Diabetic Retinopathy is a major threat to visual loss in working age adults. Microanueurysm, hard exudates, and small vessel growth (abnormalities) around optical disk are the early signs of diabetic retinopathy. Machine learning with new insights providing better results in disease detection at an early stage with given images automatically. This paper proposes a model to study the tuning of hyper parameters in transfer learning to classify different stages of diabetic retinopathy with fundus images from standard data set by extracting the features. This proposed work outperforms with KNN classifier and Adam as optimizer when compared to SGD. Experimental results show that it is the best method with low data size by tuning the hyper parameters. Model validation with an average validation accuracy of 75.21% achieved with K- fold cross validation technique using KNN classifier. Similarly, analysis of the RoC shows an accuracy of 82%. © 2021, Springer Nature Singapore Pte Ltd.