Commit
•
5fc8afd
1
Parent(s):
f04fb80
Update README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# CLIP Sparse Autoencoder Checkpoint
|
| 2 |
+
|
| 3 |
+
This model is a sparse autoencoder trained on CLIP's internal representations.
|
| 4 |
+
|
| 5 |
+
## Model Details
|
| 6 |
+
|
| 7 |
+
### Architecture
|
| 8 |
+
- **Layer**: 10
|
| 9 |
+
- **Layer Type**: hook_resid_post
|
| 10 |
+
- **Model**: open-clip:laion/CLIP-ViT-B-32-DataComp.XL-s13B-b90K
|
| 11 |
+
- **Dictionary Size**: 49152
|
| 12 |
+
- **Input Dimension**: 768
|
| 13 |
+
- **Expansion Factor**: 64
|
| 14 |
+
- **CLS Token Only**: True
|
| 15 |
+
|
| 16 |
+
### Training
|
| 17 |
+
- **Training Images**: 114683904
|
| 18 |
+
- **Learning Rate**: 0.0002
|
| 19 |
+
- **L1 Coefficient**: 0.3000
|
| 20 |
+
- **Batch Size**: 4096
|
| 21 |
+
- **Context Size**: 1
|
| 22 |
+
|
| 23 |
+
## Performance Metrics
|
| 24 |
+
|
| 25 |
+
### Sparsity
|
| 26 |
+
- **L0 (Active Features)**: 64
|
| 27 |
+
- **Dead Features**: 0
|
| 28 |
+
- **Mean Log10 Feature Sparsity**: -3.5579
|
| 29 |
+
- **Features Below 1e-5**: 83
|
| 30 |
+
- **Features Below 1e-6**: 2
|
| 31 |
+
- **Mean Passes Since Fired**: 22.8832
|
| 32 |
+
|
| 33 |
+
### Reconstruction
|
| 34 |
+
- **Explained Variance**: 0.8698
|
| 35 |
+
- **Explained Variance Std**: 0.0456
|
| 36 |
+
- **MSE Loss**: 0.0016
|
| 37 |
+
- **L1 Loss**: 0
|
| 38 |
+
- **Overall Loss**: 0.0016
|
| 39 |
+
|
| 40 |
+
## Training Details
|
| 41 |
+
- **Training Duration**: 17901.0753 seconds
|
| 42 |
+
- **Final Learning Rate**: 0.0002
|
| 43 |
+
- **Warm Up Steps**: 200
|
| 44 |
+
- **Gradient Clipping**: 1
|
| 45 |
+
|
| 46 |
+
## Additional Information
|
| 47 |
+
- **Weights & Biases Run**: https://wandb.ai/perceptual-alignment/clip/runs/10hacih1
|
| 48 |
+
- **Original Checkpoint Path**: /network/scratch/s/sonia.joseph/checkpoints/clip-b
|
| 49 |
+
- **Random Seed**: 42
|