The one true function

1. Unify MobileNet and EfficientNet APIs into a single lower level builder function
2. Further unification and consolidation of the mid level API
3. Some cleanup

Author: theabhirath

src/Metalhead.jl

@@ -20,26 +20,27 @@ using .Layers
 
 # CNN models
 ## Builders
-include("convnets/builders/core.jl")
+include("convnets/builders/irmodel.jl")
 include("convnets/builders/mbconv.jl")
 include("convnets/builders/resblocks.jl")
+include("convnets/builders/resnet.jl")
+include("convnets/builders/stages.jl")
 ## AlexNet and VGG
 include("convnets/alexnet.jl")
 include("convnets/vgg.jl")
 ## ResNets
 include("convnets/resnets/core.jl")
+include("convnets/resnets/res2net.jl")
 include("convnets/resnets/resnet.jl")
 include("convnets/resnets/resnext.jl")
 include("convnets/resnets/seresnet.jl")
-include("convnets/resnets/res2net.jl")
 ## Inceptions
 include("convnets/inceptions/googlenet.jl")
 include("convnets/inceptions/inceptionv3.jl")
 include("convnets/inceptions/inceptionv4.jl")
 include("convnets/inceptions/inceptionresnetv2.jl")
 include("convnets/inceptions/xception.jl")
 ## EfficientNets
-include("convnets/efficientnets/core.jl")
 include("convnets/efficientnets/efficientnet.jl")
 include("convnets/efficientnets/efficientnetv2.jl")
 ## MobileNets

src/convnets/builders/irmodel.jl

@@ -0,0 +1,41 @@
+function irmodelbuilder(scalings::NTuple{2, Real}, block_configs::AbstractVector{<:Tuple};
+                        inplanes::Integer = 32, connection = +, activation = relu,
+                        norm_layer = BatchNorm, divisor::Integer = 8,
+                        tail_conv::Bool = true, expanded_classifier::Bool = false,
+                        headplanes::Integer, dropout_rate = nothing,
+                        inchannels::Integer = 3, nclasses::Integer = 1000, kwargs...)
+    width_mult, _ = scalings
+    # building first layer
+    inplanes = _round_channels(inplanes * width_mult, divisor)
+    layers = []
+    append!(layers,
+            conv_norm((3, 3), inchannels, inplanes, activation; stride = 2, pad = 1,
+                      norm_layer))
+    # building inverted residual blocks
+    get_layers, block_repeats = mbconv_stage_builder(block_configs, inplanes, scalings;
+                                                     norm_layer, divisor, kwargs...)
+    append!(layers, cnn_stages(get_layers, block_repeats, connection))
+    # building last layers
+    outplanes = _round_channels(block_configs[end][3] * width_mult, divisor)
+    if tail_conv
+        # special case, supported fully only for MobileNetv3
+        if expanded_classifier
+            midplanes = _round_channels(outplanes * block_configs[end][4], divisor)
+            append!(layers,
+                    conv_norm((1, 1), outplanes, midplanes, activation; norm_layer))
+            classifier = create_classifier(midplanes, headplanes, nclasses,
+                                           (hardswish, identity); dropout_rate)
+        else
+            append!(layers,
+                    conv_norm((1, 1), outplanes, headplanes, activation; norm_layer))
+            classifier = create_classifier(headplanes, nclasses; dropout_rate)
+        end
+    else
+        classifier = create_classifier(outplanes, nclasses; dropout_rate)
+    end
+    return Chain(Chain(layers...), classifier)
+end
+
+function irmodelbuilder(width_mult::Real, block_configs::AbstractVector{<:Tuple}; kwargs...)
+    return irmodelbuilder((width_mult, 1), block_configs; kwargs...)
+end

src/convnets/builders/mbconv.jl

@@ -1,9 +1,14 @@
-function dwsepconv_builder(block_configs, inplanes::Integer, stage_idx::Integer,
-                           width_mult::Real; norm_layer = BatchNorm, kwargs...)
+# TODO - potentially make these builders more flexible to specify stuff like
+# activation functions and reductions that don't change over the stages
+
+function dwsepconv_builder(block_configs::AbstractVector{<:Tuple}, inplanes::Integer,
+                           stage_idx::Integer, scalings::NTuple{2, Real};
+                           norm_layer = BatchNorm, divisor::Integer = 8, kwargs...)
+    width_mult, depth_mult = scalings
     block_fn, k, outplanes, stride, nrepeats, activation = block_configs[stage_idx]
-    outplanes = _round_channels(outplanes * width_mult)
+    outplanes = _round_channels(outplanes * width_mult, divisor)
     if stage_idx != 1
-        inplanes = _round_channels(block_configs[stage_idx - 1][3] * width_mult)
+        inplanes = _round_channels(block_configs[stage_idx - 1][3] * width_mult, divisor)
     end
     function get_layers(block_idx::Integer)
         inplanes = block_idx == 1 ? inplanes : outplanes
@@ -12,13 +17,14 @@ function dwsepconv_builder(block_configs, inplanes::Integer, stage_idx::Integer,
                                stride, pad = SamePad(), norm_layer, kwargs...)...)
         return (block,)
     end
-    return get_layers, nrepeats
+    return get_layers, ceil(Int, nrepeats * depth_mult)
 end
+_get_builder(::typeof(dwsep_conv_norm)) = dwsepconv_builder
 
-function mbconv_builder(block_configs, inplanes::Integer, stage_idx::Integer,
-                        scalings::NTuple{2, Real}; norm_layer = BatchNorm,
-                        divisor::Integer = 8, se_from_explanes::Bool = false,
-                        kwargs...)
+function mbconv_builder(block_configs::AbstractVector{<:Tuple}, inplanes::Integer,
+                        stage_idx::Integer, scalings::NTuple{2, Real};
+                        norm_layer = BatchNorm, divisor::Integer = 8,
+                        se_from_explanes::Bool = false, kwargs...)
     width_mult, depth_mult = scalings
     block_fn, k, outplanes, expansion, stride, nrepeats, reduction, activation = block_configs[stage_idx]
     # calculate number of reduced channels for squeeze-excite layer from explanes instead of inplanes
@@ -39,69 +45,31 @@ function mbconv_builder(block_configs, inplanes::Integer, stage_idx::Integer,
     end
     return get_layers, ceil(Int, nrepeats * depth_mult)
 end
+_get_builder(::typeof(mbconv)) = mbconv_builder
 
-function mbconv_builder(block_configs, inplanes::Integer, stage_idx::Integer,
-                        width_mult::Real; norm_layer = BatchNorm, kwargs...)
-    return mbconv_builder(block_configs, inplanes, stage_idx, (width_mult, 1);
-                          norm_layer, kwargs...)
-end
-
-function fused_mbconv_builder(block_configs, inplanes::Integer, stage_idx::Integer;
-                              norm_layer = BatchNorm, kwargs...)
+function fused_mbconv_builder(block_configs::AbstractVector{<:Tuple}, inplanes::Integer,
+                              stage_idx::Integer, scalings::NTuple{2, Real};
+                              norm_layer = BatchNorm, divisor::Integer = 8, kwargs...)
+    width_mult, depth_mult = scalings
     block_fn, k, outplanes, expansion, stride, nrepeats, activation = block_configs[stage_idx]
     inplanes = stage_idx == 1 ? inplanes : block_configs[stage_idx - 1][3]
+    outplanes = _round_channels(outplanes * width_mult, divisor)
     function get_layers(block_idx::Integer)
         inplanes = block_idx == 1 ? inplanes : outplanes
-        explanes = _round_channels(inplanes * expansion, 8)
+        explanes = _round_channels(inplanes * expansion, divisor)
         stride = block_idx == 1 ? stride : 1
         block = block_fn((k, k), inplanes, explanes, outplanes, activation;
                          norm_layer, stride, kwargs...)
         return stride == 1 && inplanes == outplanes ? (identity, block) : (block,)
     end
-    return get_layers, nrepeats
-end
-
-# TODO - these builders need to be more flexible to potentially specify stuff like
-# activation functions and reductions that don't change
-function _get_builder(::typeof(dwsep_conv_bn), block_configs::AbstractVector{<:Tuple},
-                      inplanes::Integer, stage_idx::Integer;
-                      scalings::Union{Nothing, NTuple{2, Real}} = nothing,
-                      width_mult::Union{Nothing, Number} = nothing, norm_layer, kwargs...)
-    @assert isnothing(scalings) "dwsep_conv_bn does not support the `scalings` argument"
-    return dwsepconv_builder(block_configs, inplanes, stage_idx, width_mult; norm_layer,
-                             kwargs...)
-end
-
-function _get_builder(::typeof(mbconv), block_configs::AbstractVector{<:Tuple},
-                      inplanes::Integer, stage_idx::Integer;
-                      scalings::Union{Nothing, NTuple{2, Real}} = nothing,
-                      width_mult::Union{Nothing, Number} = nothing, norm_layer, kwargs...)
-    if isnothing(scalings)
-        return mbconv_builder(block_configs, inplanes, stage_idx, width_mult; norm_layer,
-                              kwargs...)
-    elseif isnothing(width_mult)
-        return mbconv_builder(block_configs, inplanes, stage_idx, scalings; norm_layer,
-                              kwargs...)
-    else
-        throw(ArgumentError("Only one of `scalings` and `width_mult` can be specified"))
-    end
-end
-
-function _get_builder(::typeof(fused_mbconv), block_configs::AbstractVector{<:Tuple},
-                      inplanes::Integer, stage_idx::Integer;
-                      scalings::Union{Nothing, NTuple{2, Real}} = nothing,
-                      width_mult::Union{Nothing, Number} = nothing, norm_layer)
-    @assert isnothing(width_mult) "fused_mbconv does not support the `width_mult` argument."
-    @assert isnothing(scalings)||scalings == (1, 1) "fused_mbconv does not support the `scalings` argument"
-    return fused_mbconv_builder(block_configs, inplanes, stage_idx; norm_layer)
+    return get_layers, ceil(Int, nrepeats * depth_mult)
 end
+_get_builder(::typeof(fused_mbconv)) = fused_mbconv_builder
 
-function mbconv_stack_builder(block_configs::AbstractVector{<:Tuple}, inplanes::Integer;
-                              scalings::Union{Nothing, NTuple{2, Real}} = nothing,
-                              width_mult::Union{Nothing, Number} = nothing,
-                              norm_layer = BatchNorm, kwargs...)
-    bxs = [_get_builder(block_configs[idx][1], block_configs, inplanes, idx; scalings,
-                        width_mult, norm_layer, kwargs...)
+function mbconv_stage_builder(block_configs::AbstractVector{<:Tuple}, inplanes::Integer,
+                              scalings::NTuple{2, Real}; kwargs...)
+    builders = _get_builder.(first.(block_configs))
+    bxs = [builders[idx](block_configs, inplanes, idx, scalings; kwargs...)
            for idx in eachindex(block_configs)]
     return (stage_idx, block_idx) -> first.(bxs)[stage_idx](block_idx), last.(bxs)
 end

src/convnets/builders/resnet.jl

@@ -0,0 +1,9 @@
+function resnetbuilder(img_dims, stem, get_layers, block_repeats::AbstractVector{<:Integer},
+                       connection, classifier_fn)
+    # Build stages of the ResNet
+    stage_blocks = cnn_stages(get_layers, block_repeats, connection)
+    backbone = Chain(stem, stage_blocks...)
+    # Add classifier to the backbone
+    nfeaturemaps = Flux.outputsize(backbone, img_dims; padbatch = true)[3]
+    return Chain(backbone, classifier_fn(nfeaturemaps))
+end

src/convnets/builders/core.jl renamed to src/convnets/builders/stages.jl

@@ -46,11 +46,11 @@ function convnext(depths::AbstractVector{<:Integer}, planes::AbstractVector{<:In
     "`planes` should have exactly one value for each block"
     downsample_layers = []
     push!(downsample_layers,
-          Chain(conv_norm((4, 4), inchannels => planes[1]; stride = 4,
+          Chain(conv_norm((4, 4), inchannels, planes[1]; stride = 4,
                           norm_layer = ChannelLayerNorm)...))
     for m in 1:(length(depths) - 1)
         push!(downsample_layers,
-              Chain(conv_norm((2, 2), planes[m] => planes[m + 1]; stride = 2,
+              Chain(conv_norm((2, 2), planes[m], planes[m + 1]; stride = 2,
                               norm_layer = ChannelLayerNorm, revnorm = true)...))
     end
     stages = []
@@ -111,7 +111,7 @@ function ConvNeXt(config::Symbol; pretrain::Bool = true, inchannels::Integer = 3
     _checkconfig(config, keys(CONVNEXT_CONFIGS))
     layers = convnext(config; inchannels, nclasses)
     if pretrain
-        layers = load_pretrained(layers, "convnext_$config")
+        layers = loadpretrain!(layers, "convnext_$config")
     end
     return ConvNeXt(layers)
 end

src/convnets/convnext.jl

@@ -31,6 +31,17 @@ const EFFICIENTNET_GLOBAL_CONFIGS = Dict(:b0 => (224, (1.0, 1.0)),
                                          :b7 => (600, (2.0, 3.1)),
                                          :b8 => (672, (2.2, 3.6)))
 
+function efficientnet(config::Symbol; norm_layer = BatchNorm,
+                      dropout_rate = nothing, inchannels::Integer = 3,
+                      nclasses::Integer = 1000)
+    _checkconfig(config, keys(EFFICIENTNET_GLOBAL_CONFIGS))
+    scalings = EFFICIENTNET_GLOBAL_CONFIGS[config][2]
+    return irmodelbuilder(scalings, EFFICIENTNET_BLOCK_CONFIGS; inplanes = 32,
+                          norm_layer, activation = swish,
+                          headplanes = EFFICIENTNET_BLOCK_CONFIGS[end][3] * 4,
+                          dropout_rate, inchannels, nclasses)
+end
+
 """
     EfficientNet(config::Symbol; pretrain::Bool = false)
 
@@ -50,10 +61,7 @@ end
 
 function EfficientNet(config::Symbol; pretrain::Bool = false, inchannels::Integer = 3,
                       nclasses::Integer = 1000)
-    _checkconfig(config, keys(EFFICIENTNET_GLOBAL_CONFIGS))
-    scalings = EFFICIENTNET_GLOBAL_CONFIGS[config][2]
-    layers = efficientnet_core(EFFICIENTNET_BLOCK_CONFIGS; inplanes = 32, scalings,
-                               inchannels, nclasses)
+    layers = efficientnet(config; inchannels, nclasses)
     if pretrain
         loadpretrain!(layers, string("efficientnet-", config))
     end

src/convnets/efficientnets/core.jl

@@ -35,6 +35,15 @@ const EFFNETV2_CONFIGS = Dict(:small => [(fused_mbconv, 3, 24, 1, 1, 2, swish),
                                   (mbconv, 3, 512, 6, 2, 32, 4, swish),
                                   (mbconv, 3, 768, 6, 1, 8, 4, swish)])
 
+function efficientnetv2(config::Symbol; norm_layer = BatchNorm, dropout_rate = nothing,
+                        inchannels::Integer = 3, nclasses::Integer = 1000)
+    _checkconfig(config, keys(EFFNETV2_CONFIGS))
+    block_configs = EFFNETV2_CONFIGS[config]
+    return irmodelbuilder((1, 1), block_configs; activation = swish, norm_layer,
+                          inplanes = block_configs[1][3], headplanes = 1280,
+                          dropout_rate, inchannels, nclasses)
+end
+
 """
     EfficientNetv2(config::Symbol; pretrain::Bool = false, width_mult::Real = 1,
                    inchannels::Integer = 3, nclasses::Integer = 1000)
@@ -57,10 +66,7 @@ end
 
 function EfficientNetv2(config::Symbol; pretrain::Bool = false,
                         inchannels::Integer = 3, nclasses::Integer = 1000)
-    _checkconfig(config, sort(collect(keys(EFFNETV2_CONFIGS))))
-    block_configs = EFFNETV2_CONFIGS[config]
-    layers = efficientnet_core(block_configs; inplanes = block_configs[1][3],
-                               headplanes = 1280, inchannels, nclasses)
+    layers = efficientnetv2(config; inchannels, nclasses)
     if pretrain
         loadpretrain!(layers, string("efficientnetv2-", config))
     end

src/convnets/efficientnets/efficientnet.jl

@@ -34,7 +34,7 @@ function xception_block(inchannels::Integer, outchannels::Integer, nrepeats::Int
         end
         push!(layers, relu)
         append!(layers,
-                dwsep_conv_bn((3, 3), inc, outc; pad = 1, use_norm = (false, false)))
+                dwsep_conv_norm((3, 3), inc, outc; pad = 1, use_norm = (false, false)))
         push!(layers, BatchNorm(outc))
     end
     layers = start_with_relu ? layers : layers[2:end]
@@ -62,8 +62,8 @@ function xception(; dropout_rate = 0.0, inchannels::Integer = 3, nclasses::Integ
                      xception_block(256, 728, 2; stride = 2),
                      [xception_block(728, 728, 3) for _ in 1:8]...,
                      xception_block(728, 1024, 2; stride = 2, grow_at_start = false),
-                     dwsep_conv_bn((3, 3), 1024, 1536; pad = 1)...,
-                     dwsep_conv_bn((3, 3), 1536, 2048; pad = 1)...)
+                     dwsep_conv_norm((3, 3), 1024, 1536; pad = 1)...,
+                     dwsep_conv_norm((3, 3), 1536, 2048; pad = 1)...)
     return Chain(backbone, create_classifier(2048, nclasses; dropout_rate))
 end