Diabetic Retinopathy is a complication that affects the retina of a diabetic patient which is characterized by blood leakage and growth of new blood vessels inside the retina that can cause blindness if left untreated. Early Identification of DR is becoming the focus for physicians and researchers in the medical engineering field. Different DL techniques were proposed for this problem. But most of them did not achieve good performance and assume there is enough data set, thus the issue of limited training set and the overfitting problem is not well discussed. In the medical field finding adequate data is difficult due to privacy, sensitivity & expert labeling. To overcome this problem many researchers use traditional augmentation techniques to generate additional data & balance the class distribution, but traditional augmentation has its problems in perfectly representing the real features in the image, processes like zooming, shifting, and flipping has a tendency to distract, invert and hide significant features, this has a great impact on model performance. This thesis comes up with a different method combining GAN & transfer learning for building an efficient classification model. GANs were used to generate class-specific images for balancing the class distribution & overcome overfitting. Pre-trained models were employed to allow the DL model to work on small data set for better performance, and speed than training neural networks from scratch, and also it can decrease the need for hardware resources. Several experiments done with balanced dataset & fine-tuned ResNet152, MobileNetV2, EfficientNetB1, DenseNet, InceptionResNetV2, InceptionV3. DensNet201 achieved 88% testing accuracy and EffiientNetB1 achieved 87% testing accuracy. Both models achieved the best performance with fewer datasets and lite weight pre-trained models, compared to previously done works. Key words: GAN, CNN, DR, Transfer Learning, Adam, Adamax