Magician’s Corner: 6. TensorFlow and TensorBoard

Abstract

I n a previous article, we provided an introduction to TensorFlow and used it to build a U-Net for the purpose of image segmentation (1). TensorFlow is a popular framework for deep learning applications, developed by Google and first released in 2015. TensorFlow version 2 was released in late 2019 and has several important design changes that make it much more approachable for those new to deep learning, such as direct integration with Keras, while also supporting features for advanced users. The Ten-sorFlow name comes from the fact that tensors are the fundamental computational objects in deep learning: A tensor is an n-dimensional array. Computations on those tensors are expressed as stateful dataflow graphs. TensorFlow enjoys broad hardware and software support, with versions available for tiny microcontrollers up to some of the most powerful computers in the world.

Magician’s Corner: 9. Performance Metrics for Machine Learning Models

“…While the above metrics are very useful for understanding the performance of a deep learning tool (and indeed almost any diagnostic tool), there are other metrics that are often used while training a deep learning algorithm. For instance, binary cross entropy is a very popular metric for measuring a network's performance, and the loss curves that we display (such as in Vogelsang et al [3]) reflect these metrics. There isn't room here to discuss cross-entropy calculations, but a nice explanation is provided in https://machinelearningmastery.com/cross-entropy-for-machine-learning/.…”

Automatic driving model based on machine learning agents

“…2. Tensorboard is used to view the training results [15]. By checking the cumulative reward, it could find that the model has converged and the reward has reached the maximum value.…”

Banknote Denomination Recognition on Mobile Devices

“…The MANI application, which is currently available on Google Play Store, is very accurate. The application (18) was developed by Daffodil software using a dataset of 150,000 images of Indian banknotes belonging to different classes (10,20,50,100,200, 500, and 2,000) using the image classification technique. The application (set up by a non-blind and non-deaf user) can be configured to be used by a visually impaired or visual and hearing impaired person.…”