Simplified Configuration File for Auto Channel Pruning Search

The basic_info.proto file describes configurations for auto channel pruning search. Find the file in /amct_pytorch/proto/basic_info.proto under the AMCT installation directory.

Message	Required/Optional	Type	Field	Description
AutoMixedPrecisionConfig	-	-	-	Simplified configuration for AMCT automatic mixed-precision search. This feature is not supported in the current version.
	Optional	float	compress_ratio	Compression ratio. The computation amount of all quantizable layers is used as a reference compression multiple.
	Repeated	QuantBitLimit	quant_bit_limit	Quantization bit width search range of specified layers.
	Optional	string	ptq_cfg	Simplified PTQ configuration file, which is used to obtain the quantization factors under the quantized data types of INT4 and INT8 during calibration. If this parameter is not set, the default PTQ configuration is used.
	Optional	int64	test_iteration	Number of batches to dump. The dump data is used to measure the quantization impact and computation amount. The data size should be large enough to be representative.
	Optional	string	override_qat_cfg	Simplified QAT configuration file. The output of automatic mixed-precision search will overwrite the bit width of involved layers, and other parameters remain unchanged. If this parameter is not set, the simplified QAT configuration file (.proto) is used to generate a .cfg configuration file with quantized data type information.
AutoChannelPruneConfig	-	-	-	Simplified configuration for AMCT auto channel pruning search
	Required	float	compress_ratio	Compression ratio. The computation amount of all quantizable layers is used as a reference compression multiple.
	Optional	bool	ascend_optimized	Whether to perform Ascend affinity optimization. If the pruned model needs to be deployed on the AI processor, you are advised to set this parameter to true.
	Optional	float	max_prune_ratio	Maximum sparsity ratio of a single layer. The default value is 1.
	Optional	int64	test_iteration	Batch number of the input test data.
	Optional	string	override_prune_cfg	Simplified configuration file for channel pruning set by the user. It can contain only the skip and override configurations. The configured layers use the specified configuration and will not be rewritten by the auto channel pruning search APIs.
QuantBitLimit	-	-	-	Quantization bit width search range of specified layers.
	Optional	string	layer_name	Layer name.
	Repeated	DataType	data_range	Quantization bit width range.
DataType	-	-	-	Quantization bit width range, which is of the enumeration type. The current version supports only INT8 quantization.
	-	-	FLOAT	Float. Not quantized.
	-	-	INT8	INT8 quantization.
	-	-	INT4	INT4 quantization.

The following is an example of the simplified configuration file (amc.cfg) for auto channel pruning:

compress_ratio: 1.5
ascend_optimized: true
max_prune_ratio: 0.8
test_iteration: 1
override_prune_cfg: 'your/path/to/override_channel_prune.cfg'

Parent topic: Reference