Configure SageMaker Debugger to Save Tensors - Amazon SageMaker
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Configure SageMaker Debugger to Save Tensors

Tensors are data collections of updated parameters from the backward and forward pass of each training iteration. SageMaker Debugger collects the output tensors to analyze the state of a training job. SageMaker Debugger's CollectionConfig and DebuggerHookConfig API operations provide methods for grouping tensors into collections and saving them to a target S3 bucket.

Note

After properly configured and activated, SageMaker Debugger saves the output tensors in a default S3 bucket, unless otherwise specified. The format of the default S3 bucket URI is s3://sagemaker-<region>-<12digit_account_id>/<training-job-name>/debug-output/.

While constructing a SageMaker estimator, activate SageMaker Debugger by specifying the debugger_hook_config parameter. The following steps include examples of how to set up the debugger_hook_config using the CollectionConfig and DebuggerHookConfig API operations to pull tensors out of your training jobs and save them.

Configure Tensor Collections Using the CollectionConfig API

Use the CollectionConfig API operation to configure tensor collections. Debugger provides pre-built tensor collections that cover a variety of regular expressions (regex) of parameters if using Debugger-supported deep learning frameworks and machine learning algorithms. As shown in the following example code, add the built-in tensor collections you want to debug.

from sagemaker.debugger import CollectionConfig collection_configs=[ CollectionConfig(name="weights"), CollectionConfig(name="gradients") ]

The preceding collections set up the Debugger hook to save the tensors every 500 steps based on the default "save_interval" value.

For a full list of available Debugger built-in collections, see Debugger Built-in Collections.

If you want to customize the built-in collections, such as changing the save intervals and tensor regex, use the following CollectionConfig template to adjust parameters.

from sagemaker.debugger import CollectionConfig collection_configs=[ CollectionConfig( name="tensor_collection", parameters={ "key_1": "value_1", "key_2": "value_2", ... "key_n": "value_n" } ) ]

For more information about available parameter keys, see CollectionConfig in the Amazon SageMaker Python SDK. For example, the following code example shows how you can adjust the save intervals of the "losses" tensor collection at different phases of training: save loss every 100 steps in training phase and validation loss every 10 steps in validation phase.

from sagemaker.debugger import CollectionConfig collection_configs=[ CollectionConfig( name="losses", parameters={ "train.save_interval": "100", "eval.save_interval": "10" } ) ]
Tip

This tensor collection configuration object can be used for both DebuggerHookConfig and Rule API operations.

Configure the DebuggerHookConfig API to Save Tensors

Use the DebuggerHookConfig API to create a debugger_hook_config object using the collection_configs object you created in the previous step.

from sagemaker.debugger import DebuggerHookConfig debugger_hook_config=DebuggerHookConfig( collection_configs=collection_configs )

Debugger saves the model training output tensors into the default S3 bucket. The format of the default S3 bucket URI is s3://sagemaker-<region>-<12digit_account_id>/<training-job-name>/debug-output/.

If you want to specify an exact S3 bucket URI, use the following code example:

from sagemaker.debugger import DebuggerHookConfig debugger_hook_config=DebuggerHookConfig( s3_output_path="specify-your-s3-bucket-uri" collection_configs=collection_configs )

For more information, see DebuggerHookConfig in the Amazon SageMaker Python SDK.

Example Notebooks and Code Samples to Configure Debugger Hook

The following sections provide notebooks and code examples of how to use Debugger hook to save, access, and visualize output tensors.

Tensor Visualization Example Notebooks

The following two notebook examples show advanced use of Amazon SageMaker Debugger for visualizing tensors. Debugger provides a transparent view into training deep learning models.

  • Interactive Tensor Analysis in SageMaker Studio Notebook with MXNet

    This notebook example shows how to visualize saved tensors using Amazon SageMaker Debugger. By visualizing the tensors, you can see how the tensor values change while training deep learning algorithms. This notebook includes a training job with a poorly configured neural network and uses Amazon SageMaker Debugger to aggregate and analyze tensors, including gradients, activation outputs, and weights. For example, the following plot shows the distribution of gradients of a convolutional layer that is suffering from a vanishing gradient problem.

    
                        A graph plotting the distribution of gradients of a convolutional
                            layer suffering from a vanishing gradient problem

    This notebook also illustrates how a good initial hyperparameter setting improves the training process by generating the same tensor distribution plots.

  • Visualizing and Debugging Tensors from MXNet Model Training

    This notebook example shows how to save and visualize tensors from an MXNet Gluon model training job using Amazon SageMaker Debugger. It illustrates that Debugger is set to save all tensors to an Amazon S3 bucket and retrieves ReLu activation outputs for the visualization. The following figure shows a three-dimensional visualization of the ReLu activation outputs. The color scheme is set to blue to indicate values close to 0 and yellow to indicate values close to 1.

    
                        A visualization of the ReLU activation outputs

    In this notebook, the TensorPlot class imported from tensor_plot.py is designed to plot convolutional neural networks (CNNs) that take two-dimensional images for inputs. The tensor_plot.py script provided with the notebook retrieves tensors using Debugger and visualizes the CNN. You can run this notebook on SageMaker Studio to reproduce the tensor visualization and implement your own convolutional neural network model.

  • Real-time Tensor Analysis in a SageMaker Notebook with MXNet

    This example guides you through installing required components for emitting tensors in an Amazon SageMaker training job and using the Debugger API operations to access those tensors while training is running. A gluon CNN model is trained on the Fashion MNIST dataset. While the job is running, you will see how Debugger retrieves activation outputs of the first convolutional layer from each of 100 batches and visualizes them. Also, this will show you how to visualize weights after the job is done.

Save Tensors Using Debugger Built-in Collections

You can use built-in collections of tensors using the CollectionConfig API and save them using the DebuggerHookConfig API. The following example shows how to use the default settings of Debugger hook configurations to construct a SageMaker TensorFlow estimator. You can also utilize this for MXNet, PyTorch, and XGBoost estimators.

Note

In the following example code, the s3_output_path parameter for DebuggerHookConfig is optional. If you do not specify it, Debugger saves the tensors at s3://<output_path>/debug-output/, where the <output_path> is the default output path of SageMaker training jobs. For example:

"s3://sagemaker-us-east-1-111122223333/sagemaker-debugger-training-YYYY-MM-DD-HH-MM-SS-123/debug-output"
import sagemaker from sagemaker.tensorflow import TensorFlow from sagemaker.debugger import DebuggerHookConfig, CollectionConfig # use Debugger CollectionConfig to call built-in collections collection_configs=[ CollectionConfig(name="weights"), CollectionConfig(name="gradients"), CollectionConfig(name="losses"), CollectionConfig(name="biases") ] # configure Debugger hook # set a target S3 bucket as you want sagemaker_session=sagemaker.Session() BUCKET_NAME=sagemaker_session.default_bucket() LOCATION_IN_BUCKET='debugger-built-in-collections-hook' hook_config=DebuggerHookConfig( s3_output_path='s3://{BUCKET_NAME}/{LOCATION_IN_BUCKET}'. format(BUCKET_NAME=BUCKET_NAME, LOCATION_IN_BUCKET=LOCATION_IN_BUCKET), collection_configs=collection_configs ) # construct a SageMaker TensorFlow estimator sagemaker_estimator=TensorFlow( entry_point='directory/to/your_training_script.py', role=sm.get_execution_role(), base_job_name='debugger-demo-job', instance_count=1, instance_type="ml.p3.2xlarge", framework_version="2.9.0", py_version="py39", # debugger-specific hook argument below debugger_hook_config=hook_config ) sagemaker_estimator.fit()

To see a list of Debugger built-in collections, see Debugger Built-in Collections.

Save Tensors Using Debugger Modified Built-in Collections

You can modify the Debugger built-in collections using the CollectionConfig API operation. The following example shows how to tweak the built-in losses collection and construct a SageMaker TensorFlow estimator. You can also use this for MXNet, PyTorch, and XGBoost estimators.

import sagemaker from sagemaker.tensorflow import TensorFlow from sagemaker.debugger import DebuggerHookConfig, CollectionConfig # use Debugger CollectionConfig to call and modify built-in collections collection_configs=[ CollectionConfig( name="losses", parameters={"save_interval": "50"})] # configure Debugger hook # set a target S3 bucket as you want sagemaker_session=sagemaker.Session() BUCKET_NAME=sagemaker_session.default_bucket() LOCATION_IN_BUCKET='debugger-modified-collections-hook' hook_config=DebuggerHookConfig( s3_output_path='s3://{BUCKET_NAME}/{LOCATION_IN_BUCKET}'. format(BUCKET_NAME=BUCKET_NAME, LOCATION_IN_BUCKET=LOCATION_IN_BUCKET), collection_configs=collection_configs ) # construct a SageMaker TensorFlow estimator sagemaker_estimator=TensorFlow( entry_point='directory/to/your_training_script.py', role=sm.get_execution_role(), base_job_name='debugger-demo-job', instance_count=1, instance_type="ml.p3.2xlarge", framework_version="2.9.0", py_version="py39", # debugger-specific hook argument below debugger_hook_config=hook_config ) sagemaker_estimator.fit()

For a full list of CollectionConfig parameters, see Debugger CollectionConfig API.

Save Tensors Using Debugger Custom Collections

You can also save a reduced number of tensors instead of the full set of tensors (for example, if you want to reduce the amount of data saved in your Amazon S3 bucket). The following example shows how to customize the Debugger hook configuration to specify target tensors that you want to save. You can use this for TensorFlow, MXNet, PyTorch, and XGBoost estimators.

import sagemaker from sagemaker.tensorflow import TensorFlow from sagemaker.debugger import DebuggerHookConfig, CollectionConfig # use Debugger CollectionConfig to create a custom collection collection_configs=[ CollectionConfig( name="custom_activations_collection", parameters={ "include_regex": "relu|tanh", # Required "reductions": "mean,variance,max,abs_mean,abs_variance,abs_max" }) ] # configure Debugger hook # set a target S3 bucket as you want sagemaker_session=sagemaker.Session() BUCKET_NAME=sagemaker_session.default_bucket() LOCATION_IN_BUCKET='debugger-custom-collections-hook' hook_config=DebuggerHookConfig( s3_output_path='s3://{BUCKET_NAME}/{LOCATION_IN_BUCKET}'. format(BUCKET_NAME=BUCKET_NAME, LOCATION_IN_BUCKET=LOCATION_IN_BUCKET), collection_configs=collection_configs ) # construct a SageMaker TensorFlow estimator sagemaker_estimator=TensorFlow( entry_point='directory/to/your_training_script.py', role=sm.get_execution_role(), base_job_name='debugger-demo-job', instance_count=1, instance_type="ml.p3.2xlarge", framework_version="2.9.0", py_version="py39", # debugger-specific hook argument below debugger_hook_config=hook_config ) sagemaker_estimator.fit()

For a full list of CollectionConfig parameters, see Debugger CollectionConfig.