Create METHODOLOGY-adamw.md
Browse files- METHODOLOGY-adamw.md +29 -0
METHODOLOGY-adamw.md
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
This describes my first experiments with this dataset.
|
2 |
+
I will also be tweaking it with adafactor
|
3 |
+
|
4 |
+
I was initially playing around with loras, to see which combination of images did best.
|
5 |
+
Previously, I spent a LOT of time experimenting with training and getting nowhere, so I decided to stick with
|
6 |
+
adaptive ones. Sadly, even a 4090 cant do a finetune of SDXL with an adaptive optimizer, which is why I was
|
7 |
+
training Loras.
|
8 |
+
|
9 |
+
My results were very mixed. Eventually, I figured out that stable diffusion, while seemingly magic,
|
10 |
+
cannot TRUELY figure out "good" anime style, if you throw a whole bunch of MIXED styles at it. So I
|
11 |
+
decided to drastically change my strategy, and throw out everything that was not strictly in one style.
|
12 |
+
|
13 |
+
Once I got down to <200 images images, and had a reasonable lora, I decided to give a full finetune a try,
|
14 |
+
"the hard way" (ie: no adaptive optimizer)
|
15 |
+
|
16 |
+
So, I set EMA=CPU (because not enough VRAM to fit in GPU) played with the learning rate a little, and...
|
17 |
+
that was it? Nope!
|
18 |
+
|
19 |
+
I was training with 100 epoch, and having onetrainer do an image sample every few epochs.
|
20 |
+
ut, due to disk space, I was only doing saves every 20 or so.
|
21 |
+
|
22 |
+
When I looked back at the image samples, I noticed that the one I liked best, was actually around epoch 72.
|
23 |
+
didnt know what e71,72, or 73 looked like... and I didnt even have a save for 72 either!
|
24 |
+
|
25 |
+
What I did have, was a save at 70. So I configured OneTrainer to do a new run, starting with my saved model at
|
26 |
+
70. This time, however, I disabled warmup, and also EMA. I then set OneTrainer to make a save every epoch, and a preview every epoch.
|
27 |
+
|
28 |
+
The previews looked kind of like my original series! Thus encouraged, I decided to try out all 10 of the models,
|
29 |
+
just in case. It turns out that I liked epoch 78/100 the best.. so here we are :)
|