riccardosven / storch

A simple implementation of a torch-like autodiff engine.

Storch

Storch is a very simple torch-like tensor operation engine.

It includes tensor manipulation functions and graph operations for autodifferentiation; the rationale is a torch-like lazy-forward lazy-backward evaluation of a dynamic computational graph.

Build the project

 $ cmake -B build .
 $ cmake --build build

Run unit and integration tests

 $ ctest --test-dir build

Build documentation

 $ doxygen

Requirements

• cmake
• check
• openblas
• doxygen (to build documentation)

storch's Issues

Add tests for broadcasting graph operations.

When moving to broadcasting tensor operations, some graph operations may behave unexpectedly.
We need tests for all broadcasting graph operations to ensure correct functionality.

Add memcheck to CMakeLists

Run valgrind on unit and integration tests to prevent leaks.

Need to fix backpropagation for all broadcasting operations.

Backpropagation needs to be fixed for these broadcasting operations.

• G_Sum
• G_Diff
• G_Product
• G_Div
• G_Pow

After the generic gradient is computed, broadcasting needs to be detected and the appropriate sum-reduction applied.

Need to fix `-lcblas` vs `

`G_Pow_Backward` is wrong when using broadcasting operations.