FM_weights

Environment setup

conda create --name FM_weights python==3.9
conda activate FM_weights
pip install -r requirements.txt

Generate/Load Embeddings

Generate Embeddings

Get TieredImageNet

• tieredImageNet (courtesy of Kwonjoon Lee)

Local Data Directory

Download and extract data into data_root like below. For exmaple: data_root = /Users/zyxu/Documents/py/datasets

tiered_imagenet
├── tiered-imagenet-kwon
│   ├── test_images.npz
│   ├── test_labels.pkl
│   ├── train_images.npz
│   ├── train_labels.pkl
│   ├── val_images.npz
│   └── val_labels.pkl

Generate ${h_{c,t}}$

Modify data_root accordingly

python generate_pos_ctx.py --data_root /Users/zyxu/Documents/py/datasets --output_path output

Load Embeddings

Alternatively, you can download embeddings into output/ from here: 4sets_intermediate_outputs.pth

Reproducibility

After getting the embeddings output/4sets_intermediate_outputs.pth. notebook/uncover/vis.ipynb contains visulizations and notebook/uncover/rank_compute.ipynb contains rank computation.