ResMLP, gMLP,
         ViT,
         ConvNeXt
 
 # use Flux._big_show to pretty print large models
 for T in (:AlexNet, :VGG, :ResNet, :GoogLeNet, :Inception3, :SqueezeNet, :DenseNet, :ResNeXt, 
-          :MobileNetv2, :MobileNetv3, :MLPMixer, :ViT, :ConvNeXt)
+          :MobileNetv2, :MobileNetv3, :MLPMixer, :ResMLP, :gMLP, :ViT, :ConvNeXt)
   @eval Base.show(io::IO, ::MIME"text/plain", model::$T) = _maybe_big_show(io, model)
 end
 

src/convnets/convnext.jl

@@ -1,25 +1,22 @@
 """
-    convnextblock(planes, drop_path = 0., λ = 1f-6)
+    convnextblock(planes, drop_path_rate = 0., λ = 1f-6)
 
 Creates a single block of ConvNeXt.
 ([reference](https://arxiv.org/abs/2201.03545))
 
 # Arguments:
 - `planes`: number of input channels.
 - `drop_path_rate`: Stochastic depth rate.
-- `λ`: Init value for [LayerScale](https://arxiv.org/abs/2103.17239)
+- `λ`: Init value for LayerScale
 """
 function convnextblock(planes, drop_path_rate = 0., λ = 1f-6)
-  γ = Flux.ones32(planes) * λ
-  LayerScale(x) = x .* γ
-  scale = λ > 0 ? identity : LayerScale
   layers = SkipConnection(Chain(DepthwiseConv((7, 7), planes => planes; pad = 3), 
                                 x -> permutedims(x, (3, 1, 2, 4)),
                                 LayerNorm(planes; ϵ = 1f-6),
                                 mlp_block(planes, 4 * planes),
-                                scale, # LayerScale
+                                LayerScale(planes, λ),
                                 x -> permutedims(x, (2, 3, 1, 4)),
-                                Dropout(drop_path_rate, dims = 4)), +)
+                                DropPath(drop_path_rate)), +)
   return layers
 end
 
@@ -89,9 +86,12 @@ Creates a ConvNeXt model.
 - `drop_path_rate`: Stochastic depth rate.
 - `λ`: Init value for [LayerScale](https://arxiv.org/abs/2103.17239)
 - `nclasses`: number of output classes
+
+See also [`Metalhead.convnext`](#).
 """
 function ConvNeXt(mode::Symbol = :base; inchannels = 3, drop_path_rate = 0., λ = 1f-6, 
                   nclasses = 1000)
+  @assert mode in keys(convnext_configs) "`size` must be one of $(collect(keys(convnext_configs)))"
   depths = convnext_configs[mode][:depths]
   planes = convnext_configs[mode][:planes]
   layers = convnext(depths, planes; inchannels, drop_path_rate, λ, nclasses)

src/convnets/densenet.jl

@@ -80,7 +80,7 @@ function densenet(inplanes, growth_rates; reduction = 0.5, nclasses = 1000)
 end
 
 """
-    densenet(nblocks; growth_rate = 32, reduction = 0.5, num_classes = 1000)
+    densenet(nblocks; growth_rate = 32, reduction = 0.5, nclasses = 1000)
 
 Create a DenseNet model
 ([reference](https://arxiv.org/abs/1608.06993)).

src/convnets/mobilenet.jl

@@ -131,11 +131,9 @@ function mobilenetv3(width_mult, configs; max_width = 1024, nclasses = 1000)
   # building last several layers
   output_channel = max_width
   output_channel = width_mult > 1.0 ? _round_channels(output_channel * width_mult, 8) : output_channel
-  classifier = (
-    Dense(explanes, output_channel, hardswish),
-    Dropout(0.2),
-    Dense(output_channel, nclasses),
-  )
+  classifier = (Dense(explanes, output_channel, hardswish),
+                Dropout(0.2),
+                Dense(output_channel, nclasses))
 
   return Chain(Chain(layers...,
                      conv_bn((1, 1), inplanes, explanes, hardswish, bias = false)...),

src/layers/Layers.jl

@@ -13,10 +13,12 @@ include("embeddings.jl")
 include("mlp.jl")
 include("normalise.jl")
 include("conv.jl")
+include("others.jl")
 
 export Attention, MHAttention,
        PatchEmbedding, ViPosEmbedding, ClassTokens,
-       mlp_block,
+       mlp_block, gated_mlp_block,
+       LayerScale, DropPath,
        ChannelLayerNorm, prenorm,
        skip_identity, skip_projection,
        conv_bn,

src/layers/embeddings.jl

@@ -1,27 +1,34 @@
 """
-    PatchEmbedding(patch_size)
-    PatchEmbedding(patch_height, patch_width)
+    PatchEmbedding(imsize::NTuple{2} = (224, 224); inchannels = 3, patch_size = (16, 16), 
+                   embedplanes = 768, norm_layer = planes -> identity, flatten = true)
 
-Patch embedding layer used by many vision transformer-like models to split the input image into patches.
-"""
-struct PatchEmbedding
-  patch_height::Int
-  patch_width::Int
-end
+Patch embedding layer used by many vision transformer-like models to split the input image into 
+patches.
 
-PatchEmbedding(patch_size) = PatchEmbedding(patch_size, patch_size)
+# Arguments:
+- `imsize`: the size of the input image
+- `inchannels`: the number of channels in the input image
+- `patch_size`: the size of the patches
+- `embedplanes`: the number of channels in the embedding
+- `norm_layer`: the normalization layer - by default the identity function but otherwise takes a 
+                single argument constructor for a normalization layer like LayerNorm or BatchNorm
+- `flatten`: set true to flatten the input spatial dimensions after the embedding
+"""
+function PatchEmbedding(imsize::NTuple{2} = (224, 224); inchannels = 3, patch_size = (16, 16), 
+                        embedplanes = 768, norm_layer = planes -> identity, flatten = true)
 
-function (p::PatchEmbedding)(x)
-  h, w, c, n = size(x)
-  hp, wp = h ÷ p.patch_height, w ÷ p.patch_width
-  xpatch = reshape(x, hp, p.patch_height, wp, p.patch_width, c, n)
+  im_height, im_width = imsize
+  patch_height, patch_width = patch_size
 
-  return reshape(permutedims(xpatch, (1, 3, 5, 2, 4, 6)), p.patch_height * p.patch_width * c, 
-                 hp * wp, n)
+  @assert (im_height % patch_height == 0) && (im_width % patch_width == 0)
+  "Image dimensions must be divisible by the patch size."
+  
+  return Chain(Conv(patch_size, inchannels => embedplanes; stride = patch_size),
+               flatten ? x -> permutedims(reshape(x, (:, size(x, 3), size(x, 4))), (2, 1, 3)) 
+                       : identity,
+               norm_layer(embedplanes))
 end
 
-@functor PatchEmbedding
-
 """
     ViPosEmbedding(embedsize, npatches; init = (dims) -> rand(Float32, dims))
 

src/layers/mlp.jl

@@ -1,16 +1,44 @@
 """
-    mlp_block(planes, hidden_planes; dropout = 0., dense = Dense, activation = gelu)
+		mlp_block(inplanes::Integer, hidden_planes::Integer, outplanes::Integer = inplanes; 
+							dropout = 0., activation = gelu)
 
-Feedforward block used in many vision transformer-like models.
+Feedforward block used in many MLPMixer-like and vision-transformer models.
 
 # Arguments
-- `planes`: Number of dimensions in the input and output.
+- `inplanes`: Number of dimensions in the input.
 - `hidden_planes`: Number of dimensions in the intermediate layer.
+- `outplanes`: Number of dimensions in the output - by default it is the same as `inplanes`.
 - `dropout`: Dropout rate.
-- `dense`: Type of dense layer to use in the feedforward block.
 - `activation`: Activation function to use.
 """
-function mlp_block(planes, hidden_planes; dropout = 0., dense = Dense, activation = gelu)
-  Chain(dense(planes, hidden_planes, activation), Dropout(dropout),
-        dense(hidden_planes, planes), Dropout(dropout))
+function mlp_block(inplanes::Integer, hidden_planes::Integer, outplanes::Integer = inplanes; 
+									 dropout = 0., activation = gelu)
+  Chain(Dense(inplanes, hidden_planes, activation), Dropout(dropout),
+        Dense(hidden_planes, outplanes), Dropout(dropout))
 end
+
+"""
+    gated_mlp(gate_layer, inplanes::Integer, hidden_planes::Integer, 
+							outplanes::Integer = inplanes; dropout = 0., activation = gelu)
+
+Feedforward block based on the implementation in the paper "Pay Attention to MLPs".
+([reference](https://arxiv.org/abs/2105.08050))
+
+# Arguments
+- `gate_layer`: Layer to use for the gating.
+- `inplanes`: Number of dimensions in the input.
+- `hidden_planes`: Number of dimensions in the intermediate layer.
+- `outplanes`: Number of dimensions in the output - by default it is the same as `inplanes`.
+- `dropout`: Dropout rate.
+- `activation`: Activation function to use.
+"""
+function gated_mlp_block(gate_layer, inplanes::Integer, hidden_planes::Integer,
+												 outplanes::Integer = inplanes; dropout = 0., activation = gelu)
+	@assert hidden_planes % 2 == 0 "`hidden_planes` must be even for gated MLP"
+	return Chain(Dense(inplanes, hidden_planes, activation),
+							 Dropout(dropout),
+							 gate_layer(hidden_planes),
+							 Dense(hidden_planes ÷ 2, outplanes),
+							 Dropout(dropout))
+end
+gated_mlp_block(::typeof(identity), args...; kwargs...) = mlp_block(args...; kwargs...)

src/layers/others.jl

@@ -0,0 +1,38 @@
+"""
+    LayerScale(scale) 
+
+Implements LayerScale.
+([reference](https://arxiv.org/abs/2103.17239))
+
+# Arguments
+- `scale`: Scaling factor, a learnable diagonal matrix which is multiplied to the input.
+"""
+struct LayerScale{T<:AbstractVector{<:Real}}
+    scale::T
+end
+
+"""
+    LayerScale(λ, planes::Int)
+
+Implements LayerScale.
+([reference](https://arxiv.org/abs/2103.17239))
+
+# Arguments
+- `planes`: Size of channel dimension in the input.
+- `λ`: initialisation value for the learnable diagonal matrix.
+"""
+LayerScale(planes::Int, λ) = λ > 0 ? LayerScale(fill(Float32(λ), planes)) : identity
+
+@functor LayerScale
+(m::LayerScale)(x::AbstractArray) = m.scale .* x
+
+"""
+    DropPath(p)
+
+Implements Stochastic Depth - equivalent to `Dropout(p; dims = 4)` when `p` ≥ 0.
+([reference](https://arxiv.org/abs/1603.09382))
+
+# Arguments
+- `p`: rate of Stochastic Depth.
+"""
+DropPath(p) = p ≥ 0 ? Dropout(p; dims = 4) : identity