本文為英文版的機器翻譯版本,如內容有任何歧義或不一致之處,概以英文版為準。
Amazon SageMaker AI 會在您使用時自動產生譜系實體的圖形。您可以查詢此資料以回答各種問題。以下說明如何在適用於 Python 的 SDK 中查詢此資料。
如需如何在 Amazon SageMaker Studio 中檢視已註冊模型系列的詳細資訊,請參閱 在 Studio 中檢視模型系列詳細資訊。
您可以查詢歷程實體,以執行下列作業:
-
擷取建立模型時使用的所有資料集。
-
擷取建立端點時使用的所有工作。
-
擷取使用資料集的所有模型。
-
擷取使用模型的所有端點。
-
擷取從特定資料集衍生的端點。
-
擷取建立訓練工作的管道執行。
-
擷取實體之間的關係,以進行調查、治理和再現。
-
擷取使用成品的所有下游試用。
-
擷取所有使用成品的上游試用。
-
擷取使用所提供之 S3 URI 的成品清單。
-
擷取使用資料集成品的上游成品。
-
擷取使用資料集成品的下游成品。
-
擷取使用映像成品的資料集。
-
擷取使用內容的動作。
-
擷取使用端點的處理工作。
-
擷取使用端點的轉換工作。
-
擷取使用端點的試用元件。
-
擷取與模型套件群組相關聯之管道執行的 ARN。
-
擷取使用動作的所有成品。
-
擷取使用模型套件核准動作的所有上游資料集。
-
透過模型套件核准動作擷取模型套件。
-
擷取使用端點的下游端點內容。
-
擷取與試用元件相關聯之管道執行的 ARN。
-
擷取使用試用元件的資料集。
-
擷取使用試用元件的模型。
-
探索歷程以進行視覺化。
限制
-
下列區域無法使用歷程查詢:
-
非洲 (開普敦) – af-south
-
亞太區域 (雅加達) – ap-southeast-3
-
亞太區域 (大阪) - (ap-northeast-3)
-
歐洲 (米蘭) – eu-south-1
-
歐洲 (西班牙) – eu-south-2
-
以色列 (特拉維夫) – il-central-1
-
-
目前,關係探索的最大深度限制為 10。
-
篩選僅限於下列屬性:上次修改日期、建立日期、類型和歷程實體類型。
主題
開始查詢歷程實體
開始查詢歷程實體的最簡單方式是:
-
Amazon SageMaker AI SDK for Python
,已定義許多常見的使用案例。 -
如需示範如何使用 SageMaker AI Lineage APIs 查詢整個譜系圖表中關係的筆記本,請參閱 sagemaker-lineage-multihop-queries.ipynb
。
下列範例展示如何使用 LineageQuery
和 LineageFilter
API 建構查詢,以回答有關歷程圖的問題,並擷取一些使用案例中的實體關聯。
範例 使用 LineageQuery
API 尋找實體關聯
from sagemaker.lineage.context import Context, EndpointContext
from sagemaker.lineage.action import Action
from sagemaker.lineage.association import Association
from sagemaker.lineage.artifact import Artifact, ModelArtifact, DatasetArtifact
from sagemaker.lineage.query import (
LineageQuery,
LineageFilter,
LineageSourceEnum,
LineageEntityEnum,
LineageQueryDirectionEnum,
)
# Find the endpoint context and model artifact that should be used for the lineage queries.
contexts = Context.list(source_uri=endpoint_arn)
context_name = list(contexts)[0].context_name
endpoint_context = EndpointContext.load(context_name=context_name)
範例 尋找與某個端點相關聯的所有資料集
# Define the LineageFilter to look for entities of type `ARTIFACT` and the source of type `DATASET`.
query_filter = LineageFilter(
entities=[LineageEntityEnum.ARTIFACT], sources=[LineageSourceEnum.DATASET]
)
# Providing this `LineageFilter` to the `LineageQuery` constructs a query that traverses through the given context `endpoint_context`
# and find all datasets.
query_result = LineageQuery(sagemaker_session).query(
start_arns=[endpoint_context.context_arn],
query_filter=query_filter,
direction=LineageQueryDirectionEnum.ASCENDANTS,
include_edges=False,
)
# Parse through the query results to get the lineage objects corresponding to the datasets
dataset_artifacts = []
for vertex in query_result.vertices:
dataset_artifacts.append(vertex.to_lineage_object().source.source_uri)
pp.pprint(dataset_artifacts)
範例 尋找與某個端點相關聯的模型
# Define the LineageFilter to look for entities of type `ARTIFACT` and the source of type `MODEL`.
query_filter = LineageFilter(
entities=[LineageEntityEnum.ARTIFACT], sources=[LineageSourceEnum.MODEL]
)
# Providing this `LineageFilter` to the `LineageQuery` constructs a query that traverses through the given context `endpoint_context`
# and find all datasets.
query_result = LineageQuery(sagemaker_session).query(
start_arns=[endpoint_context.context_arn],
query_filter=query_filter,
direction=LineageQueryDirectionEnum.ASCENDANTS,
include_edges=False,
)
# Parse through the query results to get the lineage objects corresponding to the model
model_artifacts = []
for vertex in query_result.vertices:
model_artifacts.append(vertex.to_lineage_object().source.source_uri)
# The results of the `LineageQuery` API call return the ARN of the model deployed to the endpoint along with
# the S3 URI to the model.tar.gz file associated with the model
pp.pprint(model_artifacts)
範例 尋找與端點相關聯的試用元件
# Define the LineageFilter to look for entities of type `TRIAL_COMPONENT` and the source of type `TRAINING_JOB`.
query_filter = LineageFilter(
entities=[LineageEntityEnum.TRIAL_COMPONENT],
sources=[LineageSourceEnum.TRAINING_JOB],
)
# Providing this `LineageFilter` to the `LineageQuery` constructs a query that traverses through the given context `endpoint_context`
# and find all datasets.
query_result = LineageQuery(sagemaker_session).query(
start_arns=[endpoint_context.context_arn],
query_filter=query_filter,
direction=LineageQueryDirectionEnum.ASCENDANTS,
include_edges=False,
)
# Parse through the query results to get the ARNs of the training jobs associated with this Endpoint
trial_components = []
for vertex in query_result.vertices:
trial_components.append(vertex.arn)
pp.pprint(trial_components)
範例 變更歷程的焦點
LineageQuery
可以修改為具有不同的 start_arns
來變更歷程的焦點。此外,LineageFilter
可以採用多個來源和實體來擴充查詢的範圍。
我們在下面使用該模型作為歷程焦點,並找到與之相關聯的端點和資料集。
# Get the ModelArtifact
model_artifact_summary = list(Artifact.list(source_uri=model_package_arn))[0]
model_artifact = ModelArtifact.load(artifact_arn=model_artifact_summary.artifact_arn)
query_filter = LineageFilter(
entities=[LineageEntityEnum.ARTIFACT],
sources=[LineageSourceEnum.ENDPOINT, LineageSourceEnum.DATASET],
)
query_result = LineageQuery(sagemaker_session).query(
start_arns=[model_artifact.artifact_arn], # Model is the starting artifact
query_filter=query_filter,
# Find all the entities that descend from the model, i.e. the endpoint
direction=LineageQueryDirectionEnum.DESCENDANTS,
include_edges=False,
)
associations = []
for vertex in query_result.vertices:
associations.append(vertex.to_lineage_object().source.source_uri)
query_result = LineageQuery(sagemaker_session).query(
start_arns=[model_artifact.artifact_arn], # Model is the starting artifact
query_filter=query_filter,
# Find all the entities that ascend from the model, i.e. the datasets
direction=LineageQueryDirectionEnum.ASCENDANTS,
include_edges=False,
)
for vertex in query_result.vertices:
associations.append(vertex.to_lineage_object().source.source_uri)
pp.pprint(associations)
範例 是用 LineageQueryDirectionEnum.BOTH
尋找遞增與遞減關係
當方向設定為 BOTH
時,查詢會遍歷圖形,以尋找遞增和遞減關係。這種遍歷不僅在起始節點發生,還會在造訪的每個節點進行。例如,如果某個訓練工作執行兩次,而且訓練工作產生的兩個模型均部署到端點,則查詢結果的方向會設定為 BOTH
,以顯示兩個端點。這是因為模型訓練和部署是用了相同的映像。由於模型映像是相同的,因此 start_arn
和兩個端點都會出現在查詢結果中。
query_filter = LineageFilter(
entities=[LineageEntityEnum.ARTIFACT],
sources=[LineageSourceEnum.ENDPOINT, LineageSourceEnum.DATASET],
)
query_result = LineageQuery(sagemaker_session).query(
start_arns=[model_artifact.artifact_arn], # Model is the starting artifact
query_filter=query_filter,
# This specifies that the query should look for associations both ascending and descending for the start
direction=LineageQueryDirectionEnum.BOTH,
include_edges=False,
)
associations = []
for vertex in query_result.vertices:
associations.append(vertex.to_lineage_object().source.source_uri)
pp.pprint(associations)
範例 LineageQuery
中的方向 - ASCENDANTS
和 DESCENDANTS
要了解在歷程圖中的方向,可採取以下實體關係圖:資料集-> 訓練工作 -> 模型-> 端點
從模型到端點是遞減,從模型到資料集也是遞減。與此類似,從端點到模型是遞增。direction
參數可用來指定查詢應傳回 start_arns
中實體的遞減還是遞增實體。如果 start_arns
包含模型且方向為 DESCENDANTS
,則查詢會傳回端點。如果方向為 ASCENDANTS
,則查詢會傳回資料集。
# In this example, we'll look at the impact of specifying the direction as ASCENDANT or DESCENDANT in a `LineageQuery`.
query_filter = LineageFilter(
entities=[LineageEntityEnum.ARTIFACT],
sources=[
LineageSourceEnum.ENDPOINT,
LineageSourceEnum.MODEL,
LineageSourceEnum.DATASET,
LineageSourceEnum.TRAINING_JOB,
],
)
query_result = LineageQuery(sagemaker_session).query(
start_arns=[model_artifact.artifact_arn],
query_filter=query_filter,
direction=LineageQueryDirectionEnum.ASCENDANTS,
include_edges=False,
)
ascendant_artifacts = []
# The lineage entity returned for the Training Job is a TrialComponent which can't be converted to a
# lineage object using the method `to_lineage_object()` so we extract the TrialComponent ARN.
for vertex in query_result.vertices:
try:
ascendant_artifacts.append(vertex.to_lineage_object().source.source_uri)
except:
ascendant_artifacts.append(vertex.arn)
print("Ascendant artifacts : ")
pp.pprint(ascendant_artifacts)
query_result = LineageQuery(sagemaker_session).query(
start_arns=[model_artifact.artifact_arn],
query_filter=query_filter,
direction=LineageQueryDirectionEnum.DESCENDANTS,
include_edges=False,
)
descendant_artifacts = []
for vertex in query_result.vertices:
try:
descendant_artifacts.append(vertex.to_lineage_object().source.source_uri)
except:
# Handling TrialComponents.
descendant_artifacts.append(vertex.arn)
print("Descendant artifacts : ")
pp.pprint(descendant_artifacts)
範例 SDK 輔助函式讓歷程查詢變得更輕鬆
EndpointContext
、ModelArtifact
和 DatasetArtifact
類別都具有輔助函式,這些函式是 LineageQuery
API 上的包裝函式,可以讓某些歷程查詢變得更輕鬆。以下範例展示如何使用這些輔助函式。
# Find all the datasets associated with this endpoint
datasets = []
dataset_artifacts = endpoint_context.dataset_artifacts()
for dataset in dataset_artifacts:
datasets.append(dataset.source.source_uri)
print("Datasets : ", datasets)
# Find the training jobs associated with the endpoint
training_job_artifacts = endpoint_context.training_job_arns()
training_jobs = []
for training_job in training_job_artifacts:
training_jobs.append(training_job)
print("Training Jobs : ", training_jobs)
# Get the ARN for the pipeline execution associated with this endpoint (if any)
pipeline_executions = endpoint_context.pipeline_execution_arn()
if pipeline_executions:
for pipeline in pipelines_executions:
print(pipeline)
# Here we use the `ModelArtifact` class to find all the datasets and endpoints associated with the model
dataset_artifacts = model_artifact.dataset_artifacts()
endpoint_contexts = model_artifact.endpoint_contexts()
datasets = [dataset.source.source_uri for dataset in dataset_artifacts]
endpoints = [endpoint.source.source_uri for endpoint in endpoint_contexts]
print("Datasets associated with this model : ")
pp.pprint(datasets)
print("Endpoints associated with this model : ")
pp.pprint(endpoints)
# Here we use the `DatasetArtifact` class to find all the endpoints hosting models that were trained with a particular dataset
# Find the artifact associated with the dataset
dataset_artifact_arn = list(Artifact.list(source_uri=training_data))[0].artifact_arn
dataset_artifact = DatasetArtifact.load(artifact_arn=dataset_artifact_arn)
# Find the endpoints that used this training dataset
endpoint_contexts = dataset_artifact.endpoint_contexts()
endpoints = [endpoint.source.source_uri for endpoint in endpoint_contexts]
print("Endpoints associated with the training dataset {}".format(training_data))
pp.pprint(endpoints)
範例 取得歷程圖視覺化圖形
範例筆記本 visualizer.pyVisualizer
,能夠幫助歷程圖出圖。彩現查詢回應時,系統會顯示含有來自 StartArns
之歷程關係的圖形。從StartArns
開始,此視覺化圖形會顯示與 query_lineage
API 動作中傳回之其他歷程實體之間的關係。
# Graph APIs
# Here we use the boto3 `query_lineage` API to generate the query response to plot.
from visualizer import Visualizer
query_response = sm_client.query_lineage(
StartArns=[endpoint_context.context_arn], Direction="Ascendants", IncludeEdges=True
)
viz = Visualizer()
viz.render(query_response, "Endpoint")
query_response = sm_client.query_lineage(
StartArns=[model_artifact.artifact_arn], Direction="Ascendants", IncludeEdges=True
)
viz.render(query_response, "Model")