About the Skin Condition Classifier
Training Data Size: 602 MiB Accuracy: 49.5%
Testing Data Size: 232 MiB Number of Diseases Classified: 9
Training Data: ISIC labelled dataset
Training Hardware: NVidia RTX 3070 TI Laptop GPU
Training environment & Libraries: Ubuntu 22.04 LTS, Conda, CUDA 12.5, PyTorch
Deployed On Ubuntu 22.04 Docker 27.1 Container With FastAPI On EC2 Instance
The model is being from its own docker container. This website is designed with CI/CD framework, microservice architecture and modularity in mind.This deep learning model was built using a pre-trained ResNet-18 architecture, which is fine-tuned to classify nine types of skin lesions. The model was trained on a modest dataset of 602 MiB with advanced data augmentation techniques to simulate real-world variations, such as random horizontal flipping, rotation, and resized cropping. Despite the small dataset size, the model achieved an accuracy of 49.5% on the test set.
The training and evaluation were performed on a machine running Ubuntu with an NVIDIA RTX 3070 TI GPU with CUDA version 12.4, which allowed efficient processing of images and faster training times. I model trained over 20 epochs using a combination of Adam optimizer and a learning rate scheduler, to improve the model performance from 37% accuracy to 49.5%.
The end user, ie, my customer is always in my mind when making my projects. In this case, you, the recruiter was on my mind, and my target wasn't just to make the model, but also make it very easy for you to use it to help you make judgements about me. I really hope I succeeded.
So little details like, making it crystal clear to you about where to click, direct you, is all a part of UX design, and then writing this piece of HTML that you're reading, choosing the right font so its a pleasure to read, containerising and deploying this Django based Website, everything is done by me.