Learning backend cuda

Context

Suppport Learning with backend_cuda:

• For DataProvider, add a backend attribute to allow reading batch on supported backends
• Add operators to compute Accuracy with a graph: ArgMax, And, ReduceSum

Modified files

• DataProvider.hpp , DataProvider.cpp and pybind_DataProvider.cpp, add backend member to DataProvider;
• ArgMax.hpp , ArgMax.cpp and pybind_ArgMax.cpp, add ArgMax operator;
• And.hpp , And.cpp and pybind_And.cpp, add And operator;
• ReduceSum.hpp , ReduceSum.cpp and pybind_ReduceSum.cpp, add ReduceSum operator;

This MR also fixes: #155 (closed)

TODO

• Support backend_cuda in DataProvider
• Add ArgMax operator for Accuracy computation
• Add And operator for Accuracy computation
• Add ReduceSum operator for Accuracy computation

added 1 commit

• 59f06670 - add default values for axes argument

Compare with previous version

added 1 commit

• ed790d61 - add noop_with_empty_axes attr

Compare with previous version

added 1 commit

• 2629485b - change keepDims and noopWithEmptyAxes into bool

Compare with previous version

changed the description

added 30 commits

• 2629485b...2f0b5ba2 - 29 commits from branch eclipse/aidge:dev
• eab19d86 - Merge branch 'dev' of gitlab.eclipse.org:eclipse/aidge/aidge_core into learning_backend_cuda

Compare with previous version

added 1 commit

• f653463e - fix merge error

Compare with previous version

added 1 commit

• e316161e - fix merge error

Compare with previous version

requested review from @thibaultallenet

Review for 3 MR 'Learning backend cuda' in core, backend_cpu, backend_cuda.

Content

• Add a check gradient (based on the numerical differentiation) tool to test the backpropagation kernels. Reference : https://github.com/CEA-LIST/N2D2/blob/master/src/GradientCheck.cpp

@olivierbichler , Any idea how to incorporate this tool ? I think it is an important tool to test backward kernels, especially before trying to reproduce Quantization methods like LSQ or SAT.

Documentation

• Document each arguments of the constructors (for each operator : And, ArgMax, ReduceSum, ReduceMean) in the C++ doxygen documentation and python (pybind files). The locations have been highlighted in comments.

Tests

• In the backend_cpu MR : add tests for the forwardDims functions of each operator (especially ReduceSum, ReduceMean)
• Add tests that covers the new part of code in the TensorImpl_backend_cuda : aidge_backend_cuda!33 (diffs)

added 1 commit

• d6e9c816 - add doc for ne woperators

Compare with previous version

resolved all threads

added 26 commits

• d6e9c816...0e48b653 - 25 commits from branch eclipse/aidge:dev
• 8d287f91 - Merge branch 'dev' of gitlab.eclipse.org:eclipse/aidge/aidge_core into learning_backend_cuda

Compare with previous version