aup.Proposer.HyperbandProposer

The code is based on hyperband github commit a632209.

See license for redistribution.

Configuration

General parameters

Name	Default value	Explanation
proposer		hyperband
random_seed	0	[Optional] seed for random generator
max_iter	81	Max iterations (e.g. epochs) per configuration
eta	3	downsampling rate, choose 3 for training from scratch
skip_last	0	whether skip last element
engine	random	engine to generate configurations for hyperband

Specific parameters for parameter_config

Name	Explanation
name	name of the variable, will be used in the job config, i.e. training code
type	type of the parameter to be sampled: choose from “float”,”int”,”choice”
range	range of the parameter. For “choice”, list all the feasible values

In the returned job_config, it has n_iterations and tid for tracking the HPO process. n_iterations marks how much resource should be allocated for this round of training (e.g. epochs); tid is used to recover the previous trained model if needed (i.e. finetune)

APIs

class HyperbandProposer(config)

Bases: aup.Proposer.AbstractProposer.AbstractProposer

failed(job)

Mark job as failed in proposer history.

Parameters

job (Job) – Failed job

get_param()

Get new proposed parameter values

reload(path)

Reload Proposer state from path

Parameters

path (str) – path to reload

save(path)

Save Proposer state to path.

Some proposer can not generate new parameters after saving.

Parameters

path (str) – path to save

setup(s)

static setup_config()

update(score, job)

Update scores in proposer history

Parameters

• score (float) – score returned by Job

• job (Job) – Finished job

verify_config(config)

Verify the input configuration is enough for the proposer

Parameters

config (dict) – Experiment configuration of parameter_config

Returns

config

Return type

dict