CLIP Microscope ImageNet Dataset
This dataset contains the top activating ImageNet images for each neuron in OpenAI's CLIP RN50x4 model, designed for use with neural network interpretability tools.
Dataset Structure
clip-microscope-imagenet/
âââ neurons/ # Top activating images per neuron
â âââ neuron_0000/ # Images for neuron 0
â â âââ train_rank_00_act_0.8234_idx_12345.jpg
â â âââ val_rank_01_act_0.7891_idx_67890.jpg
â â âââ ...
â âââ neuron_0001/ # Images for neuron 1
â âââ ...
âââ lucid/ # Lucid-generated feature visualizations
â âââ neuron_0000_lucid.png
â âââ neuron_0001_lucid.png
â âââ ...
âââ metadata/ # Dataset metadata and statistics
âââ neuron_metadata.json # Detailed neuron information
âââ dataset_summary.json # High-level statistics
âââ neuron_summary.csv # Tabular summary
Usage
This dataset is designed to work with the CLIP Microscope web application. You can access images directly via URLs:
# Base URL
BASE_URL = "https://huggingface.co/datasets/your-username/clip-microscope-imagenet/resolve/main"
Access neuron images
neuron_image_url = f"{BASE_URL}/neurons/neuron_0123/train_rank_00_act_0.8234_idx_12345.jpg"
Access lucid images
lucid_image_url = f"{BASE_URL}/lucid/neuron_0123_lucid.png"
Model Information
- Model: OpenAI CLIP RN50x4
- Layer: Image encoder blocks
- Total Neurons: 2560
- Images per Neuron: Up to 100 (top activating)
- Source Dataset: ImageNet (train and validation splits)
File Naming Convention
Neuron Images
{split}_rank_{rank:02d}_act_{activation:.4f}_idx_{original_index}.{ext}
split: "train" or "val"
rank: Ranking among top activations (00 = highest)
activation: Neuron activation value
original_index: Index in the original ImageNet dataset
Lucid Images
neuron_{neuron_id:04d}_lucid.png
Citation
If you use this dataset, please cite:
- The original CLIP paper: Radford et al., "Learning Transferable Visual Representations"
- ImageNet: Deng et al., "ImageNet: A Large-Scale Hierarchical Image Database"
License
This dataset follows the ImageNet license terms. Please ensure compliance with ImageNet's usage guidelines.