The training and validation accuracy stuck in the value of 74% in Resnet 50

I'm new to the world of Deep Learning and I'm attempting to classify sequences of proteins into two classes using scalogram images [class1=192, class2=171] I've implemented Transfer Learning with ResNet 50 as my model architecture.the accuracy in The training and validation stuck in 73-78 %.Anyone can explain to me what happen because I confused

The sample data : [class1]:

The sample data : [class2]:

as you see the pictures in both groups look very similar

The code:

[imdsTrain,imdsVal]= splitEachLabel(imds,0.7,"Randomized");

aug= imageDataAugmenter("RandYReflection",true,"RandXReflection",true)

augTrain = augmentedImageDatastore([224,224],imdsTrain,"DataAugmentation",aug);

augVal = augmentedImageDatastore([224,224],imdsVal);

options =trainingOptions("sgdm",...

"InitialLearnRate",3e-4,...

"MaxEpochs",100,...

"Plots","training-progress",...

"Shuffle","every-epoch",...

"ValidationData",augVal,...

"ValidationFrequency",1,...

"MiniBatchSize",128,...

"L2Regularization",1e-4)

net = resnet50;

lgraph = layerGraph(net);

fclayer = lgraph.Layers(end-2);

newFclayer = fullyConnectedLayer(2,"Name","NewFc",'BiasL2Factor',10,"WeightL2Factor",10);

lgraph = replaceLayer(lgraph,fclayer.Name,newFclayer);

classlayer = lgraph.Layers(end);

newClassLayer = classificationLayer("Name","newClassLayer");

lgraph= replaceLayer(lgraph,classlayer.Name,newClassLayer);

net = trainNetwork(augTrain,lgraph,options);

The Training process:

Answer 1

[Shantanu Dixit]

Hi Moetez,

It seems that the model’s accuracy is stuck in a particular range. Although there’s no definitive answer to pinpoint the issue, the following could be potential workarounds or issues to consider:

1. Data Similarity: Since you mentioned that data from both samples are quite similar, this can make it difficult for the model to learn the underlying patterns from the data.
2. Small Data Size: Due to the small data size, you can try augmentation using ‘imageDataAugmenter’ to generate more variety.
3. Learning Rate: Instead of using a fixed learning rate throughout the training process, experiment with schedulers with different drop factors using ‘trainingOptions’.
4. Varying and Using Different Regularization Techniques: Vary the L2 regularization parameter and add dropout using ‘dropoutLayer’.

Refer to the below MathWorks documentation for more information regarding the above functions.

• Dropout: www.mathworks.com/help/deeplearning/ref/nnet.cnn.layer.dropoutlayer.html
• Image augmenter: www.mathworks.com/help/deeplearning/ref/imagedataaugmenter.html
• Training options: www.mathworks.com/help/deeplearning/ref/trainingoptions.html