Skip to content

Latest commit

 

History

History
83 lines (54 loc) · 4.02 KB

sigopt_strategy.md

File metadata and controls

83 lines (54 loc) · 4.02 KB
Raw

SigOpt Strategy

  1. Introduction

    1.1 Preparation

    1.2 SigOpt Platform

    1.3 Neural Compressor Configuration

  2. Performance

    2.1 Benefit of SigOpt Strategy

    2.2 Performance Comparison of Different Strategies

Introduction

SigOpt is an online model development platform that makes it easy to track runs, visualize training, and scale hyperparameter optimization for any type of model. Optimization Loop is the backbone of using SigOpt. We can set metrics and realize the interaction between the online platform and tuning configurations based on this mechanism.

Preparation

Before using the SigOpt strategy, a SigOpt account is necessary.

  • Each account has its own API token. Find your API token and then fill it in the sigopt_api_token field.
  • Create a new project and fill the corresponding name into the sigopt_project_id field.
  • Set the name of this experiment in sigopt_experiment_id field optionally. The default name is "nc-tune".

SigOpt Platform

If you are using the SigOpt products for the first time, please sign-up, if not, please login. It is free to apply for an account. Although there are certain restrictions on the model parameters and the number of experiments created, it is sufficient for ordinary customers. If you want higher capacity, please contact [email protected].

After logging in, you can use the token api to connect the local code to the online platform, corresponding to sigopt_api_token. It can be obtained here.

SigOpt has two concepts: project and experiment. Create a project before experimenting, corresponding to sigopt_project_id and sigopt_experiment_name. Multiple experiments can be created on each project. After creating the experiment, SigOpt will execute three simple steps below in a loop:

  • Receive a Suggestion from SigOpt;
  • Evaluate your metrics;
  • Report an Observation to SigOpt;

In our built-in sigopt strategy, the metrics add accuracy as a constraint and optimize for latency.

Neural Compressor Configuration

Compare to Basic strategy, sigopt_api_token and sigopt_project_id is necessary for SigOpt strategy. Before using the strategy, it is required to create the project corresponding to sigopt_project_id in your account.

from neural_compressor.config import PostTrainingQuantConfig, TuningCriterion

conf = PostTrainingQuantConfig(
    tuning_criterion=TuningCriterion(
        strategy="sigopt",
        strategy_kwargs={
            "sigopt_api_token": "YOUR-ACCOUNT-API-TOKEN",
            "sigopt_project_id": "PROJECT-ID",
            "sigopt_experiment_name": "nc-tune",
        },
    ),
)

Performance

Benefit of SigOpt Strategy

  • Metric based SigOpt is better than self-defining and easy to use. You can read the details here.
  • With the token api, results of each experiment are recorded in your account. You can use the SigOpt data analysis function to analyze the results, such as drawing a chart, calculating the F1 score, etc.

Performance Comparison of Different Strategies

  • MobileNet_v1(tensorflow)

    strategy FP32 baseline int8 accuracy int8 duration(s)
    basic 0.8266 0.8372 88.2132
    sigopt 0.8266 0.8372 83.7495

  • ResNet50_v1(tensorflow)

    strategy FP32 baseline int8 accuracy int8 duration(s)
    basic 0.8299 0.8294 85.0837
    sigopt 0.8299 0.8291 83.4469