accuracy_based_auto_calibration

Function Usage

Calibrates the input model based on the input configuration file to search for a quantization configuration that meets accuracy requirements, and outputs a model suitable for both accuracy simulation in the TensorFlow environment and inference deployment on Ascend AI Processor.

Constraints

None.

Prototype

accuracy_based_auto_calibration(model_file,outputs,record_file,config_file,save_dir,evaluator,strategy='BinarySearch',sensitivity='CosineSimilarity')

Command-Line Options

Option	Input/Return	Meaning	Restriction
model_file	Input	Original TensorFlow model file (.pb).	A string
outputs	Input	List of output nodes.	A string
config_file	Input	Quantization configuration file generated by the user.	A string
record_file	Input	Quantization factor record file. The existing file (if any) in the path will be overwritten upon this API call.	A string
save_dir	Input	Model save path. Must include the prefix of the model name, for example, *./quantized_model/model**.	A string
evaluator	Input	Python instance for automatic quantization calibration and accuracy evaluation.	Data type: Python instance
strategy	Input	Policy for searching for the quantization configuration that meets the accuracy requirements. The dichotomy policy is used by default.	Data type: string or python instance Default value: BinarySearch
sensitivity	Input	Metric used to evaluate how quantization-sensible each layer to be quantized is. By default, the cosine similarity metric is used.	Data type: string or python instance Default value: CosineSimilarity

Returns

None

Outputs

A .pb model file that serves for both accuracy simulation in the TensorFlow environment and inference deployment on Ascend AI Processor.
Quantization factor record file, quantization configuration file, layer similarity file, and automatic quantization rollback history file.

Examples

import amct_tensorflow as amct
from amct_tensorflow.accuracy_based_auto_calibration import accuracy_based_auto_calibration

def main():
    args_check(args)
    outputs = [PREDICTIONS]
    record_file = os.path.join(RESULT_DIR, 'record.txt')
    config_file = os.path.join(RESULT_DIR, 'config.json')
    with tf.io.gfile.GFile(args.model, mode='rb') as model:
        graph_def = tf.compat.v1.GraphDef()
        graph_def.ParseFromString(model.read())
    tf.import_graph_def(graph_def, name='')
    graph = tf.compat.v1.get_default_graph()
    amct.create_quant_config(config_file, graph)
    save_dir = os.path.join(RESULT_DIR, 'MobileNetV2')
    evaluator = MobileNetV2Evaluator(args.dataset, args.keyword, args.num_parallel_reads, args.batch_size)
    accuracy_based_auto_calibration(args.model, outputs, record_file, config_file, save_dir, evaluator)

Parent topic: PTQ APIs