Document experiment-to-checkpoint mapping
Browse files
README.md
CHANGED
|
@@ -5,7 +5,6 @@ tags:
|
|
| 5 |
- diffusion-models
|
| 6 |
- mnist
|
| 7 |
- cifar10
|
| 8 |
-
- imagenet64
|
| 9 |
- edm
|
| 10 |
- edm2
|
| 11 |
license: other
|
|
@@ -13,46 +12,53 @@ license: other
|
|
| 13 |
|
| 14 |
# Diffusion thesis models
|
| 15 |
|
| 16 |
-
|
| 17 |
-
samplers, noise schedules, AYS and EDM2.
|
| 18 |
|
| 19 |
-
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
|
| 22 |
|
| 23 |
-
|
| 24 |
|
| 25 |
-
|
| 26 |
-
-
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
-
|
|
|
|
| 30 |
|
| 31 |
-
|
| 32 |
|
| 33 |
-
|
| 34 |
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
|
|
|
|
|
|
| 38 |
|
| 39 |
-
|
| 40 |
-
here; see the [EDM2 repository](https://github.com/NVlabs/edm2).
|
| 41 |
|
| 42 |
## Usage
|
| 43 |
|
| 44 |
-
Download a
|
| 45 |
|
| 46 |
```python
|
| 47 |
from huggingface_hub import hf_hub_download
|
| 48 |
|
| 49 |
path = hf_hub_download(
|
| 50 |
"whistleroosh/diffusion-thesis-models",
|
| 51 |
-
"
|
| 52 |
)
|
| 53 |
```
|
| 54 |
|
| 55 |
-
Load it with the
|
| 56 |
|
| 57 |
```bash
|
| 58 |
uv run main.py generate \
|
|
@@ -65,6 +71,12 @@ uv run main.py generate \
|
|
| 65 |
--n-samples 1
|
| 66 |
```
|
| 67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
## License
|
| 69 |
|
| 70 |
The thesis checkpoints are released under the MIT license.
|
|
|
|
| 5 |
- diffusion-models
|
| 6 |
- mnist
|
| 7 |
- cifar10
|
|
|
|
| 8 |
- edm
|
| 9 |
- edm2
|
| 10 |
license: other
|
|
|
|
| 12 |
|
| 13 |
# Diffusion thesis models
|
| 14 |
|
| 15 |
+
Model weights for the bachelor's thesis:
|
|
|
|
| 16 |
|
| 17 |
+
**Score-Based Diffusion Models: Theoretical Foundations and Optimal Sampling Schedules**
|
| 18 |
+
Łukasz Leszko, Mateusz Scharmach
|
| 19 |
+
University of Warsaw, Faculty of Mathematics, Informatics and Mechanics, 2026
|
| 20 |
|
| 21 |
+
Code: [github.com/leszkolukasz/diffusers](https://github.com/leszkolukasz/diffusers)
|
| 22 |
|
| 23 |
+
## Structure
|
| 24 |
|
| 25 |
+
| Folder | Thesis experiment | Dataset | Contents |
|
| 26 |
+
|---|---|---|---|
|
| 27 |
+
| `exp1/` | Experiment 1: Prediction Target | MNIST | `noise`, `score` and `x0` predictors |
|
| 28 |
+
| `exp2/` | Experiment 2: Same network, different solvers | CIFAR-10 | EDM medium model |
|
| 29 |
+
| `exp3/` | Experiment 3: Same network, different reverse processes | CIFAR-10 | EDM medium model |
|
| 30 |
+
| `exp7/` | Pattern-dataset experiments | 2D patterns | Gaussian 8×8, 8×4, 6×6, Spiral5, Beehive |
|
| 31 |
+
| `edm2/` | External reference | ImageNet-64 | Official NVIDIA EDM2-S checkpoint |
|
| 32 |
|
| 33 |
+
Experiment 7 is not explicitly numbered in the thesis; its results are the 2D
|
| 34 |
+
pattern experiments included in the thesis materials.
|
| 35 |
|
| 36 |
+
The `exp1/` files keep their original `exp1v2_` filename prefix.
|
| 37 |
|
| 38 |
+
## File types
|
| 39 |
|
| 40 |
+
For each experiment:
|
| 41 |
+
|
| 42 |
+
- files without prefix — final EMA/inference weights,
|
| 43 |
+
- `raw_*` — raw model weights,
|
| 44 |
+
- `trainer_*` — training state (epoch, optimizer and scaler).
|
| 45 |
|
| 46 |
+
Model checkpoints contain a `state_dict` and an embedded `config`.
|
|
|
|
| 47 |
|
| 48 |
## Usage
|
| 49 |
|
| 50 |
+
Download a checkpoint:
|
| 51 |
|
| 52 |
```python
|
| 53 |
from huggingface_hub import hf_hub_download
|
| 54 |
|
| 55 |
path = hf_hub_download(
|
| 56 |
"whistleroosh/diffusion-thesis-models",
|
| 57 |
+
"exp2/exp2_x0_norm_x_predictor_edm_medium_cifar10.pth",
|
| 58 |
)
|
| 59 |
```
|
| 60 |
|
| 61 |
+
Load it with the thesis code:
|
| 62 |
|
| 63 |
```bash
|
| 64 |
uv run main.py generate \
|
|
|
|
| 71 |
--n-samples 1
|
| 72 |
```
|
| 73 |
|
| 74 |
+
## EDM2
|
| 75 |
+
|
| 76 |
+
`edm2/edm2-img64-s-1073741-0.075.pkl` is NVIDIA's official EDM2
|
| 77 |
+
ImageNet-64 S checkpoint (`edm2-img64-s-fid`, FID 1.58). It was not trained
|
| 78 |
+
by us. See the [EDM2 repository](https://github.com/NVlabs/edm2).
|
| 79 |
+
|
| 80 |
## License
|
| 81 |
|
| 82 |
The thesis checkpoints are released under the MIT license.
|