Name the weights after their experiments (Pointcept config naming)
Browse files
README.md
CHANGED
|
@@ -22,11 +22,12 @@ at a photo's native resolution, up to 100 MP.
|
|
| 22 |
|
| 23 |
| File | Model | Training data | Size | sha256 |
|
| 24 |
|---|---|---|---|---|
|
| 25 |
-
| `l-dpt-l-zs.pth` | L-DPT-L zero-shot | 8 synthetic / multi-view sets (below) | 1.08 GB | `b4e7e95301fda7557760f787e7992f4ccca3c03cb70e46eadeb48e227b1b31ca` |
|
| 26 |
-
| `l-dpt-l-mix.pth` | L-DPT-L mixed | the same + NYUv2, KITTI, DIODE train | 1.08 GB | `4852b823c68fea30b3360a249e71daf72a9c6c2c74f747602a117b42f35d36fa` |
|
| 27 |
|
| 28 |
-
Each file
|
| 29 |
-
|
|
|
|
| 30 |
License. Get the ViT-7B/16 (LVD-1689M) weights from Meta: the original file
|
| 31 |
`dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth` of the DINOv3 release, or the gated
|
| 32 |
Hugging Face repo
|
|
@@ -49,7 +50,7 @@ With the `ldpt` package of the L-DPT code release:
|
|
| 49 |
from huggingface_hub import hf_hub_download
|
| 50 |
from ldpt.models.l_dpt import api
|
| 51 |
|
| 52 |
-
decoder = hf_hub_download("pointcept-bot/L-DPT", "l-dpt-l-zs.pth")
|
| 53 |
dinov3 = "dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth" # Meta's file (or the HF folder)
|
| 54 |
model = api.load(decoder, encoder_weights=dinov3) # cuda if available
|
| 55 |
depth = api.predict(model, image) # HxWx3 uint8 RGB -> HxW float32
|
|
@@ -80,8 +81,8 @@ DIODE is the combined indoor + outdoor validation set.
|
|
| 80 |
|
| 81 |
| Model | NYUv2 | KITTI | ETH3D | DIODE |
|
| 82 |
|---|---|---|---|---|
|
| 83 |
-
| `l-dpt-l-zs.pth` | 3.55 / 98.30 | 5.98 / 97.85 | 3.75 / 99.01 | 19.34 / 85.65 |
|
| 84 |
-
| `l-dpt-l-mix.pth` | 2.89 / 99.08 | 5.39 / 98.38 | 3.57 / 98.94 | 14.98 / 88.17 |
|
| 85 |
|
| 86 |
NYUv2, KITTI and DIODE are zero-shot for the zero-shot model and in-domain for the mixed
|
| 87 |
one. On photos in the wild, the zero-shot model separates thin foreground (foliage,
|
|
|
|
| 22 |
|
| 23 |
| File | Model | Training data | Size | sha256 |
|
| 24 |
|---|---|---|---|---|
|
| 25 |
+
| `depth-l-dpt-v1m1-0-l-zs.pth` | L-DPT-L zero-shot | 8 synthetic / multi-view sets (below) | 1.08 GB | `b4e7e95301fda7557760f787e7992f4ccca3c03cb70e46eadeb48e227b1b31ca` |
|
| 26 |
+
| `depth-l-dpt-v1m1-1-l-mix.pth` | L-DPT-L mixed | the same + NYUv2, KITTI, DIODE train | 1.08 GB | `4852b823c68fea30b3360a249e71daf72a9c6c2c74f747602a117b42f35d36fa` |
|
| 27 |
|
| 28 |
+
Each file is named after the experiment (config) that produced it
|
| 29 |
+
(`configs/hybrid/<name>.py` of the code release) and holds the decoder only (fp32
|
| 30 |
+
weights, 271M parameters) and its model configuration. **The DINOv3 encoder is not included**: it is Meta's, under the DINOv3
|
| 31 |
License. Get the ViT-7B/16 (LVD-1689M) weights from Meta: the original file
|
| 32 |
`dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth` of the DINOv3 release, or the gated
|
| 33 |
Hugging Face repo
|
|
|
|
| 50 |
from huggingface_hub import hf_hub_download
|
| 51 |
from ldpt.models.l_dpt import api
|
| 52 |
|
| 53 |
+
decoder = hf_hub_download("pointcept-bot/L-DPT", "depth-l-dpt-v1m1-0-l-zs.pth")
|
| 54 |
dinov3 = "dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth" # Meta's file (or the HF folder)
|
| 55 |
model = api.load(decoder, encoder_weights=dinov3) # cuda if available
|
| 56 |
depth = api.predict(model, image) # HxWx3 uint8 RGB -> HxW float32
|
|
|
|
| 81 |
|
| 82 |
| Model | NYUv2 | KITTI | ETH3D | DIODE |
|
| 83 |
|---|---|---|---|---|
|
| 84 |
+
| `depth-l-dpt-v1m1-0-l-zs.pth` | 3.55 / 98.30 | 5.98 / 97.85 | 3.75 / 99.01 | 19.34 / 85.65 |
|
| 85 |
+
| `depth-l-dpt-v1m1-1-l-mix.pth` | 2.89 / 99.08 | 5.39 / 98.38 | 3.57 / 98.94 | 14.98 / 88.17 |
|
| 86 |
|
| 87 |
NYUv2, KITTI and DIODE are zero-shot for the zero-shot model and in-domain for the mixed
|
| 88 |
one. On photos in the wild, the zero-shot model separates thin foreground (foliage,
|
l-dpt-l-zs.pth → depth-l-dpt-v1m1-0-l-zs.pth
RENAMED
|
File without changes
|
l-dpt-l-mix.pth → depth-l-dpt-v1m1-1-l-mix.pth
RENAMED
|
File without changes
|