auto ML 获取验证结果,供后续处理使用

 在做深度机器学习实验的时候，经常需要先做超参数扫描，得到一批实验结果，然后将验证的结果拿出来进一步处理，或是与人工标定比较，手动计算一些关键指标如分类Precision和Recall，或是制作一些曲线。下面是从Azure上实验中取出验证（机器标定）结果的关键代码：

import tempfile

form pathlib import Path

from azureml.core import Dataset

from azureml.core.experiment import Experiment

from azureml.core.script_run_config importScriptrunConfig

form azureml.core.workspace import Workspace

from azureml.train.automl.run import AutoMLRun

automl_run_id="automl_run_id"

dataset_name="dataset_validation"

experiment_name="automl_experiment_name"

csv_out="data/valid.csv"

computer_target="gpu_cluster"

validate_score=True

batch_size=16

resize_size=640

crop_size=480

workspace=Workspace.from_config()

experiment=Experiment(workspace,experiment_name)

computes=workspace.compute_targets

if computer_target in computes and computes[computer_target].type=="AmlCompute":

        print(f"Found existing compute target {computer_target}.")

        compute_target=computes[computer_target]

else:

        raise ValueError(

                f"AML cluster:{computer_target} doesn't exist. Create cluster first!"

        )

parent_run=AutoMLRun(experiment,automl_run_id)

target_run=parent_run.get_best_child()

print(f"AutoML Run {target_run.id} will be used for inference")

dataset=Dataset.get_by_name(workspace,name=dataset_name)

arguments=[

        target_run.id,

        experiment,name,

        dataset.id,

        validate_score,

        batch_size,

        resize_size,

        crop_size,

]

output_pred_file="outputs/pred.jsonl"

scoring_args=arguments+["--output_file",output_pred_file]

with tempfile.TemporaryDirectory() as tmpdir:

        entry_script_name="score_script.py"

        target_run.download_file(

                "train_artifacts/"+entry_script_name,

                Path(tmpdir,entry_script_name),

        )

        script_run_config=ScriptRunConfig(

                source_directory=tmpdir,

                script=entry_script_name,

                compute_target=compute_target,

                environment=target_run.get_environment(),

                arguments=scoring_args,

        )

        scoring_run=experement.submit(script_run_config)

        print("Job is submitted. This script will wait until the remote job is finished.")

        scoring_run.wait_for_completion(show_output=True,wait_post_processing=True)

        output_dataset_prefix="output_"

        output_dataset_name=output_dataset_prefix+scoring_run.id

        output_dataset=Dataset.get_by_name(workspace,name=output_dataset_name)

        output_dataframe=output_dataset.to_pandas_dataframe()