notnot696/deep-neural-network-from-scratch

How notnot696/deep-neural-network-from-scratch works

This repository is a NumPy neural-network library for building, training, and evaluating classifiers. Its principal workflow is to load and prepare a dataset, compose a network, train it, then predict or evaluate and save results. The sampled model directly performs layer passes and uses utility losses and plotting; optimizer, scheduler, and regularizer modules provide additional capabilities, but their integration into training is not shown. Callback and logging behavior is documented and partially visible in src/utils.py.

Main entry points: model.py Neural network · data_loader.py Iris and image data · layers.py Dense and norm layers · utils.py Training callbacks

Components 18

Data preparation

Iris and image data src/data_loader.py
Encoding and splitting src/utils.py

Network runtime

Neural network src/model.py
Dense and norm layers src/layers.py
Activation functions src/activations.py
Weight initializers src/initializers.py

Training extensions

Gradient optimizers src/optimizers.py
Learning rate schedules src/schedulers.py
Weight regularization src/regularizers.py
Training callbacks src/utils.py

Evaluation and artifacts

Losses and metrics src/utils.py
Training plots src/utils.py
Experiment logging src/utils.py
Saved model weights src/model.py
JSON experiment logs src/utils.py

Other components

Model builder
Dataset providers
Local filesystem

Connections 19

Model builder → Iris and image data: requests data
Iris and image data → Dataset providers: fetches dataset
Iris and image data → Encoding and splitting: encodes and splits
Model builder → Neural network: configures and trains
Neural network → Dense and norm layers: runs forward and backward
Dense and norm layers → Activation functions: applies activation
Dense and norm layers → Weight initializers: initializes weights
Neural network → Losses and metrics: computes loss and accuracy
Neural network → Training plots: plots training history
Neural network → Saved model weights: saves or loads
Model builder → Gradient optimizers: selects optimizer
Model builder → Learning rate schedules: selects schedule
Model builder → Weight regularization: uses regularization
Model builder → Training callbacks: uses callbacks
Training callbacks → Neural network: saves checkpoint
Model builder → Experiment logging: records experiment
Experiment logging → JSON experiment logs: writes JSON
JSON experiment logs → Local filesystem: persists
Saved model weights → Local filesystem: persists

More Python diagrams: kijai/comfyui_essentials · larsvogt/knowledge-graph-building-blocks · latent-spaces/cled · leadgrowgtm/discolike-cli · lefteris12/multitask-schema-guided-dst

Drawn by GitDiagram on September 23, 2026 from the repository's file tree, README and source files. Read as Markdown ·