Note the safetensors/param-count meaning
Browse files
README.md
CHANGED
|
@@ -54,6 +54,9 @@ any direction between modalities.
|
|
| 54 |
64} dimensions with renormalization.
|
| 55 |
- **Compact distribution.** This repository ships the connector and normalization
|
| 56 |
statistics (~60 MB); the frozen towers are downloaded from their original repositories.
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
This is a **research preview**, currently at **v0.3**: the v0.2 contrastive stage (484K
|
| 59 |
pairs) followed by a connector-only in-domain fine-tune on the AudioCaps train split.
|
|
|
|
| 54 |
64} dimensions with renormalization.
|
| 55 |
- **Compact distribution.** This repository ships the connector and normalization
|
| 56 |
statistics (~60 MB); the frozen towers are downloaded from their original repositories.
|
| 57 |
+
The parameter count shown for this repository (16.4M) is the trained connector —
|
| 58 |
+
`model.safetensors` and the `.pt` checkpoint contain the same weights; `inference.py`
|
| 59 |
+
loads the `.pt`.
|
| 60 |
|
| 61 |
This is a **research preview**, currently at **v0.3**: the v0.2 contrastive stage (484K
|
| 62 |
pairs) followed by a connector-only in-domain fine-tune on the AudioCaps train split.
|