Performance comparison of deep learning frameworks in image classification problems using convolutional and recurrent networks

Deep learning methods have resulted in effective strategies for improving performance in a large number of applications, becoming one of the most used strategies by developers and researchers. In order to facilitate the implementation of those approaches, a set of software frameworks have been developed and are currently available. Selection of a specific framework is an important task, especially when computational resources are limited. In order to provide information for deciding it, this paper presents a comparative study of three of the most widely used deep learning frameworks namely Tensorflow, Theano and Torch. The comparison is carried out by implementing convolutional and recurrent architectures for classifying images from two databases: MNIST and aFAR-10. Computational costs i.e. gradient computation time, forward time and memory consumption are reported for both GPU and CPU settings.