PyTorch reproduction of TensorFlow paper underperforms by 4 pp on DermaMNIST , what cross-framework issues should I check? [R]

About this article

I'm reproducing a published paper's hybrid Gabor + CNN architecture in PyTorch. The original implementation is in TensorFlow. My reproduction consistently lands ~4 pp below the paper's reported test accuracy on DermaMNIST (73-74% vs paper's 77.01%). I'd like to know which cross-framework differences are most likely to cause this gap. Ahmed et al., "A Lightweight Hybrid Gabor Deep Learning Approach", IJCV 2026 (DOI: 10.1007/s11263-025-02658-2). The architecture is a fixed Gabor filter bank fro...