TensorBoard is a web-based application that can help you visualize the metrics, parameters, and structure of a deep neural network created in TensorFlow. It will help you debug and optimize your deep neural networks faster and easier.

As you've probably guessed by now, deep neural networks can get quite complex. That, unfortunately, means that there are quite a few things that can go wrong. I've been known to make a mistake every now and then, and when bugs happen inside a deep neural network, which is inside a framework, that runs on another framework, that runs on a GPU, it can be very hard to find these them. TensorBoard can be the flashlight you need to find the problem in an otherwise very dark room. TensorBoard will allow you to monitor the changes in metrics and parameters as your network is trained, which can greatly accelerate troubleshooting.

TensorBoard is also great for optimization. With TensorBoard, you can visually compare multiple model runs against each other. This allows you to experiment with changing architectures and hyperparameters and then evaluate those changes relative to the other runs of the network. All this can happen throughout each epoch, so you can kill model runs that aren't doing well early if you desire, which saves you time and money.  You can read more about TensorBoard at https://www.tensorflow.org/programmers_guide/summaries_and_tensorboard.