sagemaker/pr_2188/en/tutorials/sagemaker-sdk/training-sagemaker-sdk.md

Run training on Amazon SageMaker

This guide will show you how to train a 🤗 Transformers model with the HuggingFace SageMaker Python SDK. Learn how to:

• Install and setup your training environment.
• Prepare a training script.
• Create a Hugging Face Estimator.
• Run training with the fit method.
• Access your trained model.
• Perform distributed training.
• Create a spot instance.
• Load a training script from a GitHub repository.
• Collect training metrics.

Installation and setup

Before you can train a 🤗 Transformers model with SageMaker, you need to sign up for an AWS account. If you don't have an AWS account yet, learn more here.

Once you have an AWS account, get started using one of the following:

• SageMaker Studio
• SageMaker notebook instance
• Local environment

To start training locally, you need to setup an appropriate IAM role.

Upgrade to the latest sagemaker version:

pip install 'sagemaker [!WARNING]
> [SageMaker Python SDK v3 has been recently released](https://github.com/aws/sagemaker-python-sdk), so unless specified otherwise, all the documentation and tutorials are still using the [SageMaker Python SDK v2](https://github.com/aws/sagemaker-python-sdk/tree/master-v2). We are actively working on updating all the tutorials and examples, but in the meantime make sure to install the SageMaker SDK as `pip install "sagemaker 5GB), which can slow down deployment for Amazon SageMaker Inference. 
You can control how checkpoints, logs, and artifacts are saved by customization the [TrainingArguments](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments). For example by providing `save_total_limit` as `TrainingArgument` you can control the limit of the total amount of checkpoints. Deletes the older checkpoints in `output_dir` if new ones are saved and the maximum limit is reached.

In addition to the options already mentioned above, there is another option to save the training artifacts during the training session. Amazon SageMaker supports [Checkpointing](https://docs.aws.amazon.com/sagemaker/latest/dg/model-checkpoints.html), which allows you to continuously save your artifacts during training to Amazon S3 rather than at the end of your training. To enable [Checkpointing](https://docs.aws.amazon.com/sagemaker/latest/dg/model-checkpoints.html) you need to provide the `checkpoint_s3_uri` parameter pointing to an Amazon S3 location in the `HuggingFace` estimator and set `output_dir` to `/opt/ml/checkpoints`. 
_Note: If you set `output_dir` to `/opt/ml/checkpoints` make sure to call `trainer.save_model("/opt/ml/model")` or model.save_pretrained("/opt/ml/model")/`tokenizer.save_pretrained("/opt/ml/model")` at the end of your training to be able to deploy your model seamlessly to Amazon SageMaker for Inference._

## Create a Hugging Face Estimator

Run 🤗 Transformers training scripts on SageMaker by creating a [Hugging Face Estimator](https://sagemaker.readthedocs.io/en/stable/frameworks/huggingface/sagemaker.huggingface.html#huggingface-estimator). The Estimator handles end-to-end SageMaker training. There are several parameters you should define in the Estimator:

1. `entry_point` specifies which fine-tuning script to use.
2. `instance_type` specifies an Amazon instance to launch. Refer [here](https://aws.amazon.com/sagemaker/pricing/) for a complete list of instance types.
3. `hyperparameters` specifies training hyperparameters. View additional available hyperparameters in [train.py file](https://github.com/huggingface/notebooks/blob/main/sagemaker/01_getting_started_pytorch/scripts/train.py).

The following code sample shows how to train with a custom script `train.py` with three hyperparameters (`epochs`, `per_device_train_batch_size`, and `model_name_or_path`):

```python
from sagemaker.huggingface import HuggingFace

# hyperparameters which are passed to the training job
hyperparameters={'epochs': 1,
                 'per_device_train_batch_size': 32,
                 'model_name_or_path': 'distilbert-base-uncased'
                 }

# create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='train.py',
        source_dir='./scripts',
        instance_type='ml.g6.12xlarge',
        instance_count=1,
        role=role,
        transformers_version='4.26',
        pytorch_version='1.13',
        py_version='py39',
        hyperparameters = hyperparameters
)

If you are running a TrainingJob locally, define instance_type='local' or instance_type='local_gpu' for GPU usage. Note that this will not work with SageMaker Studio.

Execute training

Start your TrainingJob by calling fit on a Hugging Face Estimator. Specify your input training data in fit. The input training data can be a:

• S3 URI such as s3://my-bucket/my-training-data.
• FileSystemInput for Amazon Elastic File System or FSx for Lustre. See here for more details about using these file systems as input.

Call fit to begin training:

huggingface_estimator.fit(
  {'train': 's3://sagemaker-us-east-1-558105141721/samples/datasets/imdb/train',
   'test': 's3://sagemaker-us-east-1-558105141721/samples/datasets/imdb/test'}
)

SageMaker starts and manages all the required EC2 instances and initiates the TrainingJob by running:

/opt/conda/bin/python train.py --epochs 1 --model_name_or_path distilbert-base-uncased --per_device_train_batch_size 32

Access trained model

Once training is complete, you can access your model through the AWS console or download it directly from S3.

from sagemaker.s3 import S3Downloader

S3Downloader.download(
    s3_uri=huggingface_estimator.model_data, # S3 URI where the trained model is located
    local_path='.',                          # local path where *.targ.gz is saved
    sagemaker_session=sess                   # SageMaker session used for training the model
)

Distributed training

SageMaker provides two strategies for distributed training: data parallelism and model parallelism. Data parallelism splits a training set across several GPUs, while model parallelism splits a model across several GPUs.

Data parallelism

The Hugging Face Trainer supports SageMaker's data parallelism library. If your training script uses the Trainer API, you only need to define the distribution parameter in the Hugging Face Estimator:

# configuration for running training on smdistributed data parallel
distribution = {'smdistributed':{'dataparallel':{ 'enabled': True }}}

# create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='train.py',
        source_dir='./scripts',
        instance_type='ml.p3dn.24xlarge',
        instance_count=2,
        role=role,
        transformers_version='4.26.0',
        pytorch_version='1.13.1',
        py_version='py39',
        hyperparameters = hyperparameters,
        distribution = distribution
)

📓 Open the sagemaker-notebook.ipynb notebook for an example of how to run the data parallelism library with TensorFlow.

Model parallelism

The Hugging Face [Trainer] also supports SageMaker's model parallelism library. If your training script uses the Trainer API, you only need to define the distribution parameter in the Hugging Face Estimator (see here for more detailed information about using model parallelism):

# configuration for running training on smdistributed model parallel
mpi_options = {
    "enabled" : True,
    "processes_per_host" : 8
}

smp_options = {
    "enabled":True,
    "parameters": {
        "microbatches": 4,
        "placement_strategy": "spread",
        "pipeline": "interleaved",
        "optimize": "speed",
        "partitions": 4,
        "ddp": True,
    }
}

distribution={
    "smdistributed": {"modelparallel": smp_options},
    "mpi": mpi_options
}

 # create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='train.py',
        source_dir='./scripts',
        instance_type='ml.p3dn.24xlarge',
        instance_count=2,
        role=role,
        transformers_version='4.26.0',
        pytorch_version='1.13.1',
        py_version='py39',
        hyperparameters = hyperparameters,
        distribution = distribution
)

📓 Open the sagemaker-notebook.ipynb notebook for an example of how to run the model parallelism library.

Spot instances

The Hugging Face extension for the SageMaker Python SDK means we can benefit from fully-managed EC2 spot instances. This can help you save up to 90% of training costs!

Note: Unless your training job completes quickly, we recommend you use checkpointing with managed spot training. In this case, you need to define the checkpoint_s3_uri.

Set use_spot_instances=True and define your max_wait and max_run time in the Estimator to use spot instances:

# hyperparameters which are passed to the training job
hyperparameters={'epochs': 1,
                 'train_batch_size': 32,
                 'model_name':'distilbert-base-uncased',
                 'output_dir':'/opt/ml/checkpoints'
                 }

# create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='train.py',
        source_dir='./scripts',
        instance_type='ml.g6.12xlarge',
        instance_count=1,
        checkpoint_s3_uri=f's3://{sess.default_bucket()}/checkpoints'
        use_spot_instances=True,
        # max_wait should be equal to or greater than max_run in seconds
        max_wait=3600,
        max_run=1000,
        role=role,
        transformers_version='4.26',
        pytorch_version='1.13',
        py_version='py39',
        hyperparameters = hyperparameters
)

# Training seconds: 874
# Billable seconds: 262
# Managed Spot Training savings: 70.0%

📓 Open the sagemaker-notebook.ipynb notebook for an example of how to use spot instances.

Git repository

The Hugging Face Estimator can load a training script stored in a GitHub repository. Provide the relative path to the training script in entry_point and the relative path to the directory in source_dir.

If you are using git_config to run the 🤗 Transformers example scripts, you need to configure the correct 'branch' in transformers_version (e.g. if you use transformers_version='4.4.2 you have to use 'branch':'v4.4.2').

Tip: Save your model to S3 by setting output_dir=/opt/ml/model in the hyperparameter of your training script.

# configure git settings
git_config = {'repo': 'https://github.com/huggingface/transformers.git','branch': 'v4.4.2'} # v4.4.2 refers to the transformers_version you use in the estimator

 # create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='run_glue.py',
        source_dir='./examples/pytorch/text-classification',
        git_config=git_config,
        instance_type='ml.g6.12xlarge',
        instance_count=1,
        role=role,
        transformers_version='4.26',
        pytorch_version='1.13',
        py_version='py39',
        hyperparameters=hyperparameters
)

SageMaker metrics

SageMaker metrics automatically parses training job logs for metrics and sends them to CloudWatch. If you want SageMaker to parse the logs, you must specify the metric's name and a regular expression for SageMaker to use to find the metric.

# define metrics definitions
metric_definitions = [
    {"Name": "train_runtime", "Regex": "train_runtime.*=\D*(.*?)$"},
    {"Name": "eval_accuracy", "Regex": "eval_accuracy.*=\D*(.*?)$"},
    {"Name": "eval_loss", "Regex": "eval_loss.*=\D*(.*?)$"},
]

# create the Estimator
huggingface_estimator = HuggingFace(
        entry_point='train.py',
        source_dir='./scripts',
        instance_type='ml.g6.12xlarge',
        instance_count=1,
        role=role,
        transformers_version='4.26',
        pytorch_version='1.13',
        py_version='py39',
        metric_definitions=metric_definitions,
        hyperparameters = hyperparameters)

📓 Open the notebook for an example of how to capture metrics in SageMaker.