Class QuantAlignmentPass

ppq/quantization/optim/refine.py:292–534

## PPQ Quant Alignment Pass(通用量化对齐过程) When deploy on real hardware and inference framework, we will find that there are various restrictions or rules that we have to follow. * AVERAGE_POOL_2D: Input and outputs must all have same scale/zero_point * CONCATENATION:

class QuantAlignmentPass(QuantizationOptimizationPass):
    """
    ## PPQ Quant Alignment Pass(通用量化对齐过程)
    
    When deploy on real hardware and inference framework, 
        we will find that there are various restrictions or rules that we have to follow.

    * AVERAGE_POOL_2D: Input and outputs must all have same scale/zero_point
    
    * CONCATENATION: Input and outputs must all have same scale/zero_point
    
    * SLICE: Input and outputs must all have same scale/zero_point
    
    More detailed restrictions please refer to: https://www.tensorflow.org/lite/performance/quantization_spec

    Those restrictions, can be concluded as some quantization should share 
        the same quantization parameter with others. PPQ Quant Alignment Pass is designed
        for dealing with problems like this.

    PPQ uses Tensor Quantization Config (A data structure defined in ppq.core) to control the
        quantization logic, so to say if we want to align quantization parameters, we align
        their TQC in fact.

    The way to align TQC is simple, code like:
        tqc1.set_master(master=tqc2)
    Will make tqc1 and tqc2 share the same quantization parameters as tqc1 has, and change the
    state of tqc2 to be QuantizationState.PASSIVE

    If we access the scale of tqc2, PPQ will return its master TQC's scale instead, so does offset.

    That is tqc1 and tqc2 are bonuded with statement "tqc1.set_master(master=tqc2)".

    ### Parameters:

    * elementwise_alignment(Set[str]):
            
            Alignment method for elementwise ops.
            
            PPQ Supports 4 alignment methods:
                namely 'Align to Input', 'Align to Large', 'Align to Output', 'None'.
            
            All elementwise ops are listed in ppq.core.common.py

    * concat_alignment(Set[str])

            Alignment method for concat-like ops.
            
            PPQ Supports 4 alignment methods:
                namely 'Align to Input', 'Align to Large', 'Align to Output', 'None'.
            
            All concat-like ops are listed in ppq.core.common.py

    * averagepool_alignment(Set[str])

            Alignment method for pooling-like ops.

            PPQ Supports 4 alignment methods:
                namely 'Align to Input', 'Align to Large', 'Align to Output', 'None'.