lewisnjue/nnetflow

A minimal, NumPy-based deep learning library with a clean, educational reverse-mode autodiff engine

What's novel

A minimal, NumPy-based deep learning library with a clean, educational reverse-mode autodiff engine

Code Analysis

10 files read · 3 rounds

A NumPy-based neural network library implementing reverse-mode automatic differentiation from scratch, with layers (Linear, Conv1d/2d, BatchNorm, LayerNorm, MultiHeadAttention), optimizers (SGD, Adam, Adagrad, RMSProp), and losses for educational purposes.

Strengths

Genuine from-scratch autodiff engine with correct gradient rules, broadcasting-aware unbroadcast, and einsum-based convolution. Tests validate gradients against PyTorch as a reference, which is a strong correctness strategy.

Weaknesses

RMSProp has a crash bug (undefined variable in __init__), Conv backward uses slow Python loops for scatter-add, pyproject.toml lists torch as a runtime dependency (not just test), and setup.py/pyproject.toml dependencies are inconsistent. README claims GPU/CuPy support that isn't visible in the code read.

Score Breakdown

Innovation

4 (25%)

Craft

82 (35%)

Traction

9 (15%)

Scope

72 (25%)

Signal breakdown

Innovation

Not Fork+1

Code Novelty+1

Concept Novelty+1

Craft

Ci+5

Tests+8

Polish+3

Releases+5

Has License+5

Code Quality+19

Readme Quality+15

Recent Activity+7

Structure Quality+5

Commit Consistency+5

Has Dependency Mgmt+5

Traction

Forks+0

Stars+6

Hn Points+0

Watchers+3

Early Traction+0

Devto Reactions+0

Community Contribs+0

Scope

Commits+8

Languages+3

Subsystems+13

Bloat Penalty+0

Completeness+7

Contributors+5

Authored Files+12

Readme Code Match+3

Architecture Depth+5

Implementation Depth+8

Evidence

Commits

145

Contributors

Files

Active weeks

TestsCI/CDREADMELicenseContributing

Repository

Language

Python

Stars

Forks

License

MIT