This repository contains a set of Jupyter notebooks that demonstrate key stages of deep learning model optimization and evaluation using the RDNetTiny architecture from Hugging Face, pretrained on ImageNet-1k, and fine-tuned on the CIFAR-10 dataset.
This notebook performs transfer learning by fine-tuning the pretrained RDNetTiny model from Hugging Face on the CIFAR-10 dataset. It includes:
- Dataset preparation and augmentation
- Model loading and customization for CIFAR-10
- Training with learning rate scheduling and performance monitoring
- Evaluation of the fine-tuned model
This notebook applies post-training quantization techniques to the transfer learned RDNetTiny model to reduce model size and improve inference speed. It includes:
- Dynamic and static quantization methods using Onnxruntime
- Accuracy evaluation and size/speed comparison with the original model
This notebook evaluates the robustness of the RDNetTiny model under common adversarial attacks. It includes:
- FGSM (Fast Gradient Sign Method) and PGD (Projected Gradient Descent) attacks
- Visualization of perturbed examples
- Accuracy degradation analysis under adversarial conditions
This notebook evaluates the robustness of the RDNetTiny model under sophisticated adversarial attacks. It includes:
- FGSM (Fast Gradient Sign Method)
- Linf Projected Gradient Descent Attack
- L2DeepFoolAttack
- L2 Projected Gradient Descent Attack
- Evaluating the the peturbation applied to dataset.
- Heatmaps for visualizing the peturbation applied to specific images and their pixel intensities.
All dependencies are installed directly within each notebook using pip. No separate setup or environment configuration is needed.