I enjoy making things. Here are a selection of projects that I have worked on over the years.
Paper replications for various classification and object detection tasks, all implemented from scratch using PyTorch. Additionally, I have pruned certain networks to enhance their real-time performance, allowing for a tunable trade-off between accuracy and efficiency.
Searched for the set of efficient deep neural architectures (FPGANets) for image classification with two constraints: arithmetic intensity and latency for FPGA. This approach outperformed many existing networks in terms of both latency and accuracy on ImageNet-1k.
EfficientNetV2 Optimization Pruned 88% of channels of EfficientNetV2, resulting in a model that is 14x smaller, 2.5x faster, with 14x fewer parameters and just a 2% loss in accuracy.
The dataset contains images of 49 famous individuals from Nepal, including actors, actresses, singers, and social figures. The individuals vary in age, gender, and profession. Potential uses for this dataset include face recognition, face detection, emotion recognition, and training machine learning models related to facial analysis.
Prepared a dataset containing Normal, Tuberculosis (TB), and Pneumonia lung X-rays, and implemented a Deep Neural Network (DNN) to accurately classify these three categories. The challenge was to tackle the class imbalance inherent in the dataset.
