From-scratch training loop to embedded inference
Small neural nets, built close to the metal.
A compact MLP is trained in PyTorch with manual forward pass, BCE loss, backprop, and gradient descent. The learned weights are exported for a C++ inference path, with int8 quantization as the edge deployment target.
7-12-8-1MLP architecture
1200manual epochs
int8deployment target
$ ready model: Titanic survival MLP input features: Pclass, Sex, Age, SibSp, Parch, Fare, Embarked export: weights/*.txt + biases/*.txt
Pipeline
01Normalize
Clean tabular inputs and freeze train-set mean/std for test parity.
02Train
Manual tensor math, sigmoid output, BCE loss, and gradient updates.
03Export
Weights and biases are written into plain text files for C++ loading.
04Quantize
Inference path moves toward compact int8 arithmetic for edge devices.
The demo below simulates the training trace from the project structure; the repository contains the actual PyTorch and C++ source.