Amazon SageMaker
Developer Guide

The AWS Documentation website is getting a new look!
Try it now and let us know what you think. Switch to the new look >>

You can return to the original look by selecting English in the language selector above.

Incremental Training in Amazon SageMaker

Over time, you might find that a model generates inference that are not as good as they were in the past. With incremental training, you can use the artifacts from an existing model and use an expanded dataset to train a new model. Incremental training saves both time and resources.

Use incremental training to:

  • Train a new model using an expanded dataset that contains an underlying pattern that was not accounted for in the previous training and which resulted in poor model performance.

  • Use the model artifacts or a portion of the model artifacts from a popular publicly available model in a training job. You don't need to train a new model from scratch.

  • Resume a training job that was stopped.

  • Train several variants of a model, either with different hyperparameter settings or using different datasets.

For more information about training jobs, see Train a Model with Amazon SageMaker .

You can train incrementally using the Amazon SageMaker console or the Amazon SageMaker Python SDK.

Important

Only three built-in algorithms currently support incremental training: Object Detection Algorithm, Image Classification Algorithm, and Semantic Segmentation Algorithm.

Perform Incremental Training (Console)

To complete this procedure, you need:

  • The URL of the Amazon Simple Storage Service (Amazon S3) bucket where you've stored the training data.

  • The URL of the S3 bucket where you want to store the output of the job.

  • The Amazon Elastic Container Registry path where the training code is stored. For more information, see Common Parameters for Built-In Algorithms .

  • The URL of the S3 bucket where you've stored the model artifacts that you want to use in incremental training. To find the URL for the model artifacts, see the details page of the training job used to create the model. To find the details page, in the Amazon SageMaker console, choose Inference, choose Models, and then choose the model.

To restart a stopped training job, use the URL to the model artifacts that are stored in the details page as you would with a model or a completed training job.

To perform incremental training (console)

  1. Open the Amazon SageMaker console at https://console.aws.amazon.com/sagemaker.

  2. In the navigation pane, choose Training, then choose Training jobs.

  3. Choose Create training job.

  4. Provide a name for the training job. The name must be unique within an AWS Region in an AWS account. The training job name must have 1 to 63 characters. Valid characters: a-z, A-Z, 0-9, and . : + = @ _ % - (hyphen).

  5. Choose the algorithm that you want to use. For information about algorithms, see Use Amazon SageMaker Built-in Algorithms .

  6. (Optional) For Resource configuration, either leave the default values or increase the resource consumption to reduce computation time.

    1. (Optional) For Instance type, choose the ML compute instance type that you want to use. In most cases, ml.m4.xlarge is sufficient.

    2. For Instance count, use the default, 1.

    3. (Optional) For Additional volume per instance (GB), choose the size of the ML storage volume that you want to provision. In most cases, you can use the default, 1. If you are using a large dataset, use a larger size.

  7. Provide information about the input data for the training dataset.

    1. For Channel name, either leave the default (train) or enter a more meaningful name for the training dataset, such as expanded-training-dataset.

    2. For InputMode, choose File. For incremental training, you need to use file input mode.

    3. For S3 data distribution type, choose FullyReplicated. This causes each ML compute instance to use a full replicate of the expanded dataset when training incrementally.

    4. If the expanded dataset is uncompressed, set the Compression type to None. If the expanded dataset is compressed using Gzip, set it to Gzip.

    5. (Optional) If you are using File input mode, leave Content type empty. For Pipe input mode, specify the appropriate MIME type. Content type is the multipurpose internet mail extension (MIME) type of the data.

    6. For Record wrapper, if the dataset is saved in RecordIO format, choose RecordIO. If your dataset is not saved as a RecordIO formatted file, choose None.

    7. For S3 data type, if the dataset us stored as a single file, choose S3Prefix. If the dataset is stored as several files in a folder, choose Manifest.

    8. For S3 location, provide the URL to the path where you stored the expanded dataset.

    9. Choose Done.

  8. To use model artifacts in a training job, you need to add a new channel and provide the needed information about the model artifacts.

    1. For Input data configuration, choose Add channel.

    2. For Channel name, enter model to identify this channel as the source of the model artifacts.

    3. For InputMode, choose File. Model artifacts are stored as files.

    4. For S3 data distribution type, choose FullyReplicated. This indicates that each ML compute instance should use all of the model artifacts for training.

    5. For Compression type, choose None because we are using a model for the channel.

    6. Leave Content type empty. Content type is the multipurpose internet mail extension (MIME) type of the data. For model artifacts, we leave it empty.

    7. Set Record wrapper to None because model artifacts are not stored in RecordIO format.

    8. For S3 data type, if you are using a built-in algorithm or an algorithm that stores the model as a single file, choose S3Prefix. If you are using an algorithm that stores the model as several files, choose Manifest.

    9. For S3 location, provide the URL to the path where you stored the model artifacts. Typically, the model is stored with the name model.tar.gz. To find the URL for the model artifacts, in the navigation pane, choose Inference, then choose Models. From the list of models, choose a model to display its details page. The URL for the model artifacts is listed under Primary container .

    10. Choose Done.

  9. For Output data configuration, provide the following information:

    1. For S3 location, type the path to the S3 bucket where you want to store the output data.

    2. (Optional) For Encryption key, you can add your AWS Key Management Service (AWS KMS) encryption key to encrypt the output data at rest. Provide the key ID or its Amazon Resource Number (ARN). For more information, see KMS-Managed Encryption Keys.

  10. (Optional) For Tags, add one or more tags to the training job. A tag is metadata that you can define and assign to AWS resources. In this case, you can use tags to help you manage your training jobs. A tag consists of a key and a value, which you define. For example, you might want to create a tag with Project as a key and a value referring to a project that is related to the training job, such as Home value forecasts.

  11. Choose Create training job. Amazon SageMaker creates and runs training job.

After the training job has completed, the newly trained model artifacts are stored under the S3 output path that you provided in the Output data configuration field. To deploy the model to get predictions, see Step 6: Deploy the Model to Amazon SageMaker.

Perform Incremental Training (API)

This example shows how to use Amazon SageMaker APIs to train a model using the Amazon SageMaker image classification algorithm and the Caltech 256 Image Dataset, then train a new model using the first one. It uses Amazon S3 for input and output sources. Please see the incremental training sample notebook for more details on using incremental training.

Note

In this example we used the original datasets in the incremental training, however you can use different datasets, such as ones that contain newly added samples. Upload the new datasets to S3 and make adjustments to the data_channels variable used to train the new model.

Get an AWS Identity and Access Management (IAM) role that grants required permissions and initialize environment variables:

import sagemaker from sagemaker import get_execution_role role = get_execution_role() print(role) sess = sagemaker.Session() bucket=sess.default_bucket() print(bucket) prefix = 'ic-incr-training'

Get the training image for the image classification algorithm:

from sagemaker.amazon.amazon_estimator import get_image_uri training_image = get_image_uri(sess.boto_region_name, 'image-classification', repo_version="latest") #Display the training image print (training_image)

Download the training and validation datasets, then upload them to Amazon Simple Storage Service (Amazon S3):

import os import urllib.request import boto3 # Define a download function def download(url): filename = url.split("/")[-1] if not os.path.exists(filename): urllib.request.urlretrieve(url, filename) # Download the caltech-256 training and validation datasets download('http://data.mxnet.io/data/caltech-256/caltech-256-60-train.rec') download('http://data.mxnet.io/data/caltech-256/caltech-256-60-val.rec') # Create four channels: train, validation, train_lst, and validation_lst s3train = 's3://{}/{}/train/'.format(bucket, prefix) s3validation = 's3://{}/{}/validation/'.format(bucket, prefix) # Upload the first files to the train and validation channels !aws s3 cp caltech-256-60-train.rec $s3train --quiet !aws s3 cp caltech-256-60-val.rec $s3validation --quiet

Define the training hyperparameters:

# Define hyperparameters for the estimator hyperparams = { "num_layers": "18", "resize": "32", "num_training_samples": "50000", "num_classes": "10", "image_shape": "3,28,28", "mini_batch_size": "128", "epochs": "3", "learning_rate": "0.1", "lr_scheduler_step": "2,3", "lr_scheduler_factor": "0.1", "augmentation_type": "crop_color", "optimizer": "sgd", "momentum": "0.9", "weight_decay": "0.0001", "beta_1": "0.9", "beta_2": "0.999", "gamma": "0.9", "eps": "1e-8", "top_k": "5", "checkpoint_frequency": "1", "use_pretrained_model": "0", "model_prefix": "" }

Create an estimator object and train the first model using the training and validation datasets:

# Fit the base estimator s3_output_location = 's3://{}/{}/output'.format(bucket, prefix) ic = sagemaker.estimator.Estimator(training_image, role, train_instance_count=1, train_instance_type='ml.p2.xlarge', train_volume_size=50, train_max_run=360000, input_mode='File', output_path=s3_output_location, sagemaker_session=sess, hyperparameters=hyperparams) train_data = sagemaker.session.s3_input(s3train, distribution='FullyReplicated', content_type='application/x-recordio', s3_data_type='S3Prefix') validation_data = sagemaker.session.s3_input(s3validation, distribution='FullyReplicated', content_type='application/x-recordio', s3_data_type='S3Prefix') data_channels = {'train': train_data, 'validation': validation_data} ic.fit(inputs=data_channels, logs=True)

To use the model to incrementally train another model, create a new estimator object and use the model artifacts (ic.model_data, in this example) for the model_uri input argument:

# Given the base estimator, create a new one for incremental training incr_ic = sagemaker.estimator.Estimator(training_image, role, train_instance_count=1, train_instance_type='ml.p2.xlarge', train_volume_size=50, train_max_run=360000, input_mode='File', output_path=s3_output_location, sagemaker_session=sess, hyperparameters=hyperparams, model_uri=ic.model_data) # This parameter will ingest the previous job's model as a new channel incr_ic.fit(inputs=data_channels, logs=True)

After the training job has completed, the newly trained model artifacts are stored under the S3 output path that you provided in Output_path. To deploy the model to get predictions, see Step 6: Deploy the Model to Amazon SageMaker.