tensor

package	build/tests	coverage
C++
Python

---

This library provides a two main features:

• A class for interacting with multidimensional arrays (For backend library uses BLAS/LAPACK libraries with fallback to own naive implementations).
• Deep neural networks.

The design goal is to create a numpy/pytorch alike interface for interacting with multidimensional arrays packaged in a simple, relatively lightweight, library with limited external dependencies that could be used on platforms like android phones and microcontrollers.

Features:

• tensor:
  • classes and utilities for interacting with nd-arrays
  • sane interface to gemm routines
• tensor/nn:
  • layers: FeedForward, Conv2D(naive/im2col), RNN, LSTM, Pooling, Dropout, BatchNormalization2D
  • optimizers: SGD(with momentum), Adagrad, RMSProp, Adam
  • saving/restoring models using protobuf
• Python
  • wrapper for major of the tensor and tensor/nn functionalities
  • experimental autograd module (PyTorch alike)

Coming soon:

• tensor:
  • improved naive GEMM implementation using AVX2 intrinsics
• tensor/nn:
  • layers: SelfAttention, LayerNorm

Examples

• C++:

  • nn-planar-data example
  • char-rnn example
  • embedded application

• Python:

  • MNIST
  • autograd capabilities

How to use in your C++ project

If you're using cmake see tensor-example for example usage.

How to use Python wrapper

Install the latest release pip install https://github.com/d-kicinski/tensor/releases/download/v0.2.0/tensor-0.2.0-cp38-cp38-linux_x86_64.whl. Alternatively, clone this repo and build it by yourself.

Documentation

For basic usages see this doc