Which Deep Learning Framework Should I Use: A Comparative Study For Deep Regression Modeling

The combined impact of deep learning techniques and computing resources with an increase in the availability of databases is transforming many research fields leading to technological advances to help solve real-life and research-related problems. In the recent years, various deep learning frameworks and libraries have been developed to implement these algorithms that can operate efficiently at large scale and heterogeneous en-vironments. However, these implementations can vary depending on the framework leading to unexpected inconsistency in the results even for the same algorithm. This irregularity is seen more often in deep learning models trained using advanced parallel computing resources such as GPUs. In this study, we perform an investigation with three of the well-known deep learning frameworks: Tensorflow 1, Tensorflow 2 with Keras, and Pytorch for regression-based problems in physical sciences to analyze how the results vary depending on the framework. We implement different deep neural networks with varying complexity and perform accuracy, time, and computational based analysis to study the effect of the framework on model accuracy, training/testing times, reproducibility, and memory usage.