V-Torch

A torch library for V

V-Torch is a deep learning library for the V programming language .

Roadmap & Implementation Plan

THIS IS AN EARLY DEV STAGE, PLEASE FEEL FREE TO CONTRIBUTE

The development is prioritized as follows to reach a working prototype:

1. Tensor Module (CPU only)

   • Data storage ( []f32 )
   • Shape & Strides ( compute_strides )
   • Basic Arithmetic (Add, Sub, Mul, Div, Matmul)
   • Broadcasting support

2. Autograd Module

   • Computation Graph (DAG)
   • Backward engine (Recursive/Topological sort)
   • Function interface for custom ops

3. NN Module

   • Module interface
   • Linear Layers
   • Activations (ReLU, Sigmoid, Tanh)
   • Loss Functions (MSE, CrossEntropy)

4. Optimizer Module

   • Optimizer interface
   • SGD
   • Adam

5. Verification & Examples

   • XOR Example
   • MNIST Training Script

Project Structure

The planned directory structure for the library:

V-lib/
├─ tensor/       # Core Tensor struct + arithmetic + broadcasting
├─ autograd/     # Backward engine, DAG, Function interface
├─ nn/           # Module interface, Linear, Conv, activations, loss
├─ optim/        # SGD, Adam, Optimizer interface
├─ examples/     # Training scripts (MNIST, XOR)
├─ tests/        # _test.v files for each module
└─ main.v        # Entry point for demos

Goals

• Pure V : Focus on correctness and API ergonomics in pure V first.
• Modularity : Each module should be self-contained with its own tests.
• Performance : Start with correctness, then optimize (e.g., single allocation for intermediates).