YAML Metadata
Warning:
empty or missing yaml metadata in repo card
(https://huggingface.co/docs/hub/model-cards#model-card-metadata)
CLIP Sparse Autoencoder Checkpoint
This model is a sparse autoencoder trained on CLIP's internal representations.
Model Details
Architecture
- Layer: 10
- 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: False
Training
- Training Images: 648254
- Learning Rate: 0.0053
- L1 Coefficient: 0.0002
- Batch Size: 4096
- Context Size: 49
Performance Metrics
Sparsity
- L0 (Active Features): 128
- Dead Features: 0
- Mean Log10 Feature Sparsity: -4.8176
- Features Below 1e-5: 27457
- Features Below 1e-6: 3098
- Mean Passes Since Fired: 67.5090
Reconstruction
- Explained Variance: 0.9193
- Explained Variance Std: 0.0396
- MSE Loss: 0.0020
- L1 Loss: 0
- Overall Loss: 0.0020
Training Details
- Training Duration: 2697 seconds
- Final Learning Rate: 0.0000
- Warm Up Steps: 500
- Gradient Clipping: 1
Additional Information
- Original Checkpoint Path: /network/scratch/p/praneet.suresh/celeba_checkpoints_2/05af194f-tinyclip_sae_16_hyperparam_sweep_lr/n_images_648338.pt
- Wandb Run: https://wandb.ai/perceptual-alignment/celeba-patches_remaining_layers/runs/3lkjalrx
- Random Seed: 42
- Downloads last month
- 4