MRT V3 API

Contents

• MRT V3 API

  • mrt.V3.utils

  • mrt.V3.prepare

  • mrt.V3.calibrate

  • mrt.V3.quantize

  • mrt.V3.evaluate

  • mrt.V3.mrt_compile

  • mrt.V3.execute

mrt.V3.utils

Utility Module for MRT V3.

Collector of utility functions including YAML configuration node manipulating functions, MRT stage helper functions. Default YAML configurations for MRT Common Stage options and Command line help prompt are also included.

mrt.V3.utils.get_model_prefix(model_dir, model_name)

Get the prefix of the pre-trained MRT model name.

Parameters

• model_dir (str) – Directory of the model file.

• model_name (str) – Name of the MRT pre-trained model.

Returns

model_prefix – Prefix of the pre-trained MRT model name.

Return type

str

mrt.V3.utils.get_logger(verbosity)

Get the console logger

Parameters

verbosity (str) – verbosity level chosen from debug, info, warning, error, fatal.

Returns

logger – Console logger.

Return type

logging.RootLogger

mrt.V3.utils.set_batch(input_shape, batch)

Get the input shape with respect to a specified batch value and an original input shape.

Parameters

• input_shape (tuple) – The input shape with batch axis unset.

• batch (int) – The batch value.

Returns

ishape – The input shape with the value of batch axis equal to batch.

Return type

tuple

mrt.V3.utils.load_fname(prefix, suffix=None, with_ext=False)

Get the model files at a given stage.

Parameters

• prefix (string) – The file path without and extension.

• suffix (string) – The file suffix with respect to a given stage of MRT.

• with_ext (bool) – Whether to include ext file.

Returns

files – The loaded file names.

Return type

tuple of string

mrt.V3.utils.save_conf(fname, logger=<RootLogger root (WARN)>, **conf_map)

Save the JSON-formatted MRT configuration from configuration checkpoint file.

Parameters

• fname (str) – Path of the JSON-formatted MRT configuration checkpoint file.

• logger (logging.RootLogger) – Console logger.

• conf_map (dict) – Dictionary of the attribute-value pairs.

mrt.V3.utils.load_conf(fname, logger=<RootLogger root (WARN)>)

Load the JSON-formatted MRT configuration from configuration checkpoint file.

Parameters

• fname (str) – Path of the JSON-formatted MRT configuration checkpoint file.

• logger (logging.RootLogger) – Console logger.

Returns

conf_map – Dictionary of the attribute-value pairs.

Return type

dict

mrt.V3.utils.check_file_existance(*fpaths, logger=<RootLogger root (WARN)>)

Check the existance of the listed file paths.

Parameters

• fpaths (list of str) – List of paths to be checked.

• logger (logging.RootLogger) – Console logger.

mrt.V3.utils.get_ctx(device_type, device_ids, dctx=cpu(0))

Get the context instance of mxnet

Parameters

• device_type (str) – context type string chosen from cpu or gpu.

• device_ids (list) – list of context ids

• dctx (mx.context.Context) – default context

Returns

context – The created context with respect to the device_type and device_ids.

Return type

mx.context.Context

mrt.V3.utils.get_batch_axis(input_shape)

Get the batch axis entry of an input shape.

Parameters

input_shape (tuple) – The data shape related to dataset.

Returns

axis – The batch axis entry of an input shape.

Return type

int

mrt.V3.utils.get_cfg_defaults()

Get a yacs CfgNode object with default values for MRT.

Returns

cfg – CfgNode represents an internal node in the configuration tree. It’s a simple dict-like container that allows for attribute-based access to keys.

Return type

yacs.config.CfgNode

mrt.V3.utils.merge_cfg(yaml_file)

Override the default YAML configuration node with the provided YAML-formatted configuration file.

Parameters

yaml_file (str) – Path of the YAML-formatted configuration file.

Returns

cfg – CfgNode represents an internal node in the configuration tree. It’s a simple dict-like container that allows for attribute-based access to keys.

Return type

yacs.config.CfgNode

mrt.V3.utils.revise_cfg(cfg, stage, attr, value)

Revise MRT YAML configuration node with respect to the specified stage and attribute name into the provided value.

Parameters

• cfg (yacs.config.CfgNode) – CfgNode represents an internal node in the configuration tree. It’s a simple dict-like container that allows for attribute-based access to keys.

• stage (str) – Stage name chosen from common, prepare, calibrate, quantize, evaluate or compile.

• attr (str) – Attribute name attribute to the provided stage.

• value (int, float, str, list, tuple, bool or NoneType) – The revision value to be applied, type of value should be supported by yacs.config.CfgNode

mrt.V3.utils.override_cfg_args(cfg, mrt_argv)

Override YAML configuration node with command line optional arguments for the simplicity of MRT configuration revision.

Parameters

• cfg (yacs.config.CfgNode) – CfgNode represents an internal node in the configuration tree. It’s a simple dict-like container that allows for attribute-based access to keys.

• mrt_argv (list) – list of even length which can be resoluted as key value pairs, the key could be split into stage name and attribute name.

Returns

cfg – Overridden YAML configuration node.

Return type

yacs.config.CfgNode

mrt.V3.prepare

Preparation Module for MRT V3.

Prepare function definition, default YAML configurations for MRT preparation Stage options and Command line help prompt are also included.

mrt.V3.prepare.prepare(cm_cfg, pass_cfg, logger=None)

YAML configuration API of MRT preparation stage.

Parameters

• cm_cfg (yacs.config.CfgNode) – CfgNode of common stage.

• pass_cfg (yacs.config.CfgNode) – CfgNode of preparation stage.

• logger (logging.RootLogger) – Console logger.

mrt.V3.calibrate

Calibration Module for MRT V3.

Calibration function definition, default YAML configurations for MRT calibration Stage options and Command line help prompt are also included.

mrt.V3.calibrate.calibrate(cm_cfg, pass_cfg, logger=None)

YAML configuration API of MRT calibration stage.

Parameters

• cm_cfg (yacs.config.CfgNode) – CfgNode of common stage.

• pass_cfg (yacs.config.CfgNode) – CfgNode of calibration stage.

• logger (logging.RootLogger) – Console logger.

mrt.V3.quantize

Quantization Module for MRT V3.

Quantize function definition, default YAML configurations for MRT quantization Stage options and Command line help prompt are also included.

mrt.V3.quantize.quantize(cm_cfg, pass_cfg, logger=None)

YAML configuration API of MRT quantization stage.

Parameters

• cm_cfg (yacs.config.CfgNode) – CfgNode of common stage.

• pass_cfg (yacs.config.CfgNode) – CfgNode of evaluation stage.

• logger (logging.RootLogger) – Console logger.

mrt.V3.evaluate

Evaluation Module for MRT V3.

Evaluate function definition, default YAML configurations for MRT evaluation Stage options and Command line help prompt are also included.

mrt.V3.evaluate.forward(net, data, ctx, baxis, olen)

Multiple xpu run support.

Parameters

• net (mxnet.gluon.block.SymbolBlock) – Graph for inference.

• data (mxnet.ndarray.ndarray.NDArray) – Input data to pass into the graph.

• ctx (mx.context.Context) – Context for inference.

• baxis (int) – Axis id of batch dimension.

• olen (int) – Length of the output.

Returns

outs – inference result of the graph with respect to the given input data, for multiple outputs, outs will be a list the entry type of which is mxnet.ndarray.ndarray.NDArray.

Return type

mxnet.ndarray.ndarray.NDArray or list

mrt.V3.evaluate.get_evaluation_info(cm_cfg, pass_cfg, logger=None)

YAML configuration API to get evaluation function, quantization function and dataset iteration function

Parameters

• cm_cfg (yacs.config.CfgNode) – CfgNode of common stage.

• pass_cfg (yacs.config.CfgNode) – CfgNode of calibration stage.

• logger (logging.RootLogger) – Console logger.

mrt.V3.evaluate.evaluate(cm_cfg, pass_cfg, logger=None)

YAML configuration API of MRT evaluation stage.

Parameters

• cm_cfg (yacs.config.CfgNode) – CfgNode of common stage.

• pass_cfg (yacs.config.CfgNode) – CfgNode of calibration stage.

• logger (logging.RootLogger) – Console logger.

mrt.V3.evaluate.get_ctx_eval(ctx)

Get the context instance for evaluation stage

Parameters

ctx (mx.context.Context) – The input context.

Returns

ctx – The modified context.

Return type

mx.context.Context

mrt.V3.evaluate.inference_original_model(symbol_file, params_file, data, batch_axis=0, device_type=None, device_ids=None)

MRT Inference API for original model.

Parameters

• symbol_file (str) – Path to the quantized mxnet symbol JSON file.

• params_file (str) – Path to the quantized mxnet parameters file.

• data (mxnet.ndarray.ndarray.NDArray) – Input data to pass into the graph.

• batch_axis (int) – Axis id of batch dimension.

• device_type (str) – Context type string chosen from cpu or gpu.

• device_ids (list) – List of context ids.

Returns

outs – inference result of the graph with respect to the given input data, for multiple outputs, outs will be a list the entry type of which is mxnet.ndarray.ndarray.NDArray.

Return type

mxnet.ndarray.ndarray.NDArray or list

mrt.V3.evaluate.inference_quantized_model(qsymbol_file, qparams_file, qext_file, data, batch_axis=0, split=False, device_type=None, device_ids=None)

MRT Inference API for quantized model.

Parameters

• qsymbol_file (str) – Path to the quantized mxnet symbol JSON file.

• qparams_file (str) – Path to the quantized mxnet parameters file.

• qext_file (str) – Path to the quantized extension file which store intermediate results.

• data (mxnet.ndarray.ndarray.NDArray) – Input data to pass into the graph.

• batch_axis (int) – Axis id of batch dimension.

• split (bool) – Flag indicating whether the model is split before quantization.

• device_type (str) – Context type string chosen from cpu or gpu.

• device_ids (list) – List of context ids.

Returns

outs – inference result of the graph with respect to the given input data, for multiple outputs, outs will be a list the entry type of which is mxnet.ndarray.ndarray.NDArray.

Return type

mxnet.ndarray.ndarray.NDArray or list

mrt.V3.mrt_compile

Compilation Module for MRT V3.

Compile function definition, default YAML configurations for MRT compilation Stage options and Command line help prompt are also included.

mrt.V3.mrt_compile.mrt_compile(cm_cfg, pass_cfg, logger=None)

YAML configuration API of MRT compilation stage.

Parameters

• cm_cfg (yacs.config.CfgNode) – CfgNode of common stage.

• pass_cfg (yacs.config.CfgNode) – CfgNode of compilation stage.

• logger (logging.RootLogger) – Console logger.

mrt.V3.execute

Execution Module for MRT V3.

Definition of execution functions for either MRT complete process execution or specified stages execution.

mrt.V3.execute.yaml_main(cfg, logger=None)

Execution function to launch the complete MRT process.

Parameters

• cfg (yacs.config.CfgNode) – CfgNode of MRT.

• logger (logging.RootLogger) – Console logger.

mrt.V3.execute.run(cfg, logger=None)

Execution function to launch either the complete MRT process or a specified MRT stage.

Parameters

• cfg (yacs.config.CfgNode) – CfgNode of MRT.

• logger (logging.RootLogger) – Console logger.