Quantization Configuration Parameters

**Table 1** retrain_enable
Description	QAT enable
Type	Boolean
Value	true or false
Command-Line Options	true: on false: off
Recommended Value	true
Required/Optional	Optional

**Table 2** retrain_data_config
Description	Activation quantization configuration
Type	Dictionary
Value	-
Command-Line Options	Includes the following parameters: batch_num: batch number for quantization. clip_max: upper bound of clipping-based quantization, defaulted to be empty. clip_min: lower bound of clipping-based quantization, defaulted to be empty. fixed_min: whether to fix the minimum value of clipping-based quantization to 0, defaulted to be empty. dst_type: Currently, only INT8 quantization bit width is supported. The default value is INT8.
Recommended Value	-
Required/Optional	Optional

**Table 3** retrain_weight_config
Description	Weight quantization configuration
Type	Dictionary
Value	-
Command-Line Options	Includes the following parameters: weights_retrain_algo: quantization algorithm select, defaulted to arq_retrain. channel_wise: whether to use different quantization factors for each channel. dst_type: Currently, only INT8 quantization bit width is supported. The default value is INT8.
Recommended Value	-
Required/Optional	Optional

**Table 4** weights_retrain_algo
Description	Weight quantization algorithm
Type	String
Value	-
Command-Line Options	ulq_quantize: ULQ clipping-based quantization algorithm arq_retrain: ARQ algorithm
Recommended Value	arq_retrain
Required/Optional	Optional

**Table 5** channel_wise
Description	Whether to use different quantization factors for each channel.
Type	Boolean
Value	true or false
Command-Line Options	true: Channels are separately quantized using different quantization factors. false: All channels are quantized altogether using the same quantization factors.
Recommended Value	true
Required/Optional	Optional

**Table 6** batch_num
Description	Batch number for quantization
Type	Integer
Value	Greater than 0
Command-Line Options	Defaults to 1. You are advised to keep the calibration dataset size within 50 images. Calculate batch_num based on batch_size as follows: batch_num × batch_size = Calibration dataset size batch_size indicates the number of images per batch.
Recommended Value	1
Required/Optional	Optional

**Table 7** fixed_min
Description	Fixed lower bound switch for the activation quantization algorithm
Type	Boolean
Value	true or false
Command-Line Options	true: fixes the lower bound of the activation quantization algorithm at 0. false: does not fix the lower bound of the activation quantization algorithm.
Recommended Value	Do not include this parameter.
Required/Optional	Optional

**Table 8** clip_max
Description	Upper bound for the activation quantization algorithm
Type	Float
Value	clip_max > 0 Find the maximum max based on the distribution of the activation values at different layers. The recommended value range is [0.3 × max, 1.7 × max].
Command-Line Options	If this parameter is included, the upper bound of the clipping-based activation quantization algorithm is fixed. If this parameter is not included, the upper bound is learned using the IFMR algorithm.
Recommended Value	Do not include this parameter.
Required/Optional	Optional

**Table 9** clip_min
Description	Lower bound for the activation quantization algorithm
Type	Float
Value	clip_min < 0 Find the minimum min based on the distribution of the activation values at different layers. The recommended value range is [0.3 × min, 1.7 × min].
Command-Line Options	If this parameter is included, the lower bound of the clipping-based activation quantization algorithm is fixed. If this parameter is not included, the lower bound is learned using the IFMR algorithm.
Recommended Value	Do not include this parameter.
Required/Optional	Optional

**Table 10** dst_type
Description	Quantization bit width select
Type	String
Value	Currently, only INT8 (default) is supported.
Command-Line Options	Selects the quantization bit width during quantization.
Recommended Value	-
Required/Optional	Optional

Parent topic: Single-Operator Mode