I am implementing the model in Matlab R2018a. I have twenty categories. My training accuracy goes to 98% but validation accuracy is 60-66%. If I reduce overfiitting , then both comes to near about 66%.
Well, it is clear that it is not a matter of what algorithm you are using. Let me tell you a trick to improve accuracy when you are having good results during training but not testing.
1) It is due to less amount of data.
2) Data is not versatile
to overcome this you can use python to create more data using keras library (that makes your training data less prone to overfit)
If even after that you are unable to improve your accuracy, I recommend you to post your model architecture (i.e. no. Of cnn layers, max pool layers, dropouts etc.) And also the no. of data points you are using for training.
Yes, now the Validation Accuracy is increased to 88.40% and training accuracy goes nearly to hundred.If I increase the epochs, validation accuracy also increases, but final validation accuracy suddenly drops to above 88.40%..I am stuck here.
Deepali Bongulwar, for the fitting of your modal one your images, make sure you choose LR (learning rate) wisely. For example, you can make a Losses vs LR curve and find the values where cure is steepest/max slope as in fig. attached below it is slice(1e-4, 1e-2).
or you can follow Subash Gautam advice and use TF with half of the efforts you are putting right now.