hearmeneigh
commited on
Commit
•
a6da21e
1
Parent(s):
f091ad0
Update README.md
Browse files
README.md
CHANGED
@@ -69,3 +69,23 @@ tags:
|
|
69 |
* [Toolchain](https://github.com/hearmeneigh/dataset-rising)
|
70 |
* [Dataset](https://huggingface.co/datasets/hearmeneigh/e621-rising-v3-curated)
|
71 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
* [Toolchain](https://github.com/hearmeneigh/dataset-rising)
|
70 |
* [Dataset](https://huggingface.co/datasets/hearmeneigh/e621-rising-v3-curated)
|
71 |
|
72 |
+
|
73 |
+
## Training Procedure
|
74 |
+
|
75 |
+
[Training legend](TBD)
|
76 |
+
|
77 |
+
* 160 images per batch (epoch variant)
|
78 |
+
* `1024x1024px` image size
|
79 |
+
* Adam optimizer
|
80 |
+
* Beta1 = `0.9`
|
81 |
+
* Beta2 = `0.999`
|
82 |
+
* Weight decay = `1e-2`
|
83 |
+
* Epsilon = `1e-08`
|
84 |
+
* Constant learning rate `4e-6`
|
85 |
+
* `fp16` mixed precision
|
86 |
+
* SNR gamma set to `5.0`
|
87 |
+
* Noise offset set to `0.07`
|
88 |
+
* `cosine_with_restarts` scheduler
|
89 |
+
* 11 epochs of [V3 curated dataset](https://huggingface.co/datasets/hearmeneigh/e621-rising-v3-curated) samples resized to `< 1024x1024px` (maintain aspect ratio)
|
90 |
+
* 16 epochs of [V3 finetuner dataset](https://huggingface.co/datasets/hearmeneigh/e621-rising-v3-finetuner) samples resized to `< 1024x1024px` (maintain aspect ratio)
|
91 |
+
* Tags for each sample are shuffled for each epoch
|