In this thesis, part of the NFFA-Europe project, different deep learning techniques
are used in order to train several neural networks on high performance computing
facilities with the goal of classifying images of nanoscience structures captured by
SEM (scanning electron microscope). Using TensorFlow and TF-Slim as deep learning
frameworks, we train on multiple and different GPU cards several state-of-the-art
convolutional neural network (CNN) architectures (i.e. AlexNet, Inception, ResNet,
DenseNet) and test their performances in terms of training time and accuracy on
our SEM dataset. Furthermore, we coded a DenseNet implementation in TF-Slim.
Moreover, we apply and benchmark transfer learning, which consists of retraining
some pre-trained models. We then present some preliminary results, obtained in collaboration
with Intel and CINECA, about several tests on Neon, the deep learning
framework by Intel and Nervana-Systems optimized on Intel CPUs. Lastly, Inceptionv3
was ported from TF-Slim to Neon for future investigations.
