New feature, bug fixes, code modifications to create a CPP export of Resnet18 (quantified model)
Merged
requested to merge mick94/aidge_export_cpp:feat_add_quantized_export into feat_add_quantized_export
Context
Summary
This MR is or adapt, fix error and add new features to this module for support the ResNet18 quandfied model
Detailed major modifications
Work on operators:
-
Pooling -
[Feat] integration of rounding for an integer data type (pooling.hpp)
-
-
Rescaling -
[Feat] Add rescaling function (rescaling.hpp) -
[Feat] Add registrar file (CppRescaling.py) -
[Feat] Add config and forward jinja for rescaling operator (rescaling_config.jinja & rescaling_forward.jinja)
-
-
Elemwise -
[Refactor] Change "elemwise_op" attribute resolution (elemwise_config.py) -
[Refactor] Use rescaling.jinja file to collect rescaling attribute
-
Work on export object:
-
[Fix] Modify cpp_recipes: renaming operator to fix error (export_utils.py) -
[Feat] Add normalize function necessary to format data before PTQ (export_utils.py) -
[Feat] Add external input_tensor option. Can be use with main cpp generation fucntion (export.py) -
[Fix] Replacing a hard CAST with a CAST based on the output data type (_aidge_cmp.jinja) -
[Feat] create a new template for aidge_cmp function (for integer datatype)
Work to fix error for openmp option:
-
[Fix] Add "#ifdef _OPENMP" to enable or disable the OpenMP option of the compilation -
convolution.hpp -
fullyconnected.hpp -
leakyrelu.hpp -
pooling.hpp
-
Edited by Mickael GUIBERT
Merge request reports
Activity
Filter activity
Please register or sign in to reply