This model achieves ~60.3% test acc in this dataset: https://drive.google.com/uc?id=1wf-3sf_zkSi1VMhVSMcWfyX7wWVEY2R5