# --------------------------------------------- # This is a quantizer template for custimizing. # --------------------------------------------- from typing import Union import torch from ppq.core import (PASSIVE_OPERATIONS, OperationQuantizationConfig, QuantizationPolicy, QuantizationProperty, QuantizationStates, RoundingPolicy, TargetPlatform) from ppq.IR import BaseGraph, GraphCommandProcesser from ppq.IR.base.graph import Operation from .base import BaseQuantizer class ExtQuantizer(BaseQuantizer): """ ExtQuantizer 是一个空的量化器模板，你可以在这里面实现你自定义的平台量化逻辑 你需要实现该类所有成员函数与方法，并且使用 TargetPlatform.EXTENSION 调用这个量化器 --> quantize_torch_model(..., platform=EXTENSION) ExtQuantizer is a empty quantizer template for you to implement custimized quantization logic. Implement all member funtion and properties, and invoke this with TargetPlatform.EXTENSION --> quantize_torch_model(..., platform=EXTENSION) Args: BaseQuantizer ([type]): [description] """ def __init__( self, graph: Union[BaseGraph, GraphCommandProcesser] ) -> Union[torch.Tensor, list, dict]: # Quantization basic setting ------------------------- # update following properties: self._num_of_bits = 8 # platform bit width. self._quant_min = -128 # int value min self._quant_max = +127 # int value max # ---------------------------------------------------- super().__init__(graph=graph) # do not forget to initialize super class. def init_quantize_config(self, operation: Operation) -> OperationQuantizationConfig: """ 使用这个函数来初始化算子的量化配置，该函数会被父类作为接口调用。 我们提供了一个方便的函数 create_default_quant_config 来帮助你实现相关逻辑 调用这个函数将直接产生一个默认的量化配置，你可以在此基础上进行修改。 注意并不是所有算子都可以使用默认量化配置完成量化，有些算子明显具有更加复杂的量化逻辑 对于这种情况，你需要在此函数中手动创建它们的量化配置信息。 Initial your quantization configuration for each operation. (This function will be invoked by BaseQuantizer as an interface.) We provide a helper function called self.create_default_quant_config() use this funtion to create a default quantization configuration. However, some operations' quantization logic might be much more complex so that default configuration won't meet their requirements. In this situation a manually initialized configuration is demended. Args: operation (Opeartion): [description] Returns: OperationQuantizationConfig: [description] """ # create a basic quantization configration. config = self.create_default_quant_config( operation_meta=operation.meta_data, num_of_bits=self._num_of_bits, quant_max=self._quant_max, quant_min=self._quant_min, observer_algorithm='percentile', policy=self.quantize_policy, rounding=self.rounding_policy, ) # initialze configuration for conv manually if operation.type == 'Conv': # override initialized config if operation.num_of_input == 3: bias_config = config.input_quantization_config[-1] # bias should be quantized with 32 bits # in python3, int indicates long long in C++ # so that it has enough precision to represent a number like 2^32 # however, it may cause a scale underflow # here we give bias a 30 bits precision, which is pettery enough in all cases bias_config.num_of_bits = 30 bias_config.quant_max = int(pow(2, 30 - 1)) bias_config.quant_min = - int(pow(2, 30 - 1)) bias_config.policy = QuantizationPolicy( QuantizationProperty.SYMMETRICAL + QuantizationProperty.LINEAR + QuantizationProperty.PER_TENSOR + QuantizationProperty.POWER_OF_2) bias_config.state = QuantizationStates.PASSIVE_INIT for tensor_config in config.input_quantization_config[1: ]: tensor_config.observer_algorithm = 'minmax' # mark some operation as passive op. # all quantizations of a passive op will be overlapped during graph merge pass. if operation.type in PASSIVE_OPERATIONS: # Those op are not active op. config.is_active_quant_op = False return config @ property def target_platform(self) -> TargetPlatform: """ target_platform 属性是提供给子图切分使用的， 所有量化区的算子将被调度到这个设备上。 Property target_platform is acquired by graph dispather. It states where your quantized operation depoly. """ return TargetPlatform.EXTENSION @ property def default_platform(self) -> TargetPlatform: """ default_platform 属性是提供给子图切分使用的， 所有冲突区的算子将被调度到这个设备上。 Property default_platform is acquired by graph dispather. It states where non-quantable operation depoly. """ return TargetPlatform.FP32 @ property def quant_operation_types(self) -> set: """ quant_operation_types 指明了所有可以被量化的算子类型。 聪明的你可能想问了，如果我不是按类型进行量化，而是想特定去量化某几个算子怎么办。 我建议你通过手写一个调度表来实现这样的功能，只需要把其他算子调度到FP32上就行了。 请注意，并不是你写在这里的类型就一定会被量化，在量化之前还有调度器进行子图切分， 只有量化子图上的算子才可以被量化。 Property quant_operation_types contains all quantable operations' type. Notice that PPQ has a dispatching logic before quantizer, so to say quantizer is only in charge of quantable subgraph, those operation within non-quantable subgraph will never be quantized even their type is listed here. """ return { 'Conv', 'ConvTranspose', 'Gemm', 'Relu', 'PRelu', 'Clip', 'Pad', 'Resize', 'MaxPool', 'AveragePool', 'GlobalMaxPool', 'GlobalAveragePool', 'Mul', 'Add', 'Max', 'Sub', 'Div', 'LeakyRelu', 'Concat', 'Sigmoid', 'Slice' } @ property def quantize_policy(self) -> QuantizationPolicy: """ quantize_policy 指明了默认量化策略 被函数 create_default_quant_config 所使用 quantize_policy is used by create_default_quant_config(), to generate a default quantization configuration. Returns: QuantizationPolicy: [description] """ return QuantizationPolicy( QuantizationProperty.SYMMETRICAL + QuantizationProperty.LINEAR + QuantizationProperty.PER_TENSOR + QuantizationProperty.POWER_OF_2 ) @ property def rounding_policy(self) -> RoundingPolicy: """ rounding_policy 指明了默认取整策略 被函数 create_default_quant_config 所使用 rounding_policy is used by create_default_quant_config(), to generate a default quantization configuration. Returns: QuantizationPolicy: [description] Returns: RoundingPolicy: [description] """ return RoundingPolicy.ROUND_HALF_UP