ImageKeypoints
Classify
NetTrain
NetModel
NetChain
NetMeasurements
FeatureSpacePlot3D
FeatureSpacePlot

Machine Learning
Image Processing

Train LeNet to Classify Images of FashionMNIST

Classify images of a dataset of labeled fashion images

Open in Cloud
Download Notebook

Get the necessary data from the Wolfram Data Repository:

In[1]:=

trainingData=ResourceData["FashionMNIST","TrainingData"];testData=ResourceData["FashionMNIST","TestData"];

In[2]:=

classes={0"T-shirt/top",1"Trouser",2"Pullover",3"Dress",4"Coat",5"Sandal",6"Shirt",7"Sneaker",8"Bag",9"Ankle boot"};

Get a random sample of the dataset:

In[3]:=

GroupBy[trainingData,LastFirst,RandomChoice]//KeySort

Out[3]=

0

,1

,2

,3

,4

,5

,6

,7

,8

,9



Train a small model to classify the images, using 10% of the data as a validation set:

In[4]:=

fashionNet=NetTrain[NetModel["LeNet"],trainingData,ValidationSetScaled[.1]]

Out[4]=

NetChain



Input port:	image
Output port:	class



Use the test dataset to compute the accuracy of the final classifier:

In[5]:=

NetMeasurements[fashionNet,testData,"Accuracy"]

Out[5]=

0.9107

Visualize the confusion matrix of the classifier:

In[6]:=

NetMeasurements[fashionNet,testData,"ConfusionMatrixPlot"]

Out[6]=

actual class
	predicted class

Visualize the average receiver operating characteristic (ROC) curve for every class. This plot shows the relation between the true positive rate (or recall) and the false positive rate (or fall-out) in a binary classifier:

In[7]:=

ListLinePlotValues[Transpose/@NetMeasurements[fashionNet,testData,"ROCCurve"]],



Out[7]=

	T-shirt/top
	Trouser
	Pullover
	Dress
	Coat
	Sandal
	Shirt
	Sneaker
	Bag
	Ankle boot

Compute the average ROC across all the classes:

In[8]:=

rocList=Interpolation@*Values@*(GroupBy[#1,First,Mean]&)@*Transpose/@Values[NetMeasurements[fashionNet,testData,"ROCCurve"]];ListLinePlotTable[{i,With[{m=Mean[#]},Around[m,{m-Min[#],Max[#]-m}]]&@Quiet[Through[rocList[i]]]},{i,0.0,.1,.1/30}],



Out[8]=

Use

FeatureSpacePlot3D

with the features extracted by the net to build a 3D visualization for the dataset content:

In[9]:=

FeatureSpacePlot3D[Keys[RandomSample[testData,1000]],FeatureExtractorNetDrop[fashionNet,-1],ImageSizeMedium]

Out[9]=

Related Symbols

ImageKeypoints
Classify
NetTrain
NetModel
NetChain
NetMeasurements
FeatureSpacePlot3D
FeatureSpacePlot

Publisher Information

Contributed by: Wolfram Staff

Wolfram Language Example Repository

Train LeNet to Classify Images of FashionMNIST

See Also

Publisher Information