CLIP Sparse Autoencoder Checkpoint
This model is a sparse autoencoder trained on CLIP's internal representations.
Model Details
Architecture
- Layer: 11
- Layer Type: hook_resid_post
- Model: open-clip:laion/CLIP-ViT-B-32-DataComp.XL-s13B-b90K
- Dictionary Size: 49152
- Input Dimension: 768
- Expansion Factor: 64
- CLS Token Only: True
Training
- Training Images: 110178304
- Learning Rate: 0.0002
- L1 Coefficient: 0.3000
- Batch Size: 4096
- Context Size: 1
Performance Metrics
Sparsity
- L0 (Active Features): 64
- Dead Features: 0
- Mean Log10 Feature Sparsity: -3.4080
- Features Below 1e-5: 10
- Features Below 1e-6: 0
- Mean Passes Since Fired: 13.0446
Reconstruction
- Explained Variance: 0.8423
- Explained Variance Std: 0.0443
- MSE Loss: 0.0025
- L1 Loss: 0
- Overall Loss: 0.0025
Training Details
- Training Duration: 17866.3376 seconds
- Final Learning Rate: 0.0002
- Warm Up Steps: 200
- Gradient Clipping: 1
Additional Information
- Weights & Biases Run: https://wandb.ai/perceptual-alignment/clip/runs/b5q0wr11
- Original Checkpoint Path: /network/scratch/s/sonia.joseph/checkpoints/clip-b
- Random Seed: 42