Apprentissage non supervisé avec les auto-encodeurs: Nouveautés de Wolfram Language 11

Apprentissage non supervisé avec les auto-encodeurs

Entraînez un réseau auto-encodeur pour reconstruire des images de chiffres manuscrits après les avoir projetés à un "code" espace vectoriel de dimensions inférieures. Utilisez ces vecteurs de code pour effectuer des regroupements et des visualisations.

Tout d'abord obtenez les données d'entraînement, puis sélectionnez les images correspondant aux chiffres de 0 à 4.

In[1]:=

resource = ResourceObject["MNIST"];
trainingData = ResourceData[resource, "TrainingData"];
trainingSubset = Select[trainingData, Last[#] <= 4 &];
testData = ResourceData[resource, "TestData"];
testSubset = Select[testData, Last[#] <= 4 &];
RandomSample[trainingSubset, 8]

Out[1]=

Obtenez la "l'image moyenne" pour soustraire des données d'entraînement.

In[2]:=

trainingImages = Keys[trainingSubset];
meanImage = Image[Mean@Map[ImageData, trainingImages]]

Out[2]=

Créez un réseau pour entraîner qui produit à la fois la reconstruction et l'erreur de reconstruction.

In[3]:=

net = NetGraph[
  {FlattenLayer[], 50, Ramp, 784, Tanh, ReshapeLayer[{1, 28, 28}], 
   MeanSquaredLossLayer[]},
  {1 -> 2 -> 3 -> 4 -> 5 -> 6 -> NetPort["Output"], 
   6 -> NetPort[7, "Input"], NetPort["Input"] -> NetPort[7, "Target"]},
  "Input" -> 
   NetEncoder[{"Image", {28, 28}, "Grayscale", 
     "MeanImage" -> meanImage}],
  "Output" -> NetDecoder[{"Image", "Grayscale"}]
  ]

Out[3]=

Entraînez le réseau afin de minimiser l'erreur de reconstruction.

In[4]:=

trained = 
  If[$VersionNumber < 11.3, 
   NetTrain[net, <|"Input" -> trainingImages|>, "Loss"], 
   NetTrain[net, <|"Input" -> trainingImages|>, LossFunction -> "Loss"]
   ];

Out[4]=

Obtenez un sous-réseau qui exécute seulement les reconstructions.

In[5]:=

reconstructor = Take[trained, {NetPort["Input"], NetPort["Output"]}]

Out[5]=

Reconstruisez quelques exemples d'images.

In[6]:=

ImageAdd[reconstructor[#], meanImage] & /@ {\!\(\*
GraphicsBox[
TagBox[RasterBox[CompressedData["
1:eJxTTMoPSmNiYGAo5gASQYnljkVFiZXBAkBOaF5xZnpeaopnXklqemqRRRJI
mQwU/x9A8CWTyfQBDrlbLCxMU7BLvbLCKTfRkQUoFzbpIBY5JhaQHAuL8hkM
KU8GJiAQVQQS6FIHlED6cjYdrGdhmYYqdV8CaJxy2df//x9IsvD0/UJ3vvNr
MHMSUNUdNDmzhxDmA3N0OYQL7psyMEUjyxWzsMCYGGaqweReHQC6SvIhVrl8
oE+UDv3HJuepBJTzRvWfKhPTtm3STEzgwEGV+t8HDkoIkYMmBwwNiJyk052v
aHL/DxZC5LDH3/ZAlqAd2x9ilaMPAABwyokL
"], {{0, 28}, {28, 0}}, {0, 255},
       
ColorFunction->GrayLevel],
BoxForm`ImageTag[
      "Byte", ColorSpace -> Automatic, Interleaving -> None],
Selectable->False],
DefaultBaseStyle->"ImageGraphics",
ImageSizeRaw->{28, 28},
PlotRange->{{0, 28}, {0, 28}}]\), \!\(\*
GraphicsBox[
TagBox[RasterBox[CompressedData["
1:eJxTTMoPSmNiYGAo5gASQYnljkVFiZXBAkBOaF5xZnpeaopnXklqemqRRRJI
mQwU/x9gcJNxEk65FcxrccpV8OGUusSdiVNuNeMBnHKmCl9wSd1nVMepbQGj
DU65YsZNuKSOCRl9xyXXzBCF08gQxnW4pJ6La+DU1s6YgFMug7EQp5w0415c
UodYcMsVMRr9wSH1VYOxDZe2X5b+X3E6hS4AAAjL8xI=
"], {{0, 28}, {28, 0}}, {
       0, 255},
ColorFunction->GrayLevel],
BoxForm`ImageTag[
      "Byte", ColorSpace -> Automatic, Interleaving -> None],
Selectable->False],
DefaultBaseStyle->"ImageGraphics",
ImageSizeRaw->{28, 28},
PlotRange->{{0, 28}, {0, 28}}]\), \!\(\*
GraphicsBox[
TagBox[RasterBox[CompressedData["
1:eJxTTMoPSmNiYGAo5gASQYnljkVFiZXBAkBOaF5xZnpeaopnXklqemqRRRJI
mQwU/6cp6G4KYWDIXIRNKowJDFQfYpfSKvRnYmrFkDrNyqR7//P/n4ZMxRhy
m1h0nwGpVnamfZhmPngLIvWYsMmBQRcHk+VX7FKbOZgkDuDQVs/ElI9Dyp+T
KeEzdqlnokxid3Bos2RiKsIhtZGdyQmHiW/McWurZGIKwqHtPzsT0zMcUiC5
K6+B4Nf/X69f38rMzMz5iiwHBuEFkVBWC1wukAkB2Dg4wrq6jiEM7WxpaQFp
SWlpuYbL4gEEADLRqO8=
"], {{0, 28}, {28, 0}}, {0, 255},
ColorFunction->GrayLevel],
BoxForm`ImageTag[
      "Byte", ColorSpace -> Automatic, Interleaving -> None],
Selectable->False],
DefaultBaseStyle->"ImageGraphics",
ImageSizeRaw->{28, 28},
PlotRange->{{0, 28}, {0, 28}}]\), \!\(\*
GraphicsBox[
TagBox[RasterBox[CompressedData["
1:eJxTTMoPSmNiYGAo5gASQYnljkVFiZXBAkBOaF5xZnpeaopnXklqemqRRRJI
mQwU/x8Q8Plabu55HFJ1TExMrDFvsclVMYGB1E4scrOYmPNmThFi4qj/iiHn
xRQBJA+JMDHF/EKXY2C6AqKO2jExRf9Gk3Nhug+mTwgzMa1Ek8sFys12XLFi
RQ4TUwua3EKmjdfZIW5lUlj9AUXu4+LPd/iZYIB7Gbprt8T5wGX1rqC79s/7
9zduvn+fx8PE5IYagK+PwlhH9ZiYPJClNilxbICxP6kw8W9HklvKwcR8DMY5
xcLkgKyxm4npCIw9h5kpD8UfzkxKiTdBrAk6nEyocv8/K3EysbABAcgX5m/Q
PLHAQxriP5vmF+ge/P//+a3SrY3Lb/3AlKEHAAAeUY84
"], {{0, 28}, {28, 0}}, {
       0, 255},
ColorFunction->GrayLevel],
BoxForm`ImageTag[
      "Byte", ColorSpace -> Automatic, Interleaving -> None],
Selectable->False],
DefaultBaseStyle->"ImageGraphics",
ImageSizeRaw->{28, 28},
PlotRange->{{0, 28}, {0, 28}}]\), \!\(\*
GraphicsBox[
TagBox[RasterBox[CompressedData["
1:eJxTTMoPSmNiYGAo5gASQYnljkVFiZXBAkBOaF5xZnpeaopnXklqemqRRRJI
mQwU/x+U4Kw8lLHzEYZcmwSUkR2OLvXbBiY3X/cLmtwu5kooq5f5FarUJWG1
z1CmPbpcOMcpKOstAxOq3GpeHRiziMnpF4pcGPNUKOu+OOteFKkPcswwZiWz
DorU/5fM0QgTIlDlvhnrvYWqYmKagioHVG2xCgjqom2Y4TbDwLVQLmYgEJdg
YWb+9h8dnFsNBP//xzEzY0jBQAMz8yVccvVMTPj04ZSrYObGKScuPAGnnM9e
nFJUAQDVl858
"], {{0, 28}, {28, 0}}, {0, 255},
ColorFunction->GrayLevel],
BoxForm`ImageTag[
      "Byte", ColorSpace -> Automatic, Interleaving -> None],
Selectable->False],
DefaultBaseStyle->"ImageGraphics",
ImageSizeRaw->{28, 28},
PlotRange->{{0, 28}, {0, 28}}]\)}

Out[6]=

Obtenez un sous-réseau qui produit le vecteur de code.

In[7]:=

encoder = Take[trained, {NetPort["Input"], 4}]

Out[7]=

Calculez les codes pour toutes les images de test.

In[8]:=

testImages = Keys[testSubset];
features = encoder[testImages];

Projetez les vecteurs de code en trois dimensions et visualisez-les avec les classes originales (non vues par le réseau). Les classes de chiffres ont tendance à se regrouper.

In[9]:=

coords = DimensionReduce[features, 3];
classes = Values[testSubset];
Table[Extract[coords, Position[classes, i]], {i, 0, 4}]
ListPointPlot3D[
 Table[Extract[coords, Position[classes, i]], {i, 0, 4}], 
 PlotLegends -> PointLegend[96, Range[0, 4]], BoxRatios -> 1, 
 Axes -> None, Boxed -> True,
 PlotStyle -> Map[ColorData[96], Range[1, 5]], AspectRatio -> 1]

Out[9]=

Effectuez la classification automatique directement sur les vecteurs de code et montrez un échantillon prélevé de chaque groupe.

In[10]:=

components = ClusteringComponents[features, 5, 1];
Map[Part[testImages, RandomSample[#, 10]] &, 
 PositionIndex[components]]

Out[10]=

Visualisez une classification hiérarchique des représentants aléatoires de chaque classe.

In[11]:=

representatives = 
  Catenate@GroupBy[testSubset, Last -> First, RandomSample[#, 6] &];
ClusteringTree[
 encoder[representatives] -> Map[ImageCrop, representatives]]

In[12]:=

\!\(\*
GraphicsBox[RasterBox[CompressedData["
1:eJztnQd0FFX7xjG0QEgIoYYekCpIByXSwUIAURBE/BApKh8ofoJ8gAWkg1Gk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"], {{0, 0}, {360, 258}}, {0,
     255},
ColorFunction->RGBColor],
ImageSize->{360, 258},
PlotRange->{{0, 360}, {0, 258}}]\)

Wolfram Language™

Apprentissage non supervisé avec les auto-encodeurs

Exemples connexes