Open Internet by MindsNet

Reproducing TensorFlow Paper in PyTorch Underperforms

A PyTorch reproduction of a TensorFlow paper's hybrid Gabor + CNN architecture underperforms by 4 percentage points on DermaMNIST. The original implementation achieves 77.01% test accuracy, while the PyTorch reproduction achieves 73-74%. The goal is to identify cross-framework issues causing this gap.

Computing & Technology, Computer Science, Machine Learning