Crate scivex_nn 

scivex-nn

Neural networks with reverse-mode automatic differentiation for the Scivex ecosystem.

Modules

Module	Contents
variable	Variable<T> — autograd computation graph node
ops	Differentiable operations (add, mul, matmul, etc.)
functional	Activation functions (relu, sigmoid, tanh, softmax)
layer	Layer trait, Linear, Conv1d/2d/3d, BatchNorm1d/2d, pooling, RNN/LSTM/GRU, attention, Sequential
optim	Optimizer trait, SGD, Adam, AdamW, RMSprop, Adagrad, LR schedulers
loss	Loss functions (MSE, cross-entropy, BCE, Huber, focal, KL, hinge, smooth L1)
init	Weight initialization (Xavier, Kaiming)
persist	save_weights, load_weights — binary weight persistence
data	Dataset trait, TensorDataset, DataLoader

Re-exports

pub use error::NnError;

pub use error::Result;

pub use variable::Variable;

Modules

data

Dataset, DataLoader, and batching utilities. Dataset and data loading utilities.

error

Neural network error types.

explain

Model interpretability: Integrated Gradients, SmoothGrad. Model interpretability: Integrated Gradients and SmoothGrad.

functional

Activation functions (ReLU, sigmoid, tanh, softmax). Differentiable activation functions.

init

Weight initialization strategies (Xavier, Kaiming). Weight initialization strategies.

layer

Layer trait and built-in layers (Linear, BatchNorm, Dropout). Neural network layers.

loss

Loss functions (MSE, cross-entropy, BCE). Loss functions for training neural networks.

onnx

ONNX model loading and inference. ONNX model loading and inference.

ops

Differentiable tensor operations for the autograd graph. Differentiable operations on Variables.

optim

Optimizers (SGD, Adam, AdamW, RMSprop, Adagrad). Optimizers and learning rate schedulers for training neural networks.

persist

Weight persistence: save and load model parameters. Neural network weight persistence — save and load model parameters.

prelude

Convenience re-exports.

serialize

Model serialization formats (SafeTensors, GGUF). Model serialization formats: SafeTensors and GGUF.

serve

In-process model inference server with batching support. Lightweight in-process model inference server with batching support.

training

Training utilities (Trainer, callbacks, gradient clipping). Training utilities: trainer loop, callbacks, gradient clipping, mixed precision, and LR finder.

variable

Autograd computation graph node.