chebai needs in-code downloadable models

[sfluegel05]

Currently, we provide datasets and models for specific papers via kaggle / zenodo. This is great for people that want to reproduce experiments. However, for users that just need the results for some specific molecules (and are not interested in training their own models), we should provide a model by default.

The final result should look something like this:

python -m chebai predict --smiles "Cn1cnc2n(C)c(=O)n(C)c(=O)c12" --output my_output.json

which returns a file with prediction results. (the command is an illustration, the actual implementation might be a bit different)

The key component here is the self-downloading model. Internally, the predict function should download a model (or find a previously downloaded model) and use it.

Find out how to do this

• Where can we host our models?
• How can we integrate a hosted model into our cli?
• What options do we have to switch between different models? (this would be an advanced feature and not the "default experience")

[aditya0by0]

Where to Host Models

Platform	Pros	Cons
Hugging Face Hub	Free, versioned, CLI/API support, model card UI, widely used	Public by default (can be gated)
Zenodo	DOI support, long-term archival storage	Limited integration, no versioning
Kaggle Datasets	Good for reproducibility	No real API for self-downloading
AWS S3 / GCS	Scalable, fast, private/public options	Requires auth & infra management
GitHub Releases	Easy to maintain alongside code	Size limits (max 2 GB per file)

Hugging Face Hub seem a better option

• Model cards
• Versioning
• transformers & huggingface_hub CLI for downloads
• Private/public access
• Can serve PyTorch, ONNX, or even pickled models

Please check -- https://huggingface.co/docs/hub and https://huggingface.co/docs/hub/models

[aditya0by0]

To switch between different models and version we can have model registry in our code, which map to specific downloadable link.

MODEL_REGISTRY = {
    "chebaiv1": "your-org/chebai-v1",
    "gnncv2": "your-org/chebai-graph-v2",
}

and user have to provide model-name when call predict through cli

python -m chebai predict --smiles "..." --model-name "gnn-v2"

[aditya0by0]

One concern I have is that, in order to run the model prediction, users need to set up their environment to match the dependencies specified in our setup.py or pyproject.toml. This can be problematic—especially for our graph-based models—since certain dependencies (like torch-geometric) need to be installed in a way that is compatible with the user's existing PyTorch version.

Is it possible to automate this environment setup? Or is it better to let users handle it manually? What's the recommended approach to ensure smooth installation and compatibility, especially when dealing with version-sensitive dependencies?