How should a custom SageMaker algorithm determine if checkpoints are enabled?

Question

Per the [SageMaker Environment Variables][1] doc, algorithms should save model artifacts to the folder prescribed by `SM_MODEL_DIR`.

The [SageMaker Containers doc][2] describes additional environment variables, including `SM_OUTPUT_DATA_DIR` to write non-model training artifacts.

...But how should the algorithm determine if checkpointing has been requested?

The [Using Checkpoints in Amazon SageMaker][3] doc only specifies a default local path to save them to, and I can't see any environment variables that would indicate whether or not to checkpoint. I've seen one piece of code checking for the **existence** of that default local path, but not convinced anybody's actually checked to see whether it works (is present when checkpointing is requested and absent when not).

It's good to parameterize Checkpointing to avoid wasting EBS space (and precious seconds of IO) in jobs when it's not needed; and by the conventions for other I/O like model and data folders I would assume SageMaker to have a specific mechanism to pass this instruction, rather than just defining an algo hyperparameter?

[1]: https://docs.aws.amazon.com/sagemaker/latest/dg/docker-container-environmental-variables-user-scripts.html
  [2]: https://github.com/aws/sagemaker-containers
  [3]: https://docs.aws.amazon.com/sagemaker/latest/dg/model-checkpoints.html

Accepted Answer

Hi,

For custom Sagemaker containers or  deep learning frameworks, I tend to do this..and it works

This example is for pytorch I have tried

-  entry point file: 
   
   ```python
   # 1. Define a custom argument, say checkpointdir
    parser.add_argument("--checkpointdir", help="The checkpoint dir", type=str,
                        default=None)
   # 2. You can additional params for checkpoint frequency etc

# 3. Code for checkpointing
   if checkpointdir is not None:
      #TODO: save mode
    
   ```

-   Sagemaker estimator in Jupyter notebook, for. e.g.

```python
# 1. Define local and remote variables for checkpoints
checkpoint_s3 = "s3://{}/{}t/".format(bucket, "checkpoints")
localcheckpoint_dir="/opt/ml/checkpoints/"

hyperparameters = {

"batchsize": "8",
    "epochs" : "1000",
    "learning_rate":.0001,
    "weight_decay":5e-5,
    "momentum":.9,
    "patience": 20,
    "log-level" : "INFO",
    "commit_id":commit_id,
    "model" :"FasterRcnnFactory",
    "accumulation_steps": 8,
# 2.  define hp for checkpoint dir
    "checkpointdir": localcheckpoint_dir
}

# In the Sagemaker estimator fit, specify the local and remote path
from sagemaker.pytorch import PyTorch

estimator = PyTorch(
     entry_point='experiment_train.py',
                    source_dir = 'src',
                    dependencies =['src/datasets', 'src/evaluators', 'src/models'],
                    role=role,
                    framework_version ="1.0.0",
                    py_version='py3',
                    git_config= git_config,
                    image_name= docker_repo,
                    train_instance_count=1,
                    train_instance_type=instance_type,
# 3. The entrypoint file will pick up the checkpoint location from here
                    hyperparameters =hyperparameters,
                    output_path=s3_output_path,
                    metric_definitions=metric_definitions,
                    train_use_spot_instances = use_spot,
                    train_max_run =  train_max_run_secs,
                    train_max_wait = max_wait_time_secs,   
                    base_job_name ="object-detection",
# 4. Sagemaker knows that the checkpoints will need to be periodically copied from the localcheckpoint_dir to s3 pointed to by checkpoint_s3
                    checkpoint_s3_uri=checkpoint_s3,
                    checkpoint_local_path=localcheckpoint_dir)
```

How should a custom SageMaker algorithm determine if checkpoints are enabled?

関連するコンテンツ