Team name: Team Name Team members: Jaskaran Singh Walia (21BCE1089) Jesher Joshua (21BAI1925)
Introduction:
The ML/DL image classification competition was a challenge to identify four classes of eye
diseases from the dataset provided - cataract, diabetic_retinopathy, glaucoma, and normal. The
aim of the competition was to develop a deep learning model that could accurately classify the
images and achieve high accuracy.
Fig 1. Dataset’s class representation
Methodology:
We implemented deep learning methodologies to tabulate the results, and several models
were tried before selecting the final one. The models tested were VGG19, MobileNet,
GoogLeNet, Alexnet, and ElasticNet. The final model selected was Resnet50, which was fine-
tuned and optimized for the dataset. The model was trained with the SGD loss function after
trying out the Adam and LazyAdam optimizers.
Other hyperparameters used after extensive re-iteration and testing were
Batch size 64
Img size 32x32
Epochs 11
Optimizer SGD
Weights = imagenet (gave the best results)
Model Performance:
The Resnet50 model was fine-tuned and achieved an accuracy of 95% on the train dataset and
92% on the test dataset. This was achieved after hyperparameter tuning, including learning rate
adjustments and batch size adjustments. The model was trained for a total of 11 epochs.
Conclusion:
The ML/DL image classification competition was a challenging task that required us to develop
deep learning models capable of accurately classifying eye diseases. The Resnet50 model
proved to be the most effective, achieving a high accuracy of 95% on the train dataset and 92%
on the test dataset after performing hyperparameter tuning. The competition demonstrated
the power of deep learning methodologies and the importance of fine-tuning and optimizing
models to achieve the best results all of which are mentioned above.