Downloads · 30 days
16
31% of all-time downloads
celeba-sae-top_k-64-patches_only-layer_4-hook_resid_post
Prisma-Multimodal/celeba-sae-top_k-64-patches_only-layer_4-hook_resid_post
celeba-sae-top_k-64-patches_only-layer_4-hook_resid_post is a machine learning model from Prisma-Multimodal. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This model is a sparse autoencoder trained on CLIP's internal representations.
All-time downloads
51
Public
Repo size
302 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt302 MB · 100%
At a glance
- Access
- Public
- Created
- Jan 12, 2025
- Updated
- Jan 12, 2025
- SHA
- 360b5b7d
Model card
From the Hugging Face model README
CLIP Sparse Autoencoder Checkpoint
This model is a sparse autoencoder trained on CLIP's internal representations.
Model Details
Architecture
- Layer: 4
- 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: 324085
- Learning Rate: 0.0005
- L1 Coefficient: 0.0002
- Batch Size: 4096
- Context Size: 49
Performance Metrics
Sparsity
- L0 (Active Features): 256.0000
- Dead Features: 0
- Mean Log10 Feature Sparsity: -4.2232
- Features Below 1e-5: 575.0000
- Features Below 1e-6: 0.0000
- Mean Passes Since Fired: 7.0746
Reconstruction
- Explained Variance: 0.9047
- Explained Variance Std: 0.0229
- MSE Loss: 0.0007
- L1 Loss: 0
- Overall Loss: 0.0007
Training Details
- Training Duration: 1150 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/d1591a54-tinyclip_sae_16_hyperparam_sweep_lr/n_images_324169.pt
- Wandb Run: https://wandb.ai/perceptual-alignment/celeba-sweep-topk-patches_all_layers/runs/wgmp7e6o
- Random Seed: 42
- Created
- Jan 12, 2025
- Updated
- Jan 12, 2025