Train a Custom Image Classifier

Retrain an image classifier network to automatically distinguish dromedaries from camels

On the Wolfram Neural Net Repository, there are several models trained to identify the main object in an image. Here is one model trained on the ImageNet classification dataset:

In[1]:=

net=NetModel["Inception V1 Trained on ImageNet Competition Data"];

In its original state, it is not good at differentiating a camel (two humps) from a dromedary (one hump) because only one animal is present in the ImageNet training data:

In[2]:=

net

,"TopProbabilities"

Out[2]=



dromedary

0.999999,

dromedary

0.980626

The model can be retrained to perform better on this specific task. To do the retraining, remove the final classification layers and replace them with a two-class classifier:

In[3]:=

newNet=NetAppendNetDrop[net,-2],{"classifier"LinearLayer[2],"probabilities"SoftmaxLayer[]},"Output"NetDecoder"Class",

Bactrian camel

CONCEPT

dromedary

CONCEPT



Out[3]=

NetChain



uniniti

aliz

Input port:	image
Output port:	class



New training data can be gathered using

WebImageSearch

In[4]:=

camel=WebImageSearch[SearchQueryString["bactrian camel"],"Thumbnails",MaxItems300];

In[5]:=

dromedary=WebImageSearch[SearchQueryString["dromedary animal"],"Thumbnails",MaxItems300];

In[6]:=

trainingSet=JoinThreadcamel

Bactrian camel

CONCEPT

,Threaddromedary

dromedary

CONCEPT

;

Using

NetTrain

, the net is trained on the new data, saving 10% for validation measurements. Only the final classification layer is trained:

In[7]:=

trained=NetTrain[newNet,trainingSet,All,ValidationSetScaled[.1],LearningRateMultipliers{"classifier"1,_0},MaxTrainingRounds20]

Out[7]=

The new model gives a much better classification:

In[8]:=

trained["TrainedNet"]

,"TopProbabilities"

Out[8]=



dromedary

0.98971,

Bactrian camel

0.950061

Using the same procedure, retrain three other classification architectures trained on ImageNet and compare the accuracy of classification.

Out[8]=

	Inception V1
	Inception V3
	ResNet-50
	ResNet-101

For a simple task like this, the accuracy is similar and there is no need for the more powerful and slower models. The small ResNet is a good tradeoff between accuracy and speed:

Out[8]=

Model	TotalTrainingTime	MeanExamplesPerSecond	ValidationLoss
Inception V1	16.36	660.07	0.38
Inception V3	37.08	291.27	0.37
ResNet-50	24.32	444.11	0.36
ResNet-101	36.29	297.63	0.28

Publisher Information

Contributed by: Wolfram Staff

Train a Custom Image Classifier

Related Symbols

Publisher Information