How kylejeong2/rstorch works
RStorch is a Rust deep-learning framework with a C/CUDA tensor backend. Its principal workflow is to obtain data, run it through neural-network modules, compute a loss and gradients, then update parameters with an optimizer. Data loading, MNIST/transforms, autograd, device kernels, and distributed training are distinct supporting capabilities. Edges reflect README statements, shown calls, or SOURCE INDEX references; some backend and distributed internals were not sampled, so their details are not inferred.
Main entry points: tensor.rs Tensor API · linear.rs Neural Modules · tensor.cpp Tensor Dispatcher · dataloader.rs DataLoader
Components 22
Tensor Runtime
- Tensor API
src/tensor.rs - Tensor Dispatcher
src/csrc/tensor.cpp - CPU Kernels
src/csrc/cpu.cpp - CUDA Kernels
src/csrc/cuda.h - Autograd Functions
src/autograd/functions.rs - C Tensor Struct
src/csrc/tensor.h
Model Training
- Neural Modules
src/nn/modules/linear.rs - Module Interface
src/nn/module.rs - Activations
src/nn/activation.rs - Loss Functions
src/nn/loss.rs - Parameters
src/nn/parameter.rs - Optimizer API
src/optim/optimizer.rs - SGD
src/optim/optimizers/sgd.rs
Data Input
- Dataset
src/utils/data/dataset.rs - Examples and Fields
src/utils/data/example.rs - DataLoader
src/utils/data/dataloader.rs - Batch
src/utils/data/batch.rs - MNIST Dataset
src/torchvision/datasets/mnist.rs - Image Transforms
src/torchvision/transforms.rs
Compute Scaling
- Distributed Operations
src/distributed/distributed.rs - Data Parallel
src/nn/parallel.rs
Other components
- Model Author
Connections 26
- Model Author → DataLoader: requests batches
- DataLoader → Dataset: iterates
src/utils/data/dataloader.rs - DataLoader → Batch: forms batches
src/utils/data/dataloader.rs - Dataset → Examples and Fields: stores examples
src/utils/data/dataset.rs - MNIST Dataset → Dataset: provides dataset
src/torchvision/datasets/mnist.rs - Model Author → Neural Modules: builds model
- Neural Modules → Module Interface: implements
src/nn/modules/linear.rs - Neural Modules → Parameters: creates parameters
src/nn/modules/linear.rs - Neural Modules → Tensor API: computes with
src/nn/modules/linear.rs - Activations → Tensor API: applies operations
src/nn/activation.rs - Loss Functions → Tensor API: computes loss
src/nn/loss.rs - Model Author → Loss Functions: computes objective
- Model Author → Autograd Functions: backpropagates
- Autograd Functions → Tensor API: records gradients
- Model Author → SGD: steps optimizer
- SGD → Optimizer API: implements
src/optim/optimizers/sgd.rs - Optimizer API → Tensor API: updates parameters
src/optim/optimizer.rs - Tensor API → C Tensor Struct: uses FFI layout
src/csrc/tensor.h - Tensor Dispatcher → C Tensor Struct: uses tensor API
src/csrc/tensor.cpp - Tensor Dispatcher → CPU Kernels: dispatches CPU ops
src/csrc/tensor.cpp - Tensor Dispatcher → CUDA Kernels: dispatches GPU ops
src/csrc/tensor.cpp - Tensor API → Tensor Dispatcher: calls backend
- Image Transforms → Tensor API: transforms tensors
src/torchvision/transforms.rs - Data Parallel → Distributed Operations: uses collectives
src/nn/parallel.rs - Data Parallel → Module Interface: wraps model
src/nn/parallel.rs - Distributed Operations → Tensor API: operates on tensors