Add back ResNet family and SqueezeNet

Cleanup, but also DenseNets don't work just yet

Author: theabhirath

Artifacts.toml

@@ -1,3 +1,66 @@
+[resnet101]
+git-tree-sha1 = "68d563526ab34d3e5aa66b7d96278d2acde212f9"
+
+    [[resnet101.download]]
+    sha256 = "0725f05db5772cfab1024b8d0d6c85ac1fc5a83eb6f0fe02b67b1e689d5a28db"
+    url = "https://huggingface.co/FluxML/resnet101/resolve/980158099e6917b74ade2b0a9599359f06057d21/resnet101.tar.gz"
+
+[resnet152]
+git-tree-sha1 = "85a97464b6cef66e1217ae829d3620651cffab47"
+
+    [[resnet152.download]]
+    sha256 = "a8d30a735ef5649ec40a74a0515ee3d6774499267be06f5f2b372259c5ced8d6"
+    url = "https://huggingface.co/FluxML/resnet152/resolve/a66a3e1f5056179d167cb2165401950e3890b34d/resnet152.tar.gz"
+
+[resnet18]
+git-tree-sha1 = "4ced5a0338c0f0293940f1deb63e1c463125a6ff"
+
+    [[resnet18.download]]
+    sha256 = "9444ef2285f507bd890d2ca852d663749f079110ed19b544e8d91f67f3cc6b83"
+    url = "https://huggingface.co/FluxML/resnet18/resolve/9b1c6c4f7c5dbe734d80d7d4b5f132ef58bf2467/resnet18.tar.gz"
+
+[resnet34]
+git-tree-sha1 = "485519977f375ca1770b3ff3971f61e438823f5a"
+
+    [[resnet34.download]]
+    sha256 = "71ed75be6db0160af7f30be33e2f4a44836310949d1374267510e5803b1fb313"
+    url = "https://huggingface.co/FluxML/resnet34/resolve/0988ae2d4a86da06eefa6b61edf3e728861e286c/resnet34.tar.gz"
+
+[resnet50]
+git-tree-sha1 = "2973be0da60544080105756ecb3951cca2e007da"
+
+    [[resnet50.download]]
+    sha256 = "60ad32eaf160444f3bfdb6f6d81ec1e5c36a3769be7df22aaa75127b16bb1501"
+    url = "https://huggingface.co/FluxML/resnet50/resolve/1529d6ddca42e3e705cb708c9de6f79188ce8ad5/resnet50.tar.gz"
+
+[resnext101_32x8d]
+git-tree-sha1 = "d13a85131b2c0c62ef2af79a09137d4e0760a685"
+
+    [[resnext101_32x8d.download]]
+    sha256 = "aeb48f86f50ee8b0ca7dc01ca0ff5a2d2b2163e43c524203c4a8bd589db9bcc6"
+    url = "https://huggingface.co/FluxML/resnext101_32x8d/resolve/e060f030c445f644112efa2a00e3c544944046e1/resnext101_32x8d.tar.gz"
+
+[resnext101_64x4d]
+git-tree-sha1 = "db50f48614e673a40f98fb80a17688b34f42067a"
+
+    [[resnext101_64x4d.download]]
+    sha256 = "89764dd7dc3b3432f0424cb592cec5d9db2fb802ab1646f0e3c2cca2b2e5386b"
+    url = "https://huggingface.co/FluxML/resnext101_64x4d/resolve/0d0485da04efe5a53289a560d105c42d3ca5435c/resnext101_64x4d.tar.gz"
+
+[resnext50_32x4d]
+git-tree-sha1 = "1e7a08a4acae690b635e8d1caa06e75eeb2dd2fe"
+
+    [[resnext50_32x4d.download]]
+    sha256 = "084ccbc40fde07496c401ee2bc389b9cd1d60b1ac3b7ccbfde05479ea91ca707"
+    url = "https://huggingface.co/FluxML/resnext50_32x4d/resolve/150c52c9646fe697030d38ab2be767564fb4f28c/resnext50_32x4d.tar.gz"
+
+[squeezenet]
+git-tree-sha1 = "e2eeee109fda46470d657b13669cca09d5ef2f8c"
+
+    [[squeezenet.download]]
+    sha256 = "aebfa06f44767e5ff728b7b67b2d01352b4618bd5d305c26e603aabcd5ba593d"
+    url = "https://huggingface.co/FluxML/squeezenet/resolve/01ef4221df5260bd992c669ab587eed74df0c39f/squeezenet.tar.gz"
+
 [vgg11]
 git-tree-sha1 = "78ffe7d74c475cc28175f9e23a545ce2f17b1520"
 lazy = true
@@ -29,3 +92,17 @@ lazy = true
     [[vgg19.download]]
     sha256 = "5fe26391572b9f6ac84eaa0541d27e959f673f82e6515026cdcd3262cbd93ceb"
     url = "https://huggingface.co/FluxML/vgg19/resolve/88e9056f60b054eccdc190a2eeb23731d5c693b6/vgg19.tar.gz"
+
+[wideresnet101]
+git-tree-sha1 = "b881a9469fb230faff414ce9f983bc113061ab1c"
+
+    [[wideresnet101.download]]
+    sha256 = "defa61fd80a988bb07bb9db00c692d8d0a30d95e6276add1413fb7f1f3aa2607"
+    url = "https://huggingface.co/FluxML/wideresnet101/resolve/ad4df1016bb5eba4c10e2d37b049ca5d2a455670/wideresnet101.tar.gz"
+
+[wideresnet50]
+git-tree-sha1 = "bbc6bc632e743c992784b5121dcb0f6082c66b1f"
+
+    [[wideresnet50.download]]
+    sha256 = "7596a67b7aba762c2bfce8367055da79a8a3c117bd79ce11124c9f1f5a96c4e3"
+    url = "https://huggingface.co/FluxML/wideresnet50/resolve/5eca9979f5d9438a684b1e6a5b227f9d5611965a/wideresnet50.tar.gz"

src/convnets/inception/inceptionv3.jl

@@ -174,6 +174,7 @@ See also [`inceptionv3`](#).
 struct Inceptionv3
     layers::Any
 end
+@functor Inceptionv3
 
 function Inceptionv3(; pretrain::Bool = false, inchannels::Integer = 3,
                      nclasses::Integer = 1000)
@@ -184,8 +185,6 @@ function Inceptionv3(; pretrain::Bool = false, inchannels::Integer = 3,
     return Inceptionv3(layers)
 end
 
-@functor Inceptionv3
-
 (m::Inceptionv3)(x) = m.layers(x)
 
 backbone(m::Inceptionv3) = m.layers[1]

src/convnets/resnets/resnet.jl

@@ -11,10 +11,6 @@ Creates a ResNet model with the specified depth.
   - `inchannels`: The number of input channels.
   - `nclasses`: the number of output classes
 
-!!! warning
-    
-    `ResNet` does not currently support pretrained weights.
-
 Advanced users who want more configuration options will be better served by using [`resnet`](#).
 """
 struct ResNet
@@ -52,10 +48,6 @@ The number of channels in outer 1x1 convolutions is the same.
   - `inchannels`: The number of input channels.
   - `nclasses`: the number of output classes
 
-!!! warning
-    
-    `WideResNet` does not currently support pretrained weights.
-
 Advanced users who want more configuration options will be better served by using [`resnet`](#).
 """
 struct WideResNet

src/convnets/resnets/resnext.jl

@@ -1,23 +1,23 @@
 """
-    ResNeXt(depth::Integer; pretrain::Bool = false, cardinality = 32,
-            base_width = 4, inchannels::Integer = 3, nclasses::Integer = 1000)
+    ResNeXt(depth::Integer; pretrain::Bool = false, cardinality::Integer = 32,
+            base_width::Integer = 4, inchannels::Integer = 3, nclasses::Integer = 1000)
 
 Creates a ResNeXt model with the specified depth, cardinality, and base width.
 ((reference)[https://arxiv.org/abs/1611.05431])
 
 # Arguments
 
   - `depth`: one of `[18, 34, 50, 101, 152]`. The depth of the ResNet model.
-  - `pretrain`: set to `true` to load the model with pre-trained weights for ImageNet
+  - `pretrain`: set to `true` to load the model with pre-trained weights for ImageNet.
+    Supported configurations are:
+      - depth 50, cardinality of 32 and base width of 4.
+      - depth 101, cardinality of 32 and base width of 8.
+      - depth 101, cardinality of 64 and base width of 4.
   - `cardinality`: the number of groups to be used in the 3x3 convolution in each block.
   - `base_width`: the number of feature maps in each group.
   - `inchannels`: the number of input channels.
   - `nclasses`: the number of output classes
 
-!!! warning
-    
-    `ResNeXt` does not currently support pretrained weights.
-
 Advanced users who want more configuration options will be better served by using [`resnet`](#).
 """
 struct ResNeXt
@@ -27,8 +27,8 @@ end
 
 (m::ResNeXt)(x) = m.layers(x)
 
-function ResNeXt(depth::Integer; pretrain::Bool = false, cardinality = 32,
-                 base_width = 4, inchannels::Integer = 3, nclasses::Integer = 1000)
+function ResNeXt(depth::Integer; pretrain::Bool = false, cardinality::Integer = 32,
+                 base_width::Integer = 4, inchannels::Integer = 3, nclasses::Integer = 1000)
     _checkconfig(depth, sort(collect(keys(RESNET_CONFIGS)))[3:end])
     layers = resnet(RESNET_CONFIGS[depth]...; inchannels, nclasses, cardinality, base_width)
     if pretrain

src/convnets/resnets/seresnet.jl

@@ -30,7 +30,7 @@ function SEResNet(depth::Integer; pretrain::Bool = false, inchannels::Integer =
     layers = resnet(RESNET_CONFIGS[depth]...; inchannels, nclasses,
                     attn_fn = squeeze_excite)
     if pretrain
-        loadpretrain!(layers, string("SEResNet", depth))
+        loadpretrain!(layers, string("seresnet", depth))
     end
     return SEResNet(layers)
 end
@@ -39,8 +39,8 @@ backbone(m::SEResNet) = m.layers[1]
 classifier(m::SEResNet) = m.layers[2]
 
 """
-    SEResNeXt(depth::Integer; pretrain::Bool = false, cardinality = 32, base_width = 4,
-              inchannels::Integer = 3, nclasses::Integer = 1000)
+    SEResNeXt(depth::Integer; pretrain::Bool = false, cardinality::Integer = 32,
+              base_width::Integer = 4, inchannels::Integer = 3, nclasses::Integer = 1000)
 
 Creates a SEResNeXt model with the specified depth, cardinality, and base width.
 ((reference)[https://arxiv.org/pdf/1709.01507.pdf])
@@ -67,13 +67,13 @@ end
 
 (m::SEResNeXt)(x) = m.layers(x)
 
-function SEResNeXt(depth::Integer; pretrain::Bool = false, cardinality = 32, base_width = 4,
-                   inchannels::Integer = 3, nclasses::Integer = 1000)
+function SEResNeXt(depth::Integer; pretrain::Bool = false, cardinality::Integer = 32,
+                   base_width::Integer = 4, inchannels::Integer = 3, nclasses::Integer = 1000)
     _checkconfig(depth, sort(collect(keys(RESNET_CONFIGS)))[3:end])
     layers = resnet(RESNET_CONFIGS[depth]...; inchannels, nclasses, cardinality, base_width,
                     attn_fn = squeeze_excite)
     if pretrain
-        loadpretrain!(layers, string("SEResNeXt", depth, "_", cardinality, "x", base_width))
+        loadpretrain!(layers, string("seresnext", depth, "_", cardinality, "x", base_width))
     end
     return SEResNeXt(layers)
 end

src/convnets/squeezenet.jl

@@ -62,10 +62,6 @@ Create a SqueezeNet
   - `inchannels`: number of input channels.
   - `nclasses`: the number of output classes.
 
-!!! warning
-    
-    `SqueezeNet` does not currently support pretrained weights.
-
 See also [`squeezenet`](#).
 """
 struct SqueezeNet

test/runtests.jl

@@ -18,11 +18,7 @@ const PRETRAINED_MODELS = [
     (WideResNet, 101),
     (ResNeXt, 50, 32, 4),
     (ResNeXt, 101, 64, 4),
-    (ResNeXt, 101, 32, 8),
-    (DenseNet, 121),
-    (DenseNet, 161),
-    (DenseNet, 169),
-    (DenseNet, 201),
+    (ResNeXt, 101, 32, 8)
 ]
 
 function _gc()
@@ -33,7 +29,6 @@ end
 function gradtest(model, input)
     y, pb = Zygote.pullback(() -> model(input), Flux.params(model))
     gs = pb(ones(Float32, size(y)))
-
     # if we make it to here with no error, success!
     return true
 end
@@ -50,13 +45,12 @@ const TEST_IMG = imresize(Images.load(TEST_PATH), (224, 224))
 # CHW -> WHC
 const TEST_X = permutedims(convert(Array{Float32}, channelview(TEST_IMG)), (3, 2, 1)) |> normalize_imagenet
 
-# image net labels
+# ImageNet labels
 const TEST_LBLS = readlines(download("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt"))
 
 function acctest(model)
     ypred = model(TEST_X) |> vec
     top5 = TEST_LBLS[sortperm(ypred; rev = true)]
-
     return "acoustic guitar" ∈ top5
 end