The Importance of High Speed Storage in Deep Learning Training

With the increase of computational power and techniques over the past decades, the use of Deep Learning (DL) algorithms in the biomedical field has grown significantly. One of the remaining challenges to using deep neural networks is the proper tuning of the model's performance beyond its simple accuracy. Therefore, in this work, we implement the combination of the NVIDIA DALI API for high-speed storage access alongside the TensorFlow framework, applied to the image classification task of skin cancer. To that end, we use the VGG16 model, known to perform accurately on skin cancer classification. We compare the performance between the use of CPU, GPU and multi-GPU devices training both in terms of accuracy and runtime performance. These performances are also evaluated on additional models, as a mean for comparison. Our work shows the high importance of model choice and fine tuning tailored to a particular application. Moreover, we show that the use of high-speed storage considerably increases the performance of DL models, in particular when handling images and large databases which may be a significant improvement for larger databases.