How wiegerw/nerva-torch works
nerva-torch is a PyTorch-backed, manually differentiated MLP library. Its main workflow is to load minibatches, run feedforward and loss-gradient calculations, backpropagate through layers, and update parameters using a scheduled learning rate; training can also report evaluation statistics. The graph includes configurable layers, activations, losses, optimizers, initialization, matrix primitives, and the experiment CLI. Edges reflect README statements or shown source references; unsampled implementation details are not inferred.
Main entry points: training.py Training loop · layers.py Network layers · loss_functions.py Loss functions · datasets.py Batch loader
Components 17
Training workflow
- MLP CLI
tools/mlp.py - Training loop
src/nerva_torch/training.py - Learning rates
src/nerva_torch/learning_rate.py - Training metrics
src/nerva_torch/training.py
MLP model
- MLP
src/nerva_torch/multilayer_perceptron.py - Network layers
src/nerva_torch/layers.py - Optimizers
src/nerva_torch/optimizers.py - Weight initialization
src/nerva_torch/weight_initializers.py
Math components
- Activations
src/nerva_torch/activation_functions.py - Softmax functions
src/nerva_torch/softmax_functions.py - Loss functions
src/nerva_torch/loss_functions.py - Matrix primitives
src/nerva_torch/matrix_operations.py - Torch loss helpers
src/nerva_torch/loss_functions_torch.py
Data and utilities
- Batch loader
src/nerva_torch/datasets.py - NPZ dataset
- I/O and utilities
src/nerva_torch/utilities.py
Other components
- Experimenter
Connections 22
- Experimenter → MLP CLI: configures
- MLP CLI → Training loop: starts training
tools/mlp.py - Training loop → Batch loader: iterates batches
src/nerva_torch/training.py - Batch loader → NPZ dataset: loads batches
src/nerva_torch/datasets.py - Training loop → Learning rates: gets learning rate
src/nerva_torch/training.py - Training loop → Loss functions: computes loss gradient
src/nerva_torch/training.py - Training loop → MLP: runs and updates
README.md - Training loop → Training metrics: computes statistics
src/nerva_torch/training.py - MLP → Network layers: dispatches forward and backward
src/nerva_torch/multilayer_perceptron.py - Network layers → Activations: applies activation
src/nerva_torch/layers.py - Network layers → Softmax functions: computes softmax
src/nerva_torch/layers.py - Network layers → Optimizers: updates parameters
src/nerva_torch/layers.py - Network layers → Weight initialization: sets weights
src/nerva_torch/layers.py - Network layers → Matrix primitives: uses primitives
src/nerva_torch/layers.py - Loss functions → Softmax functions: uses log-softmax
src/nerva_torch/loss_functions.py - Loss functions → Activations: uses activation functions
src/nerva_torch/loss_functions.py - Loss functions → Matrix primitives: uses primitives
src/nerva_torch/loss_functions.py - Softmax functions → Matrix primitives: uses primitives
src/nerva_torch/softmax_functions.py - Batch loader → I/O and utilities: loads NPZ data
src/nerva_torch/datasets.py - MLP → I/O and utilities: loads and saves weights
src/nerva_torch/multilayer_perceptron.py - Training loop → I/O and utilities: formats diagnostics
src/nerva_torch/training.py - Torch loss helpers → Batch loader: uses data helpers
src/nerva_torch/loss_functions_torch.py