Hands-on Coding Session on Debugging Neural Networks Programs

The goal of this hands-on session is to practice troubleshooting neural network training programs. The provided example consists of a VGG-like ConvNet trained on CIFAR-10 classification. If we fix the injected bugs (that I know), it should achieve more than 80% accuracy on the test set after more than 20 epochs. The debugging process is more relevant than finding the bug. Hence, we will walk through the steps you should follow towards hunting bugs. It is not about doing an in-depth review of code based on the participant’s Tensorflow coding skills, but rather about codifying verification routines in order to monitor the dynamics of critical variables and components that may indicate coding errors, misconfigurations, or training anomalies.