whistleroosh commited on
Commit
dd5abbc
·
verified ·
1 Parent(s): 8f095fd

Document experiment-to-checkpoint mapping

Browse files
Files changed (1) hide show
  1. README.md +32 -20
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
- Checkpoints from experiments on prediction targets, reverse equations,
17
- samplers, noise schedules, AYS and EDM2.
18
 
19
- ## Contents
 
 
20
 
21
- ### `models_final/`
22
 
23
- Final checkpoints for experiments 1–3:
24
 
25
- - files without prefix — final EMA weights,
26
- - `raw_*` — raw model weights,
27
- - `trainer_*` — training state (epoch, optimizer, scaler).
 
 
 
 
28
 
29
- Model files contain a `state_dict` and an embedded `config`.
 
30
 
31
- ### `edm2/`
32
 
33
- Official NVIDIA EDM2 ImageNet-64 S checkpoint:
34
 
35
- ```text
36
- edm2/edm2-img64-s-1073741-0.075.pkl
37
- ```
 
 
38
 
39
- This is the `edm2-img64-s-fid` preset (`FID 1.58`). It was not trained
40
- here; see the [EDM2 repository](https://github.com/NVlabs/edm2).
41
 
42
  ## Usage
43
 
44
- Download a file with `huggingface_hub`:
45
 
46
  ```python
47
  from huggingface_hub import hf_hub_download
48
 
49
  path = hf_hub_download(
50
  "whistleroosh/diffusion-thesis-models",
51
- "models_final/exp2_x0_norm_x_predictor_edm_medium_cifar10.pth",
52
  )
53
  ```
54
 
55
- Load it with the code from the thesis project:
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.