training

docs/train.md

@@ -268,3 +268,7 @@ Because we use zero convolutions, the SD should always be able to predict meanin
 You will always find that at some iterations, the model "suddenly" be able to fit some training conditions. This means that you will get a basically usable model at about 3k to 7k steps (future training will improve it, but that model after the first "sudden converge" should be basically functional).
 
 Note that 3k to 7k steps is not very large, and you should consider larger batch size rather than more training steps. If you can observe the "sudden converge" at 3k step using batch size 4, then, rather than train it with 300k further steps, a better idea is to use 100× gradient accumulation to re-train that 3k steps with 100× batch size. Note that perhaps we should not do this *too* extremely (perhaps 100x accumulation is too extreme), but you should consider that, since "sudden converge" will *always* happen at that certain point, getting a better converge is more important.
+
+Because that "sudden converge" always happens, lets say "sudden converge" will happen at 3k step and our money can optimize 90k step, then we have two options: (1) train 3k steps, sudden converge, then train 87k steps. (2) 30x gradient accumulation, train 3k steps (90k real computation steps), then sudden converge.
+
+In my experiments, (2) is usually better than (1). However, in real cases, perhaps you may need to balance the steps before and after the "sudden converge" on your own to find a balance. The training after "sudden converge" is also important.