Simplified Retrain Configuration File

CANN V100R020C10 Development Auxiliary Tool Guide (Inference) 01

Huawei uses machine translation combined with human proofreading to translate this document to different languages in order to help you better understand the content of this document. Note: Even the most advanced machine translation cannot match the quality of professional translators. Huawei shall not bear any responsibility for translation accuracy and it is recommended that you refer to the English document (a link for which has been provided).

Table 3-37 describes the parameters in the retrain_config.proto file.

Table 3-37 Parameter description

Message	Required/Optional	Type	Field	Description
AMCTRetrainConfig	-	-	-	Simplified AMCT retrain configuration.
	Required	RetrainDataQuantConfig	retrain_data_quant_config	Data retrain configuration.
	Required	RetrainWeightQuantConfig	retrain_weight_quant_config	Weight retrain configuration.
	Repeated	string	skip_layers	Layers to skip during retrain.
	Repeated	RetrainOverrideLayer	override_layer_configs	Layer to override.
	Repeated	string	skip_layer_types	Types of layers to skip during retrain. This parameter is not supported in the current version.
	Repeated	RetrainOverrideLayerType	override_layer_types	Types of layers to override.
RetrainDataQuantConfig	-	-	-	Data retrain configuration.
RetrainDataQuantConfig	-	ULQuantize	ulq_quantize	Data quantization algorithm. Only ULQ is supported in the current version.
ULQuantize	-	-	-	ULQ algorithm configuration.
	Optional	ClipMaxMin	clip_max_min	Initial upper and lower limits. If not specified, IFMR is used for initialization.
	Optional	bool	fixed_min	Whether to fix the lower limit at 0. Set to true for ReLU or false for other algorithms.
ClipMaxMin	-	-	-	Initial upper and lower limits.
	Required	float	clip_max	Initial upper limit.
	Required	float	clip_min	Initial lower limit.
RetrainWeightQuantConfig	-	-	-	Weight retrain configuration.
RetrainWeightQuantConfig	-	ARQRetrain	arq_retrain	Weight quantization algorithm. Only ARQ is supported in the current version.
ARQRetrain	-	-	-	ARQ algorithm configuration.
ARQRetrain	Required	bool	channel_wise	Channel-wise ARQ enable.
RetrainOverrideLayer	-	-	-	Layer overriding configuration.
	Required	string	layer_name	Layer name.
	Required	RetrainDataQuantConfig	retrain_data_quant_config	Data quantization configuration to apply.
	Required	RetrainWeightQuantConfig	retrain_weight_quant_config	Weight quantization configuration to apply.
RetrainOverrideLayerType	-	-	-	Type of layers to override.
	Required	string	layer_type	Layer type.
	Required	RetrainDataQuantConfig	retrain_data_quant_config	Data quantization configuration to apply.
	Required	RetrainWeightQuantConfig	retrain_weight_quant_config	Weight quantization configuration to apply.

The following is an example simplified retrain configuration file (quant.cfg):

# global quantize parameter
 retrain_data_quant_config: {
     ulq_quantize: {
         clip_max_min: {
             clip_max: 6.0
             clip_min: -6.0
         }
     }
 }

 retrain_weight_quant_config: {
     arq_retrain: {
         channel_wise: true
     }
 }

 override_layer_types : {
     layer_type: "InnerProduct"
     retrain_weight_quant_config: {
         arq_retrain: {
            channel_wise: false
         }
     }
 }

 override_layer_configs : {   
     layer_name: "fc_5"   
     retrain_weight_quant_config: {        
         arq_retrain: {   
            channel_wise: false
         }    
      }
}

Translation

简体中文

Download

Updated: 2020-12-11

Document ID: EDOC1100164848

Downloads: 27

Average rating:

This Document Applies to these Products

Related Version

Related Documents