Class MkldnnQuantizerConfig

Class Documentation

class paddle::MkldnnQuantizerConfig

Config for mkldnn quantize.

The MkldnnQuantizerConfig is used to configure Mkldnn’s quantization parameters, including scale algorithm, warmup data, warmup batch size, quantized op list, etc.

It is not recommended to use this config directly, please refer to AnalysisConfig::mkldnn_quantizer_config()

Public Functions

MkldnnQuantizerConfig()

Construct a new Mkldnn Quantizer Config object.

void SetScaleAlgo(std::string op_type_name, std::string conn_name, ScaleAlgo algo)

Set the scale algo.

Specify a quantization algorithm for a connection (input/output) of the operator type.

Parameters

  • [in] op_type_name: the operator’s name.

  • [in] conn_name: name of the connection (input/output) of the operator.

  • [in] algo: the algorithm for computing scale.

ScaleAlgo scale_algo(const std::string &op_type_name, const std::string &conn_name) const

Get the scale algo.

Get the quantization algorithm for a connection (input/output) of the operator type.

Return

the scale algo.

Parameters

  • [in] op_type_name: the operator’s name.

  • [in] conn_name: name of the connection (input/output) of the operator.

void SetWarmupData(std::shared_ptr<std::vector<PaddleTensor>> data)

Set the warmup data.

Set the batch of data to be used for warm-up iteration.

Parameters

  • [in] data: batch of data.

std::shared_ptr<std::vector<PaddleTensor>> warmup_data() const

Get the warmup data.

Get the batch of data used for warm-up iteration.

Return

the warm up data

void SetWarmupBatchSize(int batch_size)

Set the warmup batch size.

Set the batch size for warm-up iteration.

Parameters

  • [in] batch_size: warm-up batch size

int warmup_batch_size() const

Get the warmup batch size.

Get the batch size for warm-up iteration.

Return

the warm up batch size

void SetEnabledOpTypes(std::unordered_set<std::string> op_list)

Set quantized op list.

In the quantization process, set the op list that supports quantization

Parameters

  • [in] op_list: List of quantized ops

const std::unordered_set<std::string> &enabled_op_types() const

Get quantized op list.

Return

list of quantized ops

void SetExcludedOpIds(std::unordered_set<int> op_ids_list)

Set the excluded op ids.

Parameters

  • [in] op_ids_list: excluded op ids

const std::unordered_set<int> &excluded_op_ids() const

Get the excluded op ids.

Return

exclude op ids

void SetDefaultScaleAlgo(ScaleAlgo algo)

Set default scale algorithm.

Parameters

  • [in] algo: Method for calculating scale in quantization process

ScaleAlgo default_scale_algo() const

Get default scale algorithm.

Return

Method for calculating scale in quantization process

Protected Attributes

std::map<std::string, std::map<std::string, ScaleAlgo>> rules_
std::unordered_set<std::string> enabled_op_types_
std::unordered_set<int> excluded_op_ids_
std::shared_ptr<std::vector<PaddleTensor>> warmup_data_
int warmup_bs_ = {1}
ScaleAlgo default_scale_algo_ = {ScaleAlgo::MAX}