File size: 5,358 Bytes
59aac48 4eccf06 59aac48 4eccf06 59aac48 4eccf06 59aac48 4eccf06 59aac48 4eccf06 59aac48 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 | ---
library_name: gendynamics
license: mit
tags:
- diffusion
- unconditional-generation
---
# synthetic_alpha_stable_mixture
Unconditional reference models trained with gendynamics. The weights correspond to the best validation epoch.
## Dataset
Synthetic unbalanced high-dimensional alpha-stable mixture. Generated by `gendynamics.datasets.fetch_synthetic_data('unbalanced_highdim_alpha_stable_mixture')`. Sample the configured mixture, then make a seeded random train/validation/test split.
The train/validation/test splits contain 25,000/1,000/10,000 samples
of shape `[150]`. Normalization: None
The exact generator arguments or processed-file identity, split settings, and seed are in `dataset.json`.
The source data themselves are not redistributed here; source-data terms remain with their provider.
The preparation code is in [gendynamics](https://github.com/Diffusion-Research-Lab/gendynamics). API-backed sources can change; the processed-file checksum identifies this snapshot.
## Architecture and training
All independently trained models use this network architecture:
```json
{
"name": "mlp",
"params": {
"width": 240,
"depth": 4,
"time_dim": 128,
"dropout": 0.0,
"use_norm": true,
"dim": 150
}
}
```
| Directory | Model | Prediction target | Trainable parameters | Sampling steps | Learning rate | Epochs run / best | Val loss | Test loss |
|---|---|---|---:|---:|---:|---:|---:|---:|
| ddpm-v | DDPM-V | velocity | 693,894 | 128 | 0.0005 | 562 / 462 | 8.425 | 4.067 |
| dlpm-eps | DLPM-Eps | noise | 693,894 | 128 | 0.0005 | 256 / 156 | 0.8574 | 0.9514 |
| tedm-origin | t-EDM | denoised data | 693,894 | 64 | 0.0005 | 158 / 58 | 1.412 | 3.308 |
The complete **selected final** model and training settings are below and in each directory's `config.json`.
The training settings include the maximum epoch budget; `epochs run / best` reports what actually happened.
`provenance.json` records software versions and the source revision. Each model was reloaded and checked before upload;
`export_check.json` records that check. Losses use each model's own objective, so they are not comparable sample-quality scores.
The number of held-out rows used for each test loss is in `metrics.json`. Sample previews do not establish tail accuracy.
### DDPM-V (`ddpm-v/`)
```json
{
"model_parameters": {
"n_steps": 128,
"sigma_max": 2.0,
"sampler": "ddpm"
},
"training": {
"device": "cuda",
"data_device": "cpu",
"batch_size": 1024,
"n_epochs": 640,
"early_stopping_patience": 100,
"lr": 0.0005,
"num_workers": 0,
"use_adamw": true,
"weight_decay": 1e-06,
"grad_clip_norm": 10.0,
"lr_schedule": "cosine",
"warmup_steps": 400,
"cosine_eta_min_ratio": 0.05,
"freq_logging": 50,
"stats_freq_epochs": 1,
"log_grad_norm": true,
"ckpt_freq_epochs": 64,
"ckpt_keep_last": 8
}
}
```
### DLPM-Eps (`dlpm-eps/`)
```json
{
"model_parameters": {
"n_steps": 128,
"alpha": 1.8,
"sampler": "native"
},
"training": {
"device": "cuda",
"data_device": "cpu",
"batch_size": 1024,
"n_epochs": 640,
"early_stopping_patience": 100,
"lr": 0.0005,
"num_workers": 0,
"use_adamw": true,
"weight_decay": 1e-06,
"grad_clip_norm": 10.0,
"lr_schedule": "cosine",
"warmup_steps": 400,
"cosine_eta_min_ratio": 0.05,
"freq_logging": 50,
"stats_freq_epochs": 1,
"log_grad_norm": true,
"ckpt_freq_epochs": 64,
"ckpt_keep_last": 8
}
}
```
### t-EDM (`tedm-origin/`)
```json
{
"model_parameters": {
"n_steps": 64,
"nu": 2.2,
"sigma_min": 0.005,
"sigma_max": 5.0,
"sigma_data": 1.0,
"solver": "edm_stochastic_heun"
},
"training": {
"device": "cuda",
"data_device": "cpu",
"batch_size": 1024,
"n_epochs": 640,
"early_stopping_patience": 100,
"lr": 0.0005,
"num_workers": 0,
"use_adamw": true,
"weight_decay": 1e-06,
"grad_clip_norm": 10.0,
"lr_schedule": "cosine",
"warmup_steps": 400,
"cosine_eta_min_ratio": 0.05,
"freq_logging": 50,
"stats_freq_epochs": 1,
"log_grad_norm": true,
"ckpt_freq_epochs": 64,
"ckpt_keep_last": 8
}
}
```
DDIM-V (`ddim-v/`) shares the trained DDPM-V weights. Only the reverse sampler changes to deterministic DDIM (`eta=0`); it has no separate training run. Its exact sampling settings are in `ddim-v/config.json`.
## Loading
Install the training package and its dependencies:
python -m pip install git+https://github.com/Diffusion-Research-Lab/2_training_2026_tail_reference_models.git
Then load one of the listed model directories by its underscored name:
from toolkit.huggingface import load_huggingface_model
model, normalization, data_scale = load_huggingface_model("Diffusion-Research-Lab/alpha-stable-mixture-1p7-150d-diffusion", "ddpm_v", device="cpu")
samples = model.sample(16).float().cpu()
if normalization is not None:
samples = samples * normalization["std"] + normalization["mean"]
The model-specific files are `model.safetensors`, `config.json`, `dataset.json`, `metrics.json`,
`provenance.json`, `export_check.json`, and, when applicable, `normalization.safetensors`.
This repository is released under the MIT license. Cite the model repository revision when using these weights.
|