Simple question about lowering the learning rate near the end of training

Want to understand the training process better? Got tips for which model to use and when? This is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Training a Faceswap model.

If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
cosmico
Posts: 95
Joined: Sat Jan 18, 2020 6:32 pm
Has thanked: 13 times
Been thanked: 35 times

Simple question about lowering the learning rate near the end of training

Post by cosmico »

.
The questions are super simple, just let me give a little context.
.
So I've made a handful of deepfakes up to this point, and if seeing their individual sharp k-9 tooth, and seeing the shimmering of light in their pupil crystal clear can be considered 100% trained, usually I settle for 90% trained. But for the first time, I have a really good model with a really good data, and a really good chance at hitting that "100% trained". So I had a couple questions about getting there involving learning rate and batch size
.
The guide states

A learning rate set too high will constantly swing above and below the lowest value and will never learn anything. Set the learning rate set too low and the model may hit a trough and think it has reached it's lowest point, and will stop improving.
Think of it as walking down a mountain. You want to get to the bottom, so you should always be going down. However, the way down the mountain is not always downhill, there are smaller hills and valleys on the way. The learning rate needs to be high enough to be able to get out of these smaller valleys, but not so high that you end up on top of the next mountain.

Now I'm well aware that training is an exponential process, the Loss drops less and less as you go on, But I felt like my Loss wasn't dropping after a considerably long time. It stayed with in the same range it started at after I gave the model an additional 25% time training. Or in other words, of all the time it trained, in the last quarter of training the Loss didn't change. Is this sign that its time to lower the learning rate? I have good data on both the A and B side. Also should I be lowering the batch size as I near the end similar to lowering the learning rate?

User avatar
torzdf
Posts: 2651
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: Simple question about lowering the learning rate near the end of training

Post by torzdf »

Personally, I've never needed to change the learning rate.

When you're getting towards the end though, it can help to turn off "no-flip", and also "disable warp" (in the trainer settings). Just remember to turn it back on again prior to starting a new train.

My word is final

Locked