Upload folder using huggingface_hub
Browse files
README.md
CHANGED
|
@@ -2,30 +2,12 @@
|
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
|
| 5 |
-
### Model Card
|
| 6 |
-
|
| 7 |
-
| Property | Value |
|
| 8 |
-
| ----------------------------- | --------------------------------- |
|
| 9 |
-
| **Model Type** | Vision Transformer (ViT) |
|
| 10 |
-
| **Architecture** | HEVC-Style Vision Transformer |
|
| 11 |
-
| **Hidden Size** | 1024 |
|
| 12 |
-
| **Intermediate Size** | 4096 |
|
| 13 |
-
| **Number of Layers** | 24 |
|
| 14 |
-
| **Number of Attention Heads** | 16 |
|
| 15 |
-
| **Patch Size** | 14 |
|
| 16 |
-
| **Image Resolution** | 448×448 (pre-trained) |
|
| 17 |
-
| **Video Resolution** | 224×224 with 256 tokens per frame |
|
| 18 |
-
| **Positional Encoding** | 3D RoPE (4:6:6 split for T:H:W) |
|
| 19 |
-
| **Normalization** | Layer Normalization |
|
| 20 |
-
| **Activation Function** | GELU |
|
| 21 |
-
| **License** | Apache 2.0 |
|
| 22 |
|
| 23 |
### Key Features
|
| 24 |
|
| 25 |
- **Codec-Style Patch Selection**: Instead of sampling sparse frames densely (all patches from few frames), OneVision Encoder samples dense frames sparsely (important patches from many frames).
|
| 26 |
- **3D Rotary Position Embedding**: Uses a 4:6:6 split for temporal, height, and width dimensions to capture spatiotemporal relationships.
|
| 27 |
|
| 28 |
-
### Intended Use
|
| 29 |
|
| 30 |
#### Downstream Tasks
|
| 31 |
|
|
@@ -136,3 +118,21 @@ Training on a mixed dataset of 740K samples from LLaVA-OneVision and 800K sample
|
|
| 136 |
<img alt="LMM Probe Results" src="https://raw.githubusercontent.com/anxiangsir/asset/main/OneVision/probe_lmm_github_light.png" width="800" style="max-width: 100%;">
|
| 137 |
</picture>
|
| 138 |
</p>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
### Key Features
|
| 7 |
|
| 8 |
- **Codec-Style Patch Selection**: Instead of sampling sparse frames densely (all patches from few frames), OneVision Encoder samples dense frames sparsely (important patches from many frames).
|
| 9 |
- **3D Rotary Position Embedding**: Uses a 4:6:6 split for temporal, height, and width dimensions to capture spatiotemporal relationships.
|
| 10 |
|
|
|
|
| 11 |
|
| 12 |
#### Downstream Tasks
|
| 13 |
|
|
|
|
| 118 |
<img alt="LMM Probe Results" src="https://raw.githubusercontent.com/anxiangsir/asset/main/OneVision/probe_lmm_github_light.png" width="800" style="max-width: 100%;">
|
| 119 |
</picture>
|
| 120 |
</p>
|
| 121 |
+
|
| 122 |
+
### Model Card
|
| 123 |
+
|
| 124 |
+
| Property | Value |
|
| 125 |
+
| ----------------------------- | --------------------------------- |
|
| 126 |
+
| **Model Type** | Vision Transformer (ViT) |
|
| 127 |
+
| **Architecture** | HEVC-Style Vision Transformer |
|
| 128 |
+
| **Hidden Size** | 1024 |
|
| 129 |
+
| **Intermediate Size** | 4096 |
|
| 130 |
+
| **Number of Layers** | 24 |
|
| 131 |
+
| **Number of Attention Heads** | 16 |
|
| 132 |
+
| **Patch Size** | 14 |
|
| 133 |
+
| **Image Resolution** | 448×448 (pre-trained) |
|
| 134 |
+
| **Video Resolution** | 224×224 with 256 tokens per frame |
|
| 135 |
+
| **Positional Encoding** | 3D RoPE (4:6:6 split for T:H:W) |
|
| 136 |
+
| **Normalization** | Layer Normalization |
|
| 137 |
+
| **Activation Function** | GELU |
|
| 138 |
+
| **License** | Apache 2.0 |
|