Load TorchUncertainty data in Python using dltHub

Build a TorchUncertainty-to-database or-dataframe pipeline in Python using dlt with automatic Cursor support.

In this guide, we'll set up a complete Torch Uncertainty data pipeline from API credentials to your first data load in just 10 minutes. You'll end up with a fully declarative Python pipeline based on dlt's REST API connector, like in the partial example code below:

Example code
@dlt.source
def torch_uncertainty_source(access_token=dlt.secrets.value):
    config: RESTAPIConfig = {
        "client": {
            "base_url": "https://github.com/ENSTA-U2IS-AI/torch-uncertainty/",
            "auth": {
                "type": "bearer",
                "token": access_token,
            },
        },
        "resources": [
            models,losses,metrics
            ],
    }
    [...]
    yield from rest_api_resources(config)


def get_data() -> None:
    # Connect to destination
    pipeline = dlt.pipeline(
        pipeline_name='torch_uncertainty_pipeline',
        destination='duckdb',
        dataset_name='torch_uncertainty_data', 
    )
    # Load the data
    load_info = pipeline.run(torch_uncertainty_source())
    print(load_info)

Why use dltHub Workspace with LLM Context to generate Python pipelines?

  • Accelerate pipeline development with AI-native context
  • Debug pipelines, validate schemas and data with the integrated Pipeline Dashboard
  • Build Python notebooks for end users of your data
  • Low maintenance thanks to Schema evolution with type inference, resilience and self documenting REST API connectors. A shallow learning curve makes the pipeline easy to extend by any team member
  • dlt is the tool of choice for Pythonic Iceberg Lakehouses, bringing mature data loading to pythonic Iceberg with or without catalogs

What you’ll do

We’ll show you how to generate a readable and easily maintainable Python script that fetches data from torch_uncertainty’s API and loads it into Iceberg, DataFrames, files, or a database of your choice. Here are some of the endpoints you can load:

  • Models: Deep learning models with uncertainty quantification including ResNet, SWAG, SWA, and EMA implementations
  • Losses: Uncertainty-aware loss functions like DECLoss, ELBOLoss, and Deep Evidential losses
  • Metrics: Calibration and uncertainty evaluation metrics including AURC, AUSE, and Calibration Error
  • OOD Criteria: Out-of-distribution detection methods using entropy, energy, and maximum logit approaches
  • DataModules: Dataset handlers for CIFAR, ImageNet, MUAD and other benchmark datasets with corruption support

You will then debug the Torch Uncertainty pipeline using our Pipeline Dashboard tool to ensure it is copying the data correctly, before building a Notebook to explore your data and build reports.

Setup & steps to follow

💡
Before getting started, let's make sure Cursor is set up correctly:

Now you're ready to get started!

  1. ⚙️ Set up dlt Workspace

    Install dlt with duckdb support:

    pip install dlt[workspace]

    Initialize a dlt pipeline with Torch Uncertainty support.

    dlt init dlthub:torch_uncertainty duckdb

    The init command will setup the necessary files and folders for the next step.

  2. 🤠 Start LLM-assisted coding

    Here’s a prompt to get you started:

    Prompt
    Please generate a REST API Source for Torch Uncertainty API, as specified in @torch_uncertainty-docs.yaml 
Start with endpoints models and losses and skip incremental loading for now. 
Place the code in torch_uncertainty_pipeline.py and name the pipeline torch_uncertainty_pipeline. 
If the file exists, use it as a starting point. 
Do not add or modify any other files. 
Use @dlt rest api as a tutorial. 
After adding the endpoints, allow the user to run the pipeline with python torch_uncertainty_pipeline.py and await further instructions.

  3. 🔒 Set up credentials

    The source uses API key authentication for accessing the GitHub repository and associated resources.

    To get the appropriate API keys, please visit the original source at https://github.com/ENSTA-U2IS-AI/torch-uncertainty. If you want to protect your environment secrets in a production environment, look into setting up credentials with dlt.

  4. 🏃‍♀️ Run the pipeline in the Python terminal in Cursor

    python torch_uncertainty_pipeline.py

    If your pipeline runs correctly, you’ll see something like the following:

    Pipeline torch_uncertainty load step completed in 0.26 seconds
1 load package(s) were loaded to destination duckdb and into dataset torch_uncertainty_data
The duckdb destination used duckdb:/torch_uncertainty.duckdb location to store data
Load package 1749667187.541553 is LOADED and contains no failed jobs

  5. 📈 Debug your pipeline and data with the Pipeline Dashboard

    Now that you have a running pipeline, you need to make sure it’s correct, so you do not introduce silent failures like misconfigured pagination or incremental loading errors. By launching the dlt Workspace Pipeline Dashboard, you can see various information about the pipeline to enable you to test it. Here you can see:

    • Pipeline overview: State, load metrics
    • Data’s schema: tables, columns, types, hints
    • You can query the data itself
    dlt pipeline torch_uncertainty_pipeline show --dashboard

  6. 🐍 Build a Notebook with data explorations and reports

    With the pipeline and data partially validated, you can continue with custom data explorations and reports. To get started, paste the snippet below into a new marimo Notebook and ask your LLM to go from there. Jupyter Notebooks and regular Python scripts are supported as well.

    import dlt

data = dlt.pipeline("torch_uncertainty_pipeline").dataset()
# get odel table as Pandas frame
data.odel.df().head()

Running into errors?

This is a PyTorch library for uncertainty quantification, not a traditional REST API, so endpoints are actually Python modules and classes. Some datasets have licensing restrictions and cannot be used commercially. Memory usage can be high with large datasets when using variation ratio metrics. Certain features like attention weights averaging are not yet supported.

Extra resources:

Next steps