NumPy with ambition: a deep learning framework from 2014

What it does DeepPy is a deep learning framework written in Python, built directly on NumPy with optional CUDA acceleration via its companion library, cudarray. It provides the standard toolkit: neural network layers, optimization routines, and GPU-backed array operations. The README is terse — two sentences and a link to a “preliminary website” — so the exact scope of built-in architectures and conveniences is unclear without digging into the source.

The interesting bit This is a fossil from the pre-framework era, when “Pythonic” and “NumPy-based” were genuine differentiators rather than table stakes. The author, Anders Boesen Lindbo Larsen, also built cudarray to make GPU acceleration accessible without leaving the NumPy API — a design philosophy that later frameworks would adopt, then make invisible.

Key highlights

• Pure Python API with NumPy semantics
• Optional CUDA acceleration through companion project cudarray
• 1,374 stars suggests it found an audience in the 2014–2016 window
• “Preliminary website” implies the project never reached a settled state
• No dependencies or installation details listed in README

Caveats

• README offers almost no detail on capabilities, API, or current maintenance status
• “Preliminary website” link may be stale; the project appears dormant
• Modern frameworks have absorbed and surpassed this niche entirely

Verdict Worth a look for historians of deep learning tooling or anyone studying how NumPy-centric GPU acceleration was hacked together before PyTorch. For production or learning modern deep learning, this is a curiosity, not a contender.