Manual PyTorch training to a C++ tensor loader
Trace the network, weight by weight.
A compact MLP is trained with an explicit forward pass, BCE loss, backpropagation, and gradient descent. Its learned tensors are exported for a C++ loader; the full C++ forward pass and int8 quantization remain deployment targets.
Pipeline
Clean tabular inputs and freeze train-set mean/std for test parity.
Manual tensor math, sigmoid output, BCE loss, and gradient updates.
Weights and biases are written into plain text files for C++ loading.
The C++ path currently reads the first exported weight matrix.
This trace visualizes the implemented training and export stages. C++ inference parity and int8 arithmetic are explicitly marked as work in progress.
Illustrative training curves
These curves are generated from example formulas, not recorded training runs. The displayed loss and accuracy are illustrative, not evaluation results.