Upload README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: pytorch
|
| 3 |
+
license: other
|
| 4 |
+
license_name: per-asset-terms
|
| 5 |
+
tags:
|
| 6 |
+
- scientific-machine-learning
|
| 7 |
+
- model-editing
|
| 8 |
+
- safetensors
|
| 9 |
+
- depth-estimation
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# HypoLoop
|
| 13 |
+
|
| 14 |
+
Native model weights and registered feedback data from **HypoLoop: Revising Weight-Editing Plans from Scientific Feedback**. The implementation, preprint, and loading instructions are available on GitHub. [HypoLoop repository](https://github.com/limengran98/HypoLoop) 路 [Loading guide](https://github.com/limengran98/HypoLoop/blob/main/docs/HUGGING_FACE.md)
|
| 15 |
+
|
| 16 |
+
| Paper task | Data | Model weights |
|
| 17 |
+
| :--- | :--- | :--- |
|
| 18 |
+
| PHY01 路 Burgers | Registered field histories, spatial grid, feedback and held-out targets | FNO1d base and final fixed- and dynamic-location checkpoints for both services |
|
| 19 |
+
| MATH01 路 Constrained QP | Registered right-hand sides, objective and constraint matrices, feedback and held-out solutions | DC3 base and final fixed- and dynamic-location checkpoints for both services |
|
| 20 |
+
| DEP02 路 DIODE | 24 registered RGB images, scene IDs, sampled depth targets and grid coordinates | Fixed-plan and revised-plan Claude head replacements for Depth Anything V2 Metric VKITTI ViT-S |
|
| 21 |
+
| DEP01 路 Make3D | [Original Make3D dataset](https://cs.stanford.edu/people/asaxena/make3d/) | DeepSeek HL-Global head replacements for the same Depth Anything V2 base |
|
| 22 |
+
|
| 23 |
+
[manifest.json](manifest.json) lists every checkpoint, its architecture, and the corresponding data files. [checksums.sha256](checksums.sha256) records the upload contents. [LICENSES.md](LICENSES.md) lists source attribution and terms by asset.
|
| 24 |
+
|
| 25 |
+
The registered order is initial feedback `F0`, subsequent feedback `B1` and `B2`, then held-out evaluation `E`. Use `E` for evaluation. Each target archive aligns `ids` with `y`. DIODE uses the 41脳61 sampled grid in [grid.json](data/diode/grid.json), with `NaN` for invalid depth. The 12 reference images are from the original Make3D training collection.
|
| 26 |
+
|
| 27 |
+
The two depth cases use named-tensor replacements. Load the original VKITTI ViT-S metric checkpoint before applying a replacement. [Depth Anything V2 upstream model](https://huggingface.co/depth-anything/Depth-Anything-V2-Metric-VKITTI-Small)
|