cudagrad 0.2.5

A tensor-valued autograd engine for Python

cudagrad

CUDA C++ strided float tensor automatic differentiation engine with Python bindings

Install

Available on PyPI:

pip install cudagrad

Distributed as:

• A source distribution, requiring the cmake command, and optionally the nvcc command, to be available at installation time.
• A binary distribution, targeting the Kaggle Python docker images:
  • CPU-only binary: Kaggle notebook example
  • GPU accelerator binary: Work in Progress

Examples

The following examples were written purely in Python using only cudagrad.Tensor for learning:

OR

0.52 seconds (59.5% faster than torch)

/benchmarks/_cudagrad/or.py

XOR

4.5 seconds (39.2% faster than torch)

/benchmarks/_cudagrad/xor.py

MOONS

14.25 seconds (5.8% slower than torch)

/benchmarks/_cudagrad/moons.py