Add pretrained weights for MobileNet v2 and v3

* Adds pre-trained weights for MobileNet v3 small and large models
* Adds pre-trained weights for MobileNet v2
* Adds, commented, lines to the model_list in the port_torchvision.jl
  script for generating these weights from torchvision models
* Replaces the test image (of a guitar) in the pytorch2flux.jl script,
  the link is currently unavailable

Author: paulnovo

Artifacts.toml

@@ -1,3 +1,27 @@
+[mobilenet_v2]
+git-tree-sha1 = "1552eb3a913b8ea964482bd68a8fd6e623f57b45"
+lazy = true
+
+    [[mobilenet_v2.download]]
+    sha256 = "24d8afb8897cb82825c51e65c7a02f4b2117581294de25ac7fd5222bbc15ae6e"
+    url = ""
+
+[mobilenet_v3_small]
+git-tree-sha1 = "9b1f0550051c731056cda7739fcdd115c36c04ad"
+lazy = true
+
+    [[mobilenet_v3_small.download]]
+    sha256 = "35b7f6d733bfbd1621349a0ec27391c2714add7ca80006fb4032d8bc66629c97"
+    url = ""
+
+[mobilenet_v3_large]
+git-tree-sha1 = "49971ff8327bc591885e78ff94140ee472f77329"
+lazy = true
+
+    [[mobilenet_v3_large.download]]
+    sha256 = "555fcb5f4f6574d77b603b2fc6672ab437ef60a1a20bc2f951122c91aaaf2f69"
+    url = ""
+
 [resnet101]
 git-tree-sha1 = "68d563526ab34d3e5aa66b7d96278d2acde212f9"
 lazy = true

scripts/Project.toml

@@ -3,10 +3,12 @@ ArtifactUtils = "8b73e784-e7d8-4ea5-973d-377fed4e3bce"
 BSON = "fbb218c0-5317-5bc6-957e-2ee96dd4b1f0"
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
 Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
+Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 HuggingFaceApi = "3cc741c3-0c9d-4fbe-84fa-cdec264173de"
 Images = "916415d5-f1e6-5110-898d-aaa5f9f070e0"
 JLD2 = "033835bb-8acc-5ee8-8aae-3f567f8a3819"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 Metalhead = "dbeba491-748d-5e0e-a39e-b530a07fa0cc"
 PythonCall = "6099a3de-0909-46bc-b1f4-468b9a2dfc0d"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
+TestImages = "5e47fb64-e119-507b-a336-dd2b206d9990"

scripts/port_torchvision.jl

@@ -33,6 +33,9 @@ model_list = [
             #   ("vit_b_32", "IMAGENET1K_V1", () -> ViT(:base, patch_size=(32,32)), weights -> tvmodels.vit_b_32(; weights)),
             #   ("vit_l_16", "IMAGENET1K_V1", () -> ViT(:large), weights -> tvmodels.vit_l_16(; weights)),
             #   ("vit_l_32", "IMAGENET1K_V1", () -> ViT(:large, patch_size=(32,32)), weights -> tvmodels.vit_l_32(; weights)),
+            #   ("mobilenet_v2", "IMAGENET1K_V2", () -> MobileNetv2(), weights -> tvmodels.mobilenet_v2(; weights)),
+            #   ("mobilenet_v3_small", "IMAGENET1K_V1", () -> MobileNetv3(:small), weights -> tvmodels.mobilenet_v3_small(; weights)),
+            #   ("mobilenet_v3_large", "IMAGENET1K_V2", () -> MobileNetv3(:large), weights -> tvmodels.mobilenet_v3_large(; weights)),
               ## NOT WORKING:
               ("densenet121", "IMAGENET1K_V1", () -> DenseNet(121), weights -> tvmodels.densenet121(; weights)),
               # ("squeezenet1_0", "IMAGENET1K_V1", () -> SqueezeNet(), weights -> tvmodels.squeezenet1_0(; weights)),

scripts/pytorch2flux.jl

@@ -10,19 +10,18 @@ using BSON
 using PythonCall
 using Images
 using Test
+using TestImages
 
 include("utils.jl")
 
 const torch = pyimport("torch")
 const torchvision = pyimport("torchvision")
 
-# test image
-const GUITAR_PATH = download("https://cdn.pixabay.com/photo/2015/05/07/11/02/guitar-756326_960_720.jpg")
 const IMAGENET_LABELS = readlines(download("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt"))
 
 function compare_pytorch(jlmodel, pymodel; rtol = 1e-4)
     sz = (224, 224)
-    img = Images.load(GUITAR_PATH);
+    img = testimage("monarch_color_256")
     img = imresize(img, sz);
     # CHW -> WHC
     data = permutedims(convert(Array{Float32}, channelview(img)), (3,2,1))

src/convnets/mobilenets/mobilenetv2.jl

@@ -57,10 +57,6 @@ Create a MobileNetv2 model with the specified configuration.
   - `inchannels`: The number of input channels.
   - `nclasses`: The number of output classes
 
-!!! warning
-    
-    `MobileNetv2` does not currently support pretrained weights.
-
 See also [`Metalhead.mobilenetv2`](@ref).
 """
 struct MobileNetv2

src/convnets/mobilenets/mobilenetv3.jl

@@ -78,10 +78,6 @@ Set `pretrain = true` to load the model with pre-trained weights for ImageNet.
   - `inchannels`: number of input channels
   - `nclasses`: the number of output classes
 
-!!! warning
-    
-    `MobileNetv3` does not currently support pretrained weights.
-
 See also [`Metalhead.mobilenetv3`](@ref).
 """
 struct MobileNetv3
@@ -94,7 +90,7 @@ function MobileNetv3(config::Symbol; width_mult::Real = 1, pretrain::Bool = fals
     layers = mobilenetv3(config; width_mult, inchannels, nclasses)
     model = MobileNetv3(layers)
     if pretrain
-        loadpretrain!(model, "mobilenet_v3")
+        loadpretrain!(model, string("mobilenet_v3_", config))
     end
     return model
 end