Introduce automatic adding post-processing nodes to ignored scopes for quantization

Open kshpv opened this issue 3 years ago • 1 comments

Reopend #1255

Changes

Add pre-quantization step in ONNXMinMaxQuantization to add nodes in ignored scope.

Fix incorrect default parameters creation for PostTrainingQuantization and ONNXMinMaxQuantization. Now all parameters needed for ONNXMinMaxQuantization are passed from PostTrainingQuantization to ONNXMinMaxQuantization directly.

Reason for changes

Some of the models consist of post-processing part which should be disable for quantization, e.g. ssd_mobilent_v1 or Mask RCNN. This PR adds this functional.

Related tickets

88757

Tests

The test_add_ignored_scopes is added.

Oct 11 '22 11:10 kshpv

Created a ticket to investigate a default enabling this feature for Torch and TF - 94355

Oct 14 '22 12:10 kshpv