tf_example7

tf_example7 example

This example is used to demonstrate how to utilize Neural Compressor to enabling quantization and benchmark with python-flavor config.

1. Installation

pip install -r requirements.txt

2. Prepare Model

python train.py

3. Run Command

• Run quantization

python test.py --tune

• Run benchmark

python test.py --benchmark

4. Introduction

• Use python code to set necessary parameters.

    from neural_compressor import conf
    dataloader = {
        'dataset': {'dummy_v2': {'input_shape': [28, 28]}}
    }
    conf.evaluation.performance.dataloader = dataloader
    conf.quantization.calibration.dataloader = dataloader
    conf.evaluation.accuracy.dataloader = dataloader
    conf.tuning.accuracy_criterion.absolute = 0.9

• Run quantization and benchmark.

    from neural_compressor.experimental import Quantization, common
    quantizer = Quantization(conf)
    quantizer.model = common.Model("../models/frozen_graph.pb")
    quantizer.fit()
    
    from neural_compressor.experimental import Benchmark
    evaluator = Benchmark(conf)
    evaluator.model = common.Model("../models/frozen_graph.pb")
    evaluator('performance')