【Azure AutoML】Python SDK V1 / V2で学習モデルのmetricsを取得する方法

投稿日 2023年6月27日
更新日 2023年6月28日
著者 azure-recipe-user
カテゴリー AutoML
カテゴリー Azure
カテゴリー Python

はじめに

Azure AutoMLでトレーニングしたモデルのmetricsをPython SDK v1 / v2の2パターンで取得してみます。

モデル自体はSDK V2でトレーニングしたものを使います。
トレーニングについては過去の記事にあるので、ぜひご参照ください。

【タスク】

環境

OS Windows 10(NVIDIA GTX 1650Ti,16GB RAM, i5-10300H CPU)
Visual Studio Code 1.73.1
Python 3.8 / 3.9
AzureML Python SDK v1 / v2

Python SDK V1のmetrics取得する方法

SDK V2でトレーニングしたモデルからmetricsを取得します。

V1の場合は、get_metrics関数を使うと簡単にできます。

get_metrics_v1.py
from azureml.core import Workspace
from azureml.train.automl.run import AutoMLRun
import pandas as pd

ws = Workspace.from_config()

experiment = ws.experiments['&lt;実験名&gt;']
automl_run = AutoMLRun(experiment, run_id = '&lt;ジョブID&gt;')
best_model, _ = automl_run.get_output()
print(best_model.get_metrics())

				
					
				1
2
3
4
5
6
7
8
9
10
11
12

						get_metrics_v1.py
from azureml.core import Workspace
from azureml.train.automl.run import AutoMLRun
import pandas as pd
 
ws = Workspace.from_config()
 
experiment = ws.experiments['&lt;実験名&gt;']
automl_run = AutoMLRun(experiment, run_id = '&lt;ジョブID&gt;')
best_model, _ = automl_run.get_output()
print(best_model.get_metrics())
 

					

			

結果

V2で作ったモデルを無理やりV1でmetrics取得しようとしているせいか、Warningがめちゃくちゃ出ますが、次のようにmetricsが色々返ってきました。

{'mean_absolute_percentage_error': 17.41204188234799, 'explained_variance': 0.738272704815586, 'mean_absolute_error': 2126.4896950718503, 'normalized_median_absolute_error': 0.03536353076485475, 'r2_score': 0.719627939723515, 'root_mean_squared_error': 2978.0824945509294, 'median_absolute_error': 1424.513746269879, 'normalized_root_mean_squared_log_error': 0.10514941989299778, 'normalized_root_mean_squared_error': 0.0739308498721744, 'root_mean_squared_log_error': 0.22949649762141444, 'normalized_mean_absolute_error': 0.05279007236661164, 'spearman_correlation': 0.8992768077701008, 'predicted_true': 'aml://artifactId/ExperimentRun/dcid.gray_iron_drsmn87sp0_0/predicted_true', 'residuals': 'aml://artifactId/ExperimentRun/dcid.gray_iron_drsmn87sp0_0/residuals'}
Exception ignored in: &lt;function _Win32Helper.__del__ at 0x000002671874F4C0&gt;
Traceback (most recent call last):
  File "C:\Users\anaconda3\envs\azml38v1\lib\site-packages\azureml\automl\runtime\shared\win32_helper.py", line 246, in __del__
TypeError: catching classes that do not inherit from BaseException is not allowed

				
					
				1
2
3
4
5
6

						{'mean_absolute_percentage_error': 17.41204188234799, 'explained_variance': 0.738272704815586, 'mean_absolute_error': 2126.4896950718503, 'normalized_median_absolute_error': 0.03536353076485475, 'r2_score': 0.719627939723515, 'root_mean_squared_error': 2978.0824945509294, 'median_absolute_error': 1424.513746269879, 'normalized_root_mean_squared_log_error': 0.10514941989299778, 'normalized_root_mean_squared_error': 0.0739308498721744, 'root_mean_squared_log_error': 0.22949649762141444, 'normalized_mean_absolute_error': 0.05279007236661164, 'spearman_correlation': 0.8992768077701008, 'predicted_true': 'aml://artifactId/ExperimentRun/dcid.gray_iron_drsmn87sp0_0/predicted_true', 'residuals': 'aml://artifactId/ExperimentRun/dcid.gray_iron_drsmn87sp0_0/residuals'}
Exception ignored in: &lt;function _Win32Helper.__del__ at 0x000002671874F4C0&gt;
Traceback (most recent call last):
  File "C:\Users\anaconda3\envs\azml38v1\lib\site-packages\azureml\automl\runtime\shared\win32_helper.py", line 246, in __del__
TypeError: catching classes that do not inherit from BaseException is not allowed
 

					

			

最後にTypeErrorも出ますが、なんとかmetricsを読み込むことはできました！！

続いて、素直にV2からmetricsを取得してみます。

Python SDK V2の場合

SDK V2の場合はMLFlowを使います。
参考サイト：

azureml-examples/sdk/python/jobs/automl-standalone-jobs/automl-forecasting-github-dau/auto-ml-forecasting-github-dau.ipynb at main · Azure/azureml-examples

Official community-driven Azure Machine Learning examples, tested with GitHub Actions. – azureml-examples/sdk/python/jobs/automl-standalone-jobs/automl-forecas…

github.com

まずはmlflowとazureml-mlflowをインストールします。

pip install mlflow azureml-mlflow

				1
2

						pip install mlflow azureml-mlflow

最初にワークスペースへの接続をし、MLFlowで実験を追跡するためにトラッキングURLを取得します。
その次にトレーニングしたジョブ（親ジョブ）からベストモデルを取得し、metricsを得るという流れになります。

get_metrics_v2.py
import mlflow
from azure.identity import DefaultAzureCredential
from azure.ai.ml import MLClient
from mlflow.tracking.client import MlflowClient
import pandas as pd

# workspace
credential = DefaultAzureCredential(exclude_shared_token_cache_credential=True)
subscription_id="&lt;サブスクリプションID&gt;"
resource_group="&lt;リソースグループ&gt;"
workspace_name="&lt;ワークスペース名&gt;"

ml_client = MLClient(credential, subscription_id, resource_group, workspace_name)

# Obtain the tracking URL from MLClient
MLFLOW_TRACKING_URI = ml_client.workspaces.get(
    name=ml_client.workspace_name
).mlflow_tracking_uri
print(MLFLOW_TRACKING_URI)

mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
print("\nCurrent tracking uri: {}".format(mlflow.get_tracking_uri()))

# Initialize MLFlow client
mlflow_client = MlflowClient()

job_name='&lt;ジョブID&gt;'

# Get the parent run
mlflow_parent_run = mlflow_client.get_run(job_name)

print("Parent Run: ")
print(mlflow_parent_run)
print(mlflow_parent_run.data.tags)

best_child_run_id = mlflow_parent_run.data.tags["automl_best_child_run_id"]
print("Found best child run id: ", best_child_run_id)

best_run = mlflow_client.get_run(best_child_run_id)

print("Best child run: ")
print(best_run)

				
					
				1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

						get_metrics_v2.py
import mlflow
from azure.identity import DefaultAzureCredential
from azure.ai.ml import MLClient
from mlflow.tracking.client import MlflowClient
import pandas as pd
 
# workspace
credential = DefaultAzureCredential(exclude_shared_token_cache_credential=True)
subscription_id="&lt;サブスクリプションID&gt;"
resource_group="&lt;リソースグループ&gt;"
workspace_name="&lt;ワークスペース名&gt;"
 
ml_client = MLClient(credential, subscription_id, resource_group, workspace_name)
 
# Obtain the tracking URL from MLClient
MLFLOW_TRACKING_URI = ml_client.workspaces.get(
    name=ml_client.workspace_name
).mlflow_tracking_uri
print(MLFLOW_TRACKING_URI)
 
mlflow.set_tracking_uri(MLFLOW_TRACKING_URI)
print("\nCurrent tracking uri: {}".format(mlflow.get_tracking_uri()))
 
# Initialize MLFlow client
mlflow_client = MlflowClient()
 
job_name='&lt;ジョブID&gt;'
 
# Get the parent run
mlflow_parent_run = mlflow_client.get_run(job_name)
 
print("Parent Run: ")
print(mlflow_parent_run)
print(mlflow_parent_run.data.tags)
 
best_child_run_id = mlflow_parent_run.data.tags["automl_best_child_run_id"]
print("Found best child run id: ", best_child_run_id)
 
best_run = mlflow_client.get_run(best_child_run_id)
 
print("Best child run: ")
print(best_run)
 

					

			

ワークスペースへの接続・ジョブIDを入力するだけで実行できます。

結果

親ジョブのmetrics・ベストモデルのmetricsが返ってきました。

Parent Run:
&lt;Run: data=&lt;RunData: metrics={'AUC_macro': 0.5537367792751098,
 'AUC_micro': 0.8053648205456106,
 'AUC_weighted': 0.5537367792751098,
 'accuracy': 0.7873917952578096,
 'average_precision_score_macro': 0.5301345560931308,
 'average_precision_score_micro': 0.7764476136917231,
 'average_precision_score_weighted': 0.6946791650615354,
 'balanced_accuracy': 0.5,
 'f1_score_macro': 0.44051441816730164,
 'f1_score_micro': 0.7873917952578096,
 'f1_score_weighted': 0.6937547541810158,
 'log_loss': 0.5174726757985704,
 'matthews_correlation': 0.0,
 'norm_macro_recall': 0.0,
 'precision_score_macro': 0.3936958976289048,
 'precision_score_micro': 0.7873917952578096,
 'precision_score_weighted': 0.6200497715962933,
 'recall_score_macro': 0.5,
 'recall_score_micro': 0.7873917952578096,
 'recall_score_weighted': 0.7873917952578096,
 'weighted_accuracy': 0.9318802438597572}, params={}, tags={'automl_best_child_run_id': 'hungry_chain_dyhsn2rnyj_3',
 'fit_time_000': '0.5580559999999999;4.332179999999999;3;8',
 'iteration_000': '0;1;2;3',
 'mlflow.rootRunId': 'hungry_chain_dyhsn2rnyj',
 'mlflow.runName': 'hungry_chain_dyhsn2rnyj',
 'mlflow.user': '',
 'model_explain_best_run_child_id': 'hungry_chain_dyhsn2rnyj_3',
 'model_explain_run': 'best_run',
 'predicted_cost_000': '0;0;0;0',
 'run_algorithm_000': 'LightGBM;XGBoostClassifier;VotingEnsemble;StackEnsemble',
 'run_preprocessor_000': 'MaxAbsScaler;MaxAbsScaler;;',
 'score_000': '0.7861874294316898;0.7775310500564546;0.7863379751599548;0.7873917952578096',
 'training_percent_000': '100;100;100;100'}&gt;, info=&lt;RunInfo: artifact_uri='azureml://japaneast.api.azureml.ms/mlflow/v2.0/subscriptions/XXX/resourceGroups/XXX/providers/Microsoft.MachineLearningServices/workspaces/XXX/experiments/febdb527-437a-4d38-8d26-7cb8528691ae/runs/hungry_chain_dyhsn2rnyj/artifacts', end_time=1683513330904, experiment_id='febdb527-437a-4d38-8d26-7cb8528691ae', lifecycle_stage='active', run_id='hungry_chain_dyhsn2rnyj', run_name='hungry_chain_dyhsn2rnyj', run_uuid='hungry_chain_dyhsn2rnyj', start_time=1683512458707, status='FINISHED', user_id='XXX'&gt;&gt;
{'model_explain_run': 'best_run', 'score_000': '0.7861874294316898;0.7775310500564546;0.7863379751599548;0.7873917952578096', 'predicted_cost_000': '0;0;0;0', 'fit_time_000': '0.5580559999999999;4.332179999999999;3;8', 'training_percent_000': '100;100;100;100', 'iteration_000': '0;1;2;3', 'run_preprocessor_000': 'MaxAbsScaler;MaxAbsScaler;;', 'run_algorithm_000': 'LightGBM;XGBoostClassifier;VotingEnsemble;StackEnsemble', 'automl_best_child_run_id': 'hungry_chain_dyhsn2rnyj_3', 'model_explain_best_run_child_id': 'hungry_chain_dyhsn2rnyj_3', 'mlflow.rootRunId': 'hungry_chain_dyhsn2rnyj', 'mlflow.runName': 'hungry_chain_dyhsn2rnyj', 'mlflow.user': 'XXX'}
Found best child run id:  hungry_chain_dyhsn2rnyj_3
Best child run:
&lt;Run: data=&lt;RunData: metrics={'AUC_macro': 0.5537367792751098,
 'AUC_micro': 0.8053648205456106,
 'AUC_weighted': 0.5537367792751098,
 'accuracy': 0.7873917952578096,
 'average_precision_score_macro': 0.5301345560931308,
 'average_precision_score_micro': 0.7764476136917231,
 'average_precision_score_weighted': 0.6946791650615354,
 'balanced_accuracy': 0.5,
 'f1_score_macro': 0.44051441816730164,
 'f1_score_micro': 0.7873917952578096,
 'f1_score_weighted': 0.6937547541810158,
 'log_loss': 0.5174726757985704,
 'matthews_correlation': 0.0,
 'norm_macro_recall': 0.0,
 'precision_score_macro': 0.3936958976289048,
 'precision_score_micro': 0.7873917952578096,
 'precision_score_weighted': 0.6200497715962933,
 'recall_score_macro': 0.5,
 'recall_score_micro': 0.7873917952578096,
 'recall_score_weighted': 0.7873917952578096,
 'weighted_accuracy': 0.9318802438597572}, params={}, tags={'mlflow.parentRunId': 'hungry_chain_dyhsn2rnyj',
 'mlflow.rootRunId': 'hungry_chain_dyhsn2rnyj',
 'mlflow.runName': 'kind_snake_g0p4zj5j',
 'mlflow.source.name': 'automl_driver.py',
 'mlflow.source.type': 'JOB',
 'mlflow.user': 'XXX',
 'model_explain_run_id': 'hungry_chain_dyhsn2rnyj_ModelExplain',
 'model_explanation': 'True'}&gt;, info=&lt;RunInfo: artifact_uri='azureml://japaneast.api.azureml.ms/mlflow/v2.0/subscriptions/XXX/resourceGroups/XXX/providers/Microsoft.MachineLearningServices/workspaces/XXX/experiments/febdb527-437a-4d38-8d26-7cb8528691ae/runs/hungry_chain_dyhsn2rnyj_3/artifacts', end_time=1683513329701, experiment_id='febdb527-437a-4d38-8d26-7cb8528691ae', lifecycle_stage='active', run_id='hungry_chain_dyhsn2rnyj_3', run_name='kind_snake_g0p4zj5j', run_uuid='hungry_chain_dyhsn2rnyj_3', start_time=1683513268867, status='FINISHED', user_id='XXX'&gt;&gt;

				
					
				1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66

						Parent Run:
&lt;Run: data=&lt;RunData: metrics={'AUC_macro': 0.5537367792751098,
 'AUC_micro': 0.8053648205456106,
 'AUC_weighted': 0.5537367792751098,
 'accuracy': 0.7873917952578096,
 'average_precision_score_macro': 0.5301345560931308,
 'average_precision_score_micro': 0.7764476136917231,
 'average_precision_score_weighted': 0.6946791650615354,
 'balanced_accuracy': 0.5,
 'f1_score_macro': 0.44051441816730164,
 'f1_score_micro': 0.7873917952578096,
 'f1_score_weighted': 0.6937547541810158,
 'log_loss': 0.5174726757985704,
 'matthews_correlation': 0.0,
 'norm_macro_recall': 0.0,
 'precision_score_macro': 0.3936958976289048,
 'precision_score_micro': 0.7873917952578096,
 'precision_score_weighted': 0.6200497715962933,
 'recall_score_macro': 0.5,
 'recall_score_micro': 0.7873917952578096,
 'recall_score_weighted': 0.7873917952578096,
 'weighted_accuracy': 0.9318802438597572}, params={}, tags={'automl_best_child_run_id': 'hungry_chain_dyhsn2rnyj_3',
 'fit_time_000': '0.5580559999999999;4.332179999999999;3;8',
 'iteration_000': '0;1;2;3',
 'mlflow.rootRunId': 'hungry_chain_dyhsn2rnyj',
 'mlflow.runName': 'hungry_chain_dyhsn2rnyj',
 'mlflow.user': '',
 'model_explain_best_run_child_id': 'hungry_chain_dyhsn2rnyj_3',
 'model_explain_run': 'best_run',
 'predicted_cost_000': '0;0;0;0',
 'run_algorithm_000': 'LightGBM;XGBoostClassifier;VotingEnsemble;StackEnsemble',
 'run_preprocessor_000': 'MaxAbsScaler;MaxAbsScaler;;',
 'score_000': '0.7861874294316898;0.7775310500564546;0.7863379751599548;0.7873917952578096',
 'training_percent_000': '100;100;100;100'}&gt;, info=&lt;RunInfo: artifact_uri='azureml://japaneast.api.azureml.ms/mlflow/v2.0/subscriptions/XXX/resourceGroups/XXX/providers/Microsoft.MachineLearningServices/workspaces/XXX/experiments/febdb527-437a-4d38-8d26-7cb8528691ae/runs/hungry_chain_dyhsn2rnyj/artifacts', end_time=1683513330904, experiment_id='febdb527-437a-4d38-8d26-7cb8528691ae', lifecycle_stage='active', run_id='hungry_chain_dyhsn2rnyj', run_name='hungry_chain_dyhsn2rnyj', run_uuid='hungry_chain_dyhsn2rnyj', start_time=1683512458707, status='FINISHED', user_id='XXX'&gt;&gt;
{'model_explain_run': 'best_run', 'score_000': '0.7861874294316898;0.7775310500564546;0.7863379751599548;0.7873917952578096', 'predicted_cost_000': '0;0;0;0', 'fit_time_000': '0.5580559999999999;4.332179999999999;3;8', 'training_percent_000': '100;100;100;100', 'iteration_000': '0;1;2;3', 'run_preprocessor_000': 'MaxAbsScaler;MaxAbsScaler;;', 'run_algorithm_000': 'LightGBM;XGBoostClassifier;VotingEnsemble;StackEnsemble', 'automl_best_child_run_id': 'hungry_chain_dyhsn2rnyj_3', 'model_explain_best_run_child_id': 'hungry_chain_dyhsn2rnyj_3', 'mlflow.rootRunId': 'hungry_chain_dyhsn2rnyj', 'mlflow.runName': 'hungry_chain_dyhsn2rnyj', 'mlflow.user': 'XXX'}
Found best child run id:  hungry_chain_dyhsn2rnyj_3
Best child run:
&lt;Run: data=&lt;RunData: metrics={'AUC_macro': 0.5537367792751098,
 'AUC_micro': 0.8053648205456106,
 'AUC_weighted': 0.5537367792751098,
 'accuracy': 0.7873917952578096,
 'average_precision_score_macro': 0.5301345560931308,
 'average_precision_score_micro': 0.7764476136917231,
 'average_precision_score_weighted': 0.6946791650615354,
 'balanced_accuracy': 0.5,
 'f1_score_macro': 0.44051441816730164,
 'f1_score_micro': 0.7873917952578096,
 'f1_score_weighted': 0.6937547541810158,
 'log_loss': 0.5174726757985704,
 'matthews_correlation': 0.0,
 'norm_macro_recall': 0.0,
 'precision_score_macro': 0.3936958976289048,
 'precision_score_micro': 0.7873917952578096,
 'precision_score_weighted': 0.6200497715962933,
 'recall_score_macro': 0.5,
 'recall_score_micro': 0.7873917952578096,
 'recall_score_weighted': 0.7873917952578096,
 'weighted_accuracy': 0.9318802438597572}, params={}, tags={'mlflow.parentRunId': 'hungry_chain_dyhsn2rnyj',
 'mlflow.rootRunId': 'hungry_chain_dyhsn2rnyj',
 'mlflow.runName': 'kind_snake_g0p4zj5j',
 'mlflow.source.name': 'automl_driver.py',
 'mlflow.source.type': 'JOB',
 'mlflow.user': 'XXX',
 'model_explain_run_id': 'hungry_chain_dyhsn2rnyj_ModelExplain',
 'model_explanation': 'True'}&gt;, info=&lt;RunInfo: artifact_uri='azureml://japaneast.api.azureml.ms/mlflow/v2.0/subscriptions/XXX/resourceGroups/XXX/providers/Microsoft.MachineLearningServices/workspaces/XXX/experiments/febdb527-437a-4d38-8d26-7cb8528691ae/runs/hungry_chain_dyhsn2rnyj_3/artifacts', end_time=1683513329701, experiment_id='febdb527-437a-4d38-8d26-7cb8528691ae', lifecycle_stage='active', run_id='hungry_chain_dyhsn2rnyj_3', run_name='kind_snake_g0p4zj5j', run_uuid='hungry_chain_dyhsn2rnyj_3', start_time=1683513268867, status='FINISHED', user_id='XXX'&gt;&gt;
 

					

			

親ジョブ・ベストモデルのmetricsともに取得することができました。
ここからmetricsだけ取り出すには、下記をコードの最後に貼り付けます。

data = best_run.data.metrics
res = json.dumps(data)
print(res)

				1
2
3
4

						data = best_run.data.metrics
res = json.dumps(data)
print(res)

結果はこうなりました。

{
  "AUC_weighted": 0.5537367792751098,
  "AUC_micro": 0.8053648205456106,
  "average_precision_score_macro": 0.5301345560931308,
  "matthews_correlation": 0,
  "weighted_accuracy": 0.9318802438597572,
  "norm_macro_recall": 0,
  "log_loss": 0.5174726757985704,
  "f1_score_weighted": 0.6937547541810158,
  "balanced_accuracy": 0.5,
  "recall_score_macro": 0.5,
  "precision_score_weighted": 0.6200497715962933,
  "f1_score_macro": 0.44051441816730164,
  "average_precision_score_micro": 0.7764476136917231,
  "precision_score_macro": 0.3936958976289048,
  "f1_score_micro": 0.7873917952578096,
  "recall_score_weighted": 0.7873917952578096,
  "AUC_macro": 0.5537367792751098,
  "average_precision_score_weighted": 0.6946791650615354,
  "accuracy": 0.7873917952578096,
  "precision_score_micro": 0.7873917952578096,
  "recall_score_micro": 0.7873917952578096
}

				
					
				1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23

						{
  "AUC_weighted": 0.5537367792751098,
  "AUC_micro": 0.8053648205456106,
  "average_precision_score_macro": 0.5301345560931308,
  "matthews_correlation": 0,
  "weighted_accuracy": 0.9318802438597572,
  "norm_macro_recall": 0,
  "log_loss": 0.5174726757985704,
  "f1_score_weighted": 0.6937547541810158,
  "balanced_accuracy": 0.5,
  "recall_score_macro": 0.5,
  "precision_score_weighted": 0.6200497715962933,
  "f1_score_macro": 0.44051441816730164,
  "average_precision_score_micro": 0.7764476136917231,
  "precision_score_macro": 0.3936958976289048,
  "f1_score_micro": 0.7873917952578096,
  "recall_score_weighted": 0.7873917952578096,
  "AUC_macro": 0.5537367792751098,
  "average_precision_score_weighted": 0.6946791650615354,
  "accuracy": 0.7873917952578096,
  "precision_score_micro": 0.7873917952578096,
  "recall_score_micro": 0.7873917952578096
}

					

			

この記事を書いた人

azure-recipe-user

記事一覧

【Azure AutoML】Python SDK V1 / V2で学習モデルのmetricsを取得する方法

はじめに

環境

Python SDK V1のmetrics取得する方法

結果

Python SDK V2の場合

結果

この記事を書いた人

azure-recipe-user

Distributed Data Systems with Azure Databricksやーる

【AzureML】モデルカタログからWhisperを使ってみる

【実機なしでも出来る】HoloLens 2 チュートリアルの Azure Cloud Services やってみた

【Azure Cognitive Searchで始めるTwitter分析】Part3. Power BI レポートの作成