Training was performed using NVIDIA Titan RTX video card with 24 GB of memory. With presented dataset of ~2700 RGB images, each 320 by 320 pixels, training speed was 2 minutes per one epoch. It has to be outlined that NNABLA library is using CUDA and cuDNN architecture very efficiently, with GPU utilization rate reaching 90%-95%.
Below is the example of the learning curve:
References:
- Sony Corporation. Neural Network Console : Not just train and evaluate. You can design neural networks with fast and intuitive GUI. https://dl.sony.com/
- Sony Corporation. Neural Network Libraries : An open source software to make research, development and implementation of neural network more efficient. https://nnabla.org/
- BatchNormalization – Ioffe and Szegedy, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. https://arxiv.org/abs/1502.03167
- Convolution – Chen et al., DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. https://arxiv.org/abs/1606.00915
- Yu et al., Multi-Scale Context Aggregation by Dilated Convolutions. https://arxiv.org/abs/1511.07122
- ReLU – Vinod Nair, Geoffrey E. Hinton. Rectified Linear Units Improve Restricted Boltzmann Machines. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.165.6419&rep=rep1&type=pdf
- Momentum – Matthew D. Zeiler. ADADELTA: An Adaptive Learning Rate Method. https://arxiv.org/abs/1212.5701
- U-Net: Convolutional Networks for Biomedical Image Segmentation, Olaf Ronneberger, Philipp Fischer, Thomas Brox. https://arxiv.org/abs/1505.04597