nGeneDL: Deep Learning Engine

nGeneDL is a purpose-built neural-network runtime engineered for PyScript. Its compact WebAssembly bundle (< 120 KB) circumvents the size and binary-compatibility limits that exclude heavyweight frameworks from the Pyodide whitelist.

Compatibility highlights

Library	Compressed bundle size	PyScript support
nGeneDL	< 120 KB	Fully supported
PyTorch	> 100 MB	Unsupported — exceeds whitelist limits
TensorFlow	> 70 MB	Unsupported — exceeds whitelist limits

Edition guide

nGeneDL_Prototype – minimal reference implementation for educational inspection.
nGeneDL_LinearAlgebra – explicit matrix operations to facilitate custom layer development.
nGeneDL_LinearAlgebra_SGD – default edition; fused SGD kernels maximise browser-side throughput while preserving accuracy.

Roadmap note

Designed for low-latency, in-browser inference, nGeneDL is oriented toward clinical workloads such as ventilator pressure–flow–volume waveform classification at the point of care. An optional Flask gateway, scheduled for a future release, will extend training to CUDA hardware while preserving browser-based inference. This enhancement is secondary to the core in-browser focus.

Practical `input_nodes` choices

Scenario	How the features are formed	Typical `input_nodes` value
Raw, flattened MNIST – no CNN, no pooling	28 × 28 grayscale pixels flattened to 1-D	784
Half-resolution flattening	Down-sample image to 14 × 14 (e.g., average pooling) then flatten	196
Quarter-resolution flattening	Down-sample to 7 × 7 → very light model; good for demos	49
Single-layer CNN Prototype	After 3 × 3 conv + max-pool to 14 × 14 (1 filter) then flatten	14 × 14 = 196
Multiple-filter CNN	`num_filters × (spatial_h × spatial_w)` e.g., 3 filters pooled to 14 × 14	3 × 196 = 588
Feature extractor (PCA / autoencoder)	Dimensionality-reduced latent vector	Any number you choose (commonly 32 – 256)
Hand-engineered statistics	e.g., intensity histogram (32 bins) + 4 Zernike moments	36

Rules of thumb

784: best for textbook-style comparisons and linear-algebra editions.
196: quick browser demos without touching the CNN pipeline.
When a CNN pre-processor is enabled, let the code compute input_nodes automatically (leave the HTML field unchanged or set to –1).
49: ultra-small demo; still ~90 % accuracy on MNIST.
For a custom feature extractor, set input_nodes to the latent dimension (often 64 or 128).

As PyScript cannot load heavyweight frameworks such as PyTorch or TensorFlow in-browser, a custom lightweight engine enables direct deep learning execution within the browser.

Control Panel

Script-mode Testcases (Working)

Major:

Minor:

Major: Testcases_nGeneDL_LinearAlgebra Minor: test_early_stopping

Traditional Dashboard with Customizable Deep Learning Parameters

Input Nodes:

Output Nodes:

Hidden Nodes:

Learning Rate: 0.01

Deep Learning Engine:

Optimizer:

CNN Prep:

Activation Function:

temporarily disabled

Use Glorot

Visualize

Debug Backprop

Task Classification

MNIST Handwriting Recognition Pipeline

Train Dataset:

Test Dataset:

▼▼▼ Development Section Below ▼▼▼

PyScript is Loading…

Wait for the “Log Console” to initialize.

nGeneDL (ver 0.6.9-dev): Custom-Built Lightweight Open-Source Deep Learning Engine for In-Browser Use