How to access model.tar file before sagemaker uploads it to the output path?

0

I have a training step (sample code below) in my sagemaker pipeline. once the training is done, anything inside the /opt/ml/model file is zipped up into model.tar file and sagemaker uploads it to an s3 location. can one get access this model.tar file before sagemaker uploads it to the s3 bucket. say from my training step, i wanted to access model.tar file , once the training is done and before sagemaker uploads it to the s3 output location . is it saved locally in the training instance , before uploading it to the s3 location ? if this is not possible, can i define another processing step to run after the trainign step , to download this model, wherever sagemaker saved it ( s3 uri) . to do this, can i define a processing step, such that if i give it a s3 location of the model, sagemaker will automatically download the model , or do i need to write code to download the model?


from sagemaker.estimator import Estimator
from sagemaker.inputs import TrainingInput

xgb_train = Estimator(
    image_uri="some_uri",
    instance_type=instance_type,
    instance_count=1,
    output_path=model_path,
    role=role,
    sagemaker_session=pipeline_session,
)

#training code 
train_args = xgb_train.fit(
    inputs={
        "train": TrainingInput(
            s3_data=step_process.properties.ProcessingOutputConfig.Outputs["train"].S3Output.S3Uri,
            content_type="text/csv",
        ),
        "validation": TrainingInput(
            s3_data=step_process.properties.ProcessingOutputConfig.Outputs[
                "validation"
            ].S3Output.S3Uri,
            content_type="text/csv",
        ),
    }
)

  • Can you share what you are planning to do once you access the tar file? You can write a script to download the S3 model in a processing step, but it will download it to the processing instance, which is also ephemeral.

質問済み 1年前322ビュー
1回答
0

You can try to do something like this or copy object as the download may happen on training instance

s3 = boto3.client('s3') s3.download_file('BUCKET_NAME', 'OBJECT_NAME', 'FILE_NAME')

profile pictureAWS
エキスパート
回答済み 1年前

ログインしていません。 ログイン 回答を投稿する。

優れた回答とは、質問に明確に答え、建設的なフィードバックを提供し、質問者の専門分野におけるスキルの向上を促すものです。

質問に答えるためのガイドライン

関連するコンテンツ