tensorstudio 0.1.0

TensorStudio is a compact C++ tensor and autograd engine with a Python API for learning, experimentation, and lightweight ML workloads.

TensorStudio

TensorStudio is a compact C++ tensor and autograd engine with a Python API for learning, experimentation, and lightweight ML workloads.

TensorStudio is experimental v0.1.0 software. It is CPU-only, eager-only, and intentionally small enough to read and modify.

Install

From a source checkout:

python -m pip install -U pip
python -m pip install -e ".[dev]"

Build source and wheel distributions:

python -m build

Quickstart

import tensorstudio as ts

x = ts.tensor([[1.0, 2.0], [3.0, 4.0]])
y = ts.ones((2, 2))

print((x + y).tolist())
print((x @ y).numpy())

Autograd

import tensorstudio as ts

x = ts.tensor([1.0, 2.0, 3.0], requires_grad=True)
loss = ((x * x).sum())
loss.backward()

print(x.grad.tolist())  # [2.0, 4.0, 6.0]

Neural Networks

import tensorstudio as ts
from tensorstudio import nn, optim

model = nn.Sequential(nn.Linear(1, 8), nn.Tanh(), nn.Linear(8, 1))
optimizer = optim.SGD(model.parameters(), lr=0.05)
criterion = nn.MSELoss()

x = ts.tensor([[0.0], [1.0], [2.0], [3.0]])
y = ts.tensor([[1.0], [3.0], [5.0], [7.0]])

for _ in range(50):
    optimizer.zero_grad()
    loss = criterion(model(x), y)
    loss.backward()
    optimizer.step()

print(loss.item())

Development

python -m pip install -e ".[dev]"
pytest
ruff check .
mypy python/tensorstudio
python -m build

The native extension is built with CMake, pybind11, scikit-build-core, and C++20.

Documentation

The documentation lives in docs/ and is configured with MkDocs:

python -m pip install -e ".[docs]"
mkdocs serve
mkdocs build

The docs cover tensor semantics, broadcasting, NumPy interop, autograd, neural network modules, training loops, CPU backend details, benchmarks, development, publishing, and the project roadmap.

Publishing

Releases are intended to be built by GitHub Actions. The publish workflow uses PyPI trusted publishing through pypa/gh-action-pypi-publish; do not commit PyPI API tokens.

Current Limitations

• CPU backend only
• Eager execution only
• No CUDA kernels or mixed precision
• No graph compiler or distributed training
• Limited dtype casting and no advanced indexing
• No sparse tensors
• Pickle serialization is for trusted TensorStudio objects only

Roadmap

• CUDA backend
• Graph/JIT mode
• Convolution ops
• Dataset utilities
• Model zoo examples
• ONNX import/export
• Improved memory allocator
• SIMD kernels
• Multithreaded ops

License

TensorStudio is licensed under the MIT License.