Diabetic retinopathy (DR) is a retinal disease that affects persons with long-time diabetes. It is a condition in which the blood vessels of the retina get damaged and a fatal condition that causes vision loss in people. The primary cautioning signs of DR aid in vision loss detection. Detecting DR manually consumes time and is expensive. To overcome this, an automated system is required that can significantly classify the different levels of DR, where deep learning (DL) became more prominent and massively used for the early-stage detection. This study proposes a novel MobileNet V2 classifier employing a transfer learning (TL) which plays a crucial role in achieving better results in classifying the stages of DR. The training and test sets were randomly generated and split into percentages of 70 and 30% images from Kaggle database. Feature extraction is done by performing depth-wise separable convolution, which leads to computations and memory requirement reduction, The work supports the function of deep learning (DL) classifier in enhancing detection capabilities for complicated medical problems and adds to the continuing progress in medical image analysis. The proposed model achieved 85.7% of accuracy and sensitivity of 85.82%.