README.md

metadata

license: other

Uploading Pretrained Checkpoints to Hugging Face

This guide explains how to upload model checkpoints to a shared Hugging Face repository using huggingface_hub.

---

1. Prerequisites

Activate your conda environment, then install the Hugging Face Hub client inside it:

conda activate <your-env>
pip install huggingface_hub

Verify it is installed correctly:

python -c "import huggingface_hub; print(huggingface_hub.__version__)"

Note on CLI command name: In huggingface_hub v1.0+, the CLI was renamed from huggingface-cli to hf. If you are on v1.x (check with the command above), use hf everywhere in this guide. If the command is still not found despite the package being installed, use the module fallback:

python -m huggingface_hub.cli.hf --version

Use python -m huggingface_hub.cli.hf as a drop-in replacement for hf anywhere below.

---

2. Create a Hugging Face Account

Go to huggingface.co and create a personal account if you do not have one. Ask the organisation owner to add you as a member of the shared organisation so you have write access to the repository.

---

3. Log in

hf auth login

You will be prompted for an access token. Generate one at: huggingface.co → Settings → Access Tokens → New token (role: Write)

Paste the token when prompted. Your credentials are stored locally and persist across sessions.

---

4. Create a Repository in the Organisation

Before uploading, you need a repository to upload to. Please use the same name as the github repo. Create one under the organisation with:

hf repos create AICM-HD/<repo-name> --repo-type model

Add --private if the repository should not be publicly visible:

hf repos create AICM-HD/<repo-name> --repo-type model --private

---

6. Checkpoint Naming Convention

Before uploading, rename checkpoints to make them unambiguous. A clear convention is:

{model}_{dataset}.pth

For example: vqgan_chestct.pth, ldm_chestct.pth, etc.

---

7. Upload Checkpoints

Replace <repo> with the actual repository name on Hugging Face.

Upload a single file:

hf upload AICM-HD/<repo> \
    /path/to/local/checkpoint.pth \
    checkpoint.pth \
    --repo-type model

The third argument is the destination filename inside the repository.

Upload all checkpoints at once from a folder:

hf upload AICM-HD/<repo> \
    /path/to/checkpoints/ \
    . \
    --repo-type model

This uploads the entire folder contents to the root of the repository.

---

8. Verify the Upload

Open the repository in a browser:

https://huggingface.co/<org>/<repo>

You should see the uploaded files listed under Files and versions. Click any file to confirm its size matches the local checkpoint.

You can also verify from the terminal:

hf repo info AICM-HD/<repo> --repo-type model

---

9. Downloading Checkpoints

Anyone with access to the repository can download checkpoints with:

pip install huggingface_hub

hf download AICM-HD/<repo> checkpoint.pth --local-dir checkpoints/

Or download everything at once:

hf download AICM-HD/<repo> --local-dir checkpoints/

If the repository is private, members must log in first (hf auth login) with their own access token.

---