hypersonic-eda 0.1.1

Generate structured EDA notebooks from datasets.

Hypersonic — EDA Notebook Generator

Build a complete Exploratory Data Analysis notebook in minutes.
Data scientists and ML folks are often evaluated on their EDA chops—but wiring up the same scaffolding each time is repetitive. Hypersonic lets you quick-build a clean, structured Jupyter notebook so you can spend your time on insights, not boilerplate.

• ⚡ Fast: point it at CSV/Parquet/SQLite and get an EDA notebook.
• 🧭 Structured: title/overview, reproducible loading, helper plots, and per-feature sections.
• 🧹 Practical: includes basic string normalization and “check/correct category typos” prompts.
• 🧰 Flexible: use it as a CLI or import as a Python library.

Name inspiration: hypersonic jets—because your EDA should take off quickly.

---

Installation

Python 3.9+ recommended.

From source (this repo)

# from the repo root
pip install -e .

This installs a console command hy and the hypersonic-eda Python package.

---

Quick Start (CLI)

Basic (CSV)

hy --input data/my_data.csv --output eda.ipynb

With a target column

hy --input data/my_data.csv --target "label" --output eda_label.ipynb

Parquet

hy --input data/my_data.parquet --output eda_parquet.ipynb

SQLite (.db), first table auto-detected

hy --input data/my_database.db --output eda_db.ipynb

SQLite with a specific table

hy --input data/my_database.db --table events --output eda_events.ipynb

Remote files (HTTP/HTTPS)

hy --input "https://example.com/data.csv" --output eda_remote.ipynb

---

Command-line Options

hy --help

• --input (required): CSV/Parquet/SQLite .db (path or URL)
• --table (optional): Table name if input is SQLite
• --target (optional): Target column name (excluded from feature loops)
• --max-cat (default: 30): Max number of categorical features to include
• --max-num (default: 30): Max number of numeric features to include
• --output (default: eda.ipynb): Output notebook path

---

What the Generated Notebook Contains

1. Title & Overview
   Source info (path/URL, table, target) and timestamp.

2. Data Loading (reproducible)
   Handles CSV/Parquet/SQLite; downloads remote files to a temp path for SQLite.

3. Dataset Snapshot
   A compact describe(include="all") view (transposed) of the dataframe.

4. Category Text Hygiene
   A cell that normalizes string columns (lowercase, trims whitespace, removes stray spaces/underscores/periods).
   A Markdown reminder to “carefully check and correct category typos.”
   Per-categorical-feature cells that print unique values and counts to help you spot issues.

5. Helper Functions
   Simple plotting utilities (plot_categorical, plot_numeric) and tabulation helpers.

6. Per-Feature Sections
   For each categorical feature: a plot cell and a value-counts table.
   For each numeric feature: a histogram and a summary-stats table (with IQR fences).

7. Notes
   Pointers and next-steps you can extend.

Tip: For very long-tail categoricals, consider grouping infrequent classes into “Other” right in the notebook.

---

Python API

Use Hypersonic (hy) programmatically if you prefer.

import pandas as pd
from hypersonic import hy as hnb

# Infer feature types from a small preview (mimics CLI behavior)
df = pd.read_csv("data/my_data.csv")
cats, nums = hnb.infer_feature_types(df, max_unique_for_categorical=40)

# Build the notebook
hnb.build_notebook(
    input_source="data/my_data.csv",  # path or URL
    output_path="eda.ipynb",
    target=None,                      # or "label"
    table=None,                       # set for SQLite
    features_categorical=cats,
    features_numeric=nums,
)

Or just call the CLI entrypoint from Python:

from hypersonic import hy

hy(["--input","data.csv","--output","eda.ipynb"])

---

Examples

Create a notebook focused on a target, with more features:

hy --input data/train.csv --target Outcome --max-cat 50 --max-num 50 --output eda_outcome.ipynb

Generate EDA for the first table in a remote SQLite DB:

hy --input "https://host/path/data.db" --output eda_db.ipynb

---

Why this design?

• One file in, one notebook out. Clear contract, low ceremony.
• Reproducibility first. The “Data Loading” cell records exactly how the data was pulled.
• No heavy dependencies. Pure pandas/matplotlib/nbformat (+ requests for remote files).

---

Limitations & Notes

• Extremely wide datasets (thousands of columns) will generate large notebooks—use --max-cat/--max-num to cap feature counts.
• The default string cleaning is intentionally simple; adapt it in the notebook if your domain needs different rules.
• Plots are basic on purpose—tune them to your style post-generation.

---

Roadmap

• Subcommands for different flows (e.g., hy eda target ...)
• Optional profiling (timings, memory)
• “Other” binning for long-tail categories
• Skew detection and auto-transform hints

---

Contributing

Issues and PRs welcome!
Please open an issue with a small sample dataset and the expected behavior.

---

License

MIT © Krish Ambady