code/upload_to_hf.py

import os
import json
from huggingface_hub import HfApi

def get_param_count():
    # The default log path from train_cifar10.sh and train.py
    log_file = "./runs/cifar10_rtm/cifar10/log.jsonl"
    if not os.path.exists(log_file):
        return "unknown"
    with open(log_file, "r") as f:
        for line in f:
            try:
                data = json.loads(line)
                if "message" in data and "Number of parameters in IMLE:" in data["message"]:
                    parts = data["message"].split("Number of parameters in IMLE:")
                    if len(parts) > 1:
                        params_str = parts[1].strip()
                        return params_str
            except:
                pass
    return "unknown"

def main():
    api = HfApi()
    
    # 1. Get parameter count and format repo name
    param_count = get_param_count()
    repo_id = f"JerMa88/rtm-latent-refinement-cifar10-{param_count}params"
    print(f"Creating repository: {repo_id}")
    
    # Create the repo (it will use the implicitly logged-in user)
    try:
        api.create_repo(repo_id, exist_ok=True)
    except Exception as e:
        print(f"Error creating repo: {e}")
        
    # Upload training checkpoints and logs
    run_dir = "./runs/cifar10_rtm/cifar10"
    if os.path.exists(run_dir):
        print(f"Uploading run directory {run_dir}...")
        api.upload_folder(
            folder_path=run_dir,
            repo_id=repo_id,
            path_in_repo="runs/cifar10",
        )
    else:
        print(f"Run directory {run_dir} not found! Check if training finished successfully.")

    # Upload current repository code
    print("Uploading repository code...")
    api.upload_folder(
        folder_path=".",
        repo_id=repo_id,
        path_in_repo="code",
        ignore_patterns=["runs/*", "datasets/*", "venv/*", ".git/*"]
    )
    
    # Upload README to Hugging Face
    readme_content = f"""
# RTM Latent Refinement (Fast Variant)

This model was trained with a reduced architecture ({param_count} parameters) compared to the original paper's baseline to speed up training.

- **Dataset**: CIFAR-10
- **W&B Logs**: Check the W&B project `rtm-latent-refinement` under the account `JerMa88`.
- **Source Code**: Included in the `code/` directory.
"""
    with open("HF_README.md", "w") as f:
        f.write(readme_content)
    
    api.upload_file(
        path_or_fileobj="HF_README.md",
        path_in_repo="README.md",
        repo_id=repo_id,
    )
    
    print("Upload complete! You can view the model on Hugging Face.")

if __name__ == "__main__":
    main()