Manual Tuning

If the PTQ accuracy does not meet the requirements, you can manually adjust the parameters in the config.json file. This section provides the adjustment principles and parameter description.

Workflow

If you find that the accuracy of the model quantized based on the initial config.json file generated by the create_quant_config API call is not as expected, you can tune the configuration parameters until the accuracy meets your requirement. The workflow for manually tuning the parameters in the PTQ configuration file config.json goes through the following three phases:

Tune the amount of data used for calibration.
Skip quantizing certain layers.
Tune the quantization algorithm and parameters.

Specifically,

Run quantization based on the initial config.json file generated by the create_quant_config API. If the accuracy of the quantized model is satisfactory, stop tuning the configuration parameters. Otherwise, go to 2.
Tweak the value of batch_num to tune the amount of data used for calibration.
batch_num controls the batch number for quantization. Tune it based on the batch size and the dataset size. Generally:

A larger value of batch_num indicates more data samples used for quantization and a smaller accuracy drop of the quantized model. However, excessive data does not necessarily improve accuracy, but certainly consumes more memory and reduces the quantization speed, hence resulting in insufficient memory, video RAM, and thread resources. An optimal tradeoff is achieved when the product of batch_num and batch_size (the number of images per batch) is 16 or 32.
Run quantization based on the new configuration generated in 2. If the accuracy of the quantized model is satisfactory, stop tuning the configuration parameters. Otherwise, go to 4.
Tweak the value of quant_enable to skip quantizing certain layers.
quant_enable is the quantization switch of a specified layer. The value false indicates that the layer will be skipped during quantization; true, otherwise. Removing the layer configuration can also skip the layer.

Quantizing a model can have a negative effect on accuracy. Layers sensitive to quantization will suffer from remarkable error increases once quantized and therefore should be left dequantized. Spot these layers as follows:
1. In a model, the input layer, output layer, and layers with especially fewer parameters are likely to be quantization-sensitive.
2. Use the Model Accuracy Analyzer to compare the output errors between the original model and the quantized model layer-wise (a cosine similarity of at least 0.99, for example) to locate the layers that reduce accuracy the most and dequantize them with priority.
Run quantization based on the new configuration generated in 4. If the accuracy of the quantized model is satisfactory, stop tuning the configuration parameters. Otherwise, go to 6.
Tweak the values of activation_quant_params and weight_quant_params to tune the quantization algorithms and parameters.
For details about the algorithm parameters, see the parameter description in Quantization Configuration File. For details about the algorithm, see PTQ Algorithms.
Run quantization based on the new configuration generated in 6. If the accuracy of the quantized model is satisfactory, stop tuning the configuration parameters. Otherwise, it indicates that your model is not suitable for quantization and the quantization configuration should be removed.

Figure 1 Configuration tuning workflow

Quantization Configuration File

If inference based on the config.json quantization configuration file generated by the create_quant_config call has significant accuracy drop, tune the config.json file until the accuracy is as expected. For details about the JSON quantization configuration file example, see Example. Keep the layer names unique in the file.

The following tables describe the parameters in the configuration file.

**Table 1** version
Description	Version number of the quantization configuration file
Type	Integer
Value	1
Command-Line Options	Currently, only version 1 is available.
Recommended Value	1
Required/Optional	Optional

**Table 2** 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 3** activation_offset
Description	Symmetric quantization or asymmetric quantization select for activation quantization. It is a global configuration parameter. The asymmetric parameter takes precedence over the activation_offset parameter if both of them exist in the configuration file.
Type	Boolean
Value	true or false
Command-Line Options	true: asymmetric quantization false: symmetric quantization
Recommended Value	true
Required/Optional	Optional

**Table 4** do_fusion
Description	Fusion switch
Type	Boolean
Value	true or false
Command-Line Options	true: on false: off Currently, only Conv+BN fusion is supported.
Recommended Value	true
Required/Optional	Optional

**Table 5** skip_fusion_layers
Description	Layers to skip fusion
Type	String
Value	Must be names of fusible layers. Currently, only Conv+BN fusion is supported.
Command-Line Options	Sets the layers to skip fusion.
Recommended Value	-
Required/Optional	Optional

**Table 6** layer_config
Description	Quantization configuration of a network layer
Type	Object
Value	-
Command-Line Options	Includes the following parameters: quant_enable activation_quant_params weight_quant_params
Recommended Value	-
Required/Optional	Optional

**Table 7** quant_enable
Description	Quantization enable
Type	Boolean
Value	true or false
Command-Line Options	true: on false: off
Recommended Value	true
Required/Optional	Optional

**Table 8** dmq_balancer_param
Description	Migration strength of the DMQ Balancer
Type	Float
Value	[0.2, 0.8]
Command-Line Options	Degree to which the quantization difficulty of activations is migrated to weights. Set the migration strength to a small value if there are many outliers in the activation distribution.
Recommended Value	0.5
Required/Optional	Optional

**Table 9** activation_quant_params
Description	Activation quantization parameters
Type	Object
Value	-
Command-Line Options	Includes the following parameters. (Beware that IFMR algorithm parameters are mutually exclusive with HFMG ones at the same layer.) IFMR algorithm parameters: max_percentile min_percentile search_range search_step act_algo num_bits asymmetric HFMG algorithm parameters: act_algo num_of_bins num_bits asymmetric
Recommended Value	-
Required/Optional	Optional

**Table 10** weight_quant_params
Description	Weight quantization parameters
Type	Object
Value	-
Command-Line Options	Includes the following parameters in uniform quantization: num_bits wts_algo channel_wise
Recommended Value	-
Required/Optional	Optional

**Table 11** num_bits
Description	Quantization bit width
Type	Integer
Value	8 or 16
Command-Line Options	Currently, the value can only be 8, indicating that the INT8 quantization bit width is used.
Recommended Value	-
Required/Optional	Required

**Table 12** act_algo
Description	Activation quantization algorithm
Type	String
Value	ifmr or hfmg
Command-Line Options	ifmr: IFMR algorithm for activation quantization hfmg: HFMG algorithm for activation quantization
Recommended Value	-
Required/Optional	Optional

**Table 13** asymmetric
Description	Symmetric quantization or asymmetric quantization select for activation quantization. It is used to select the layer-wise quantization algorithm. The asymmetric parameter takes precedence over the activation_offset parameter if both of them exist in the configuration file.
Type	Boolean
Value	true or false
Command-Line Options	true: asymmetric quantization false: symmetric quantization
Recommended Value	true
Required/Optional	Optional

**Table 14** max_percentile
Description	Upper bound for searching for the largest of the IFMR activation quantization algorithm
Type	Float
Value	(0.5,1]
Command-Line Options	For example, given 100 numeric values in descending order, the upper bound 1.0 indicates that the value indexed 0 (100 – 100 × 1.0) is considered as the largest. A larger value indicates that the upper bound for clipping-based quantization is closer to the maximum value of the data to be quantized.
Recommended Value	0.999999
Required/Optional	Optional

**Table 15** min_percentile
Description	Lower bound for searching for the smallest of the IFMR activation quantization algorithm
Type	Float
Value	(0.5,1]
Command-Line Options	For example, given 100 numeric values in ascending order, the lower bound 1.0 indicates that the value indexed 0 (100 – 100 × 1.0) is considered as the smallest. A larger value indicates that the lower bound for clipping-based quantization is closer to the minimum value of the data to be quantized.
Recommended Value	0.999999
Required/Optional	Optional

**Table 16** search_range
Description	Quantization factor search range ([search_range_start, search_range_end]) of the IFMR algorithm
Type	A list of two floats
Value	0 < search_range_start < search_range_end
Command-Line Options	Sets the quantization factor search range. search_range_start: search start search_range_end: search end
Recommended Value	[0.7, 1.3]
Required/Optional	Optional

**Table 17** search_step
Description	Quantization factor search step of the IFMR algorithm
Type	Float
Value	(0, (search_range_end – search_range_start)]
Command-Line Options	Sets the fluctuation step of the upper bound for clipping-based quantization. A smaller value indicates a smaller quantization factor search step. The number of search iterations is calculated as: search_iteration = (search_range_end – search_range_start)/search_step. Increasing the number of search iterations will increase the search time and lead to process suspension.
Recommended Value	0.01
Required/Optional	Optional

**Table 18** num_of_bins
Description	Number of bins (the minimum unit in a histogram) of the HFMG algorithm
Type	Unsigned integer
Value	{1024, 2048, 4096, 8192}
Command-Line Options	A larger value of num_of_bins leads to better distribution fitting of the histogram and better quantization effect, but it also incurs longer PTQ time.
Recommended Value	4096
Required/Optional	Optional for quantization using the HFMG algorithm.

**Table 19** wts_algo
Description	Weight quantization algorithm
Type	String
Value	arq_quantize
Command-Line Options	arq_quantize: ARQ algorithm
Recommended Value	-
Required/Optional	Optional

**Table 20** channel_wise
Description	Whether to use different quantization factors for each channel in the ARQ algorithm.
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

Parent topic: Manual Quantization