Loss value depends on data you put through the network and its current state (i.e. weights). Caffemodel & solverstate only contain information on weights (plus gradients, which do not matter here), therefore you have no way of reconstructing the loss.
You can check your train log (if you saved it) for intermediate loss values, but other than that you have to evaluate it manually (load the network in TRAIN phase, propagate an image forward and check the loss).
I suppose Alex meant just that in his quality post ;)