Forbu14 commited on
Commit
762d41b
·
verified ·
1 Parent(s): d85412e

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +3 -62
README.md CHANGED
@@ -1,62 +1,3 @@
1
- ---
2
- language: en
3
- tags:
4
- - weather-forecasting
5
- - diffusion-models
6
- - rectified-flow
7
- - meteorology
8
- - pytorch
9
- - deep-learning
10
- license: mit
11
- datasets:
12
- - meteolibre
13
- ---
14
-
15
-
16
-
17
-
18
- # MeteoLibre Rectified Flow Model
19
-
20
- In the folder models_shortcut/:
21
-
22
- ## Model Description
23
-
24
- - **Model type**: Rectified Flow Diffusion Model
25
- - **Architecture**: 3D U-Net with FiLM conditioning
26
- - **Input**: Meteorological data patches (12 channels + 1 lightning channels, 3D spatio-temporal)
27
- - **Output**: Generated weather forecast data
28
- - **Training data**: MeteoLibre meteorological dataset
29
- - **Language(s)**: Python
30
- - **License**: MIT
31
-
32
- ## Training
33
-
34
- The model was trained using:
35
- - **Framework**: PyTorch with Hugging Face Accelerate
36
- - **Optimizer**: Adam (lr=5e-4) OR SOAP
37
- - **Batch size**: 64
38
- - **Epochs**: 200
39
- - **Precision**: Mixed precision (bf16)
40
- - **Distributed training**: Multi-GPU support
41
-
42
- And there is different video exemple for the inference.
43
-
44
- Performance summary for the first wave of shortcut model:
45
-
46
- ## Performance Summary
47
-
48
- | Model | Optimizer | Steps | sat_mse | sat_psnr | sat_ssim | light_mae | light_precision | light_recall | light_f1 | light_iou |
49
- |----------------|-----------|--------|---------|----------|----------|-----------|-----------------|--------------|----------|-----------|
50
- | RF (Run 1) | - | 128 | 0.0952 | 28.5327 | 0.8042 | 0.0221 | 0.5482 | 0.6535 | 0.5950 | - |
51
- | RF (Run 2) | - | 128 | 0.1076 | 27.8870 | 0.8000 | 0.0221 | 0.5157 | 0.6454 | 0.5724 | - |
52
- | Baseline | Persistence| baseline| 0.2368 | 24.5138 | 0.7266 | 0.0154 | 0.6714 | 0.6665 | 0.6678 | 0.1023 |
53
- | Shortcut | Adam | 16 | 0.0981 | 28.3788 | 0.8106 | 0.0216 | 0.6339 | 0.5192 | 0.5686 | 0.0791 |
54
- | Shortcut | Adam | 64 | 0.0983 | 28.3702 | 0.8114 | 0.0207 | 0.6609 | 0.5304 | 0.5860 | 0.0791 |
55
- | Shortcut | Adam | 128 | 0.0983 | 28.3581 | 0.8112 | 0.0208 | 0.6518 | 0.5208 | 0.5769 | 0.0791 |
56
- | Shortcut | SOAP | 16 | 0.0601 | 30.5008 | 0.8663 | 0.0156 | 0.8654 | 0.6958 | 0.7710 | 0.0818 |
57
- | Shortcut | SOAP | 64 | 0.0606 | 30.4786 | 0.8661 | 0.0151 | 0.8658 | 0.6879 | 0.7663 | 0.0818 |
58
- | Shortcut | SOAP | 128 | 0.0605 | 30.4848 | 0.8660 | 0.0151 | 0.8635 | 0.6886 | 0.7656 | 0.0818 |
59
-
60
- Metrics from evaluation on 64x20 elements (satellite and lightning data).
61
-
62
-
 
1
+ FCI+radar models summery:
2
+ - radar_finetune_v2_sigma0dot02.safetensors: good model for long term, train with noise on input data (sigma: 0.02)
3
+ - radar_finetune_v3_sigma0dot2.safetensors: train with noise on input data (sigma: 0.2) but in progressive / generative mode (between t = 0./1.)