Update README.md
Browse files
README.md
CHANGED
|
@@ -197,25 +197,6 @@ If you would like to use the regularization images from this training, see the d
|
|
| 197 |
|
| 198 |
- `landscape`: https://huggingface.co/datasets/3ee/regularization-landscape
|
| 199 |
|
| 200 |
-
---
|
| 201 |
-
|
| 202 |
-
## 📊 Training Stats
|
| 203 |
-
|
| 204 |
-
🔋 GPU: `Nvidia 3080 16GB`
|
| 205 |
-
|
| 206 |
-
🔬 Learning Rate: `1e-6` & Training steps per image: `96` (`13400` total)
|
| 207 |
-
|
| 208 |
-
💬 Text Encoder Training: `0.2`
|
| 209 |
-
|
| 210 |
-
🎨 Total input images per concept: `14` (total = `14 * 4 = 56`)
|
| 211 |
-
|
| 212 |
-
☕ cups of coffee consumed: `21`
|
| 213 |
-
|
| 214 |
-
- I spent a lot of time experimenting and fine tuning. I started with `25` steps per image at first and went on from there.
|
| 215 |
-
|
| 216 |
-
- I went as high as `176` steps per image and attempted many different `batch sizes` with `Gradient Accumulation`. The results were different (but really good!) than what I had planned for this model which is: capture the style of SPOP. I plan to use the knowledge I gained from those "failed" models to great use on future ones!
|
| 217 |
-
|
| 218 |
-
- In conclusion, the `13400` version came out to be the version that captured the Dreamworks style.
|
| 219 |
|
| 220 |
---
|
| 221 |
|
|
|
|
| 197 |
|
| 198 |
- `landscape`: https://huggingface.co/datasets/3ee/regularization-landscape
|
| 199 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 200 |
|
| 201 |
---
|
| 202 |
|