Caffe snapshots: .solverstate vs .caffemodel

Question

When training the network, pictures taken every N iterations are combined in two forms. One of them is the .solverstate file, which, I believe, is exactly what it sounds, preserving the state of the loss and gradient functions, etc. Another .caffemodel file that I know stores prepared parameters.

.caffemodel is the file you need if you want a pre-prepared model, so I assume it is also the file you want if you are going to test your network.

WWhat is .solverstate okay? In this tutorial, it looks like you can restart it, but how is this different than using .caffemodel? Does .solverstate also have the same information as .caffemodel? In other words, is this .caffemodel just a subset of .solverstate?

Answer 1

The solverstate file, as its name conveys, saves the state of the solver, and not the information associated with the classification results. The model is saved as a caffemodel file that you can use to obtain classification results for your data. If you want to set up your network, you can use the pre-prepared caffemodel file. This will save time, since your network does not need to learn from scratch. But in the event that your current training needs to be stopped due to a power outage or an unexpected reboot, you can resume your training form with the previous solver snapshot. The difference between using the solverstate file and the caffemodel file is that the former allows you to complete your training in advance, while the latter may require you to modify certain training parameters, such as the maximum number of iterations.