MachineLearningNotebooks/training

Training ML models with Azure ML SDK

These notebook tutorials cover the various scenarios for training machine learning and deep learning models with Azure Machine Learning.

Sample notebooks

• 01.train-hyperparameter-tune-deploy-with-pytorch
  Train, hyperparameter tune, and deploy a PyTorch image classification model that distinguishes bees vs. ants using transfer learning. Azure ML concepts covered:

  • Create a remote compute target (Batch AI cluster)
  • Upload training data using Datastore
  • Run a single-node PyTorch training job
  • Hyperparameter tune model with HyperDrive
  • Find and register the best model
  • Deploy model to ACI

• 02.distributed-pytorch-with-horovod
  Train a PyTorch model on the MNIST dataset using distributed training with Horovod. Azure ML concepts covered:

  • Create a remote compute target (Batch AI cluster)
  • Run a two-node distributed PyTorch training job using Horovod

• 03.train-hyperparameter-tun-deploy-with-tensorflow
  Train, hyperparameter tune, and deploy a TensorFlow model on the MNIST dataset. Azure ML concepts covered:

  • Create a remote compute target (Batch AI cluster)
  • Upload training data using Datastore
  • Run a single-node TensorFlow training job
  • Leverage features of the Run object
  • Download the trained model
  • Hyperparameter tune model with HyperDrive
  • Find and register the best model
  • Deploy model to ACI

• 04.distributed-tensorflow-with-horovod
  Train a TensorFlow word2vec model using distributed training with Horovod. Azure ML concepts covered:

  • Create a remote compute target (Batch AI cluster)
  • Upload training data using Datastore
  • Run a two-node distributed TensorFlow training job using Horovod

• 05.distributed-tensorflow-with-parameter-server
  Train a TensorFlow model on the MNIST dataset using native distributed TensorFlow (parameter server). Azure ML concepts covered:

  • Create a remote compute target (Batch AI cluster)
  • Run a two workers, one parameter server distributed TensorFlow training job

• 06.distributed-cntk-with-custom-docker
  Train a CNTK model on the MNIST dataset using the Azure ML base Estimator with custom Docker image and distributed training. Azure ML concepts covered:

  • Create a remote compute target (Batch AI cluster)
  • Upload training data using Datastore
  • Run a base Estimator training job using a custom Docker image from Docker Hub
  • Distributed CNTK two-node training job via MPI using base Estimator

• 07.tensorboard
  Train a TensorFlow MNIST model locally, on a DSVM, and on Batch AI and view the logs live on TensorBoard. Azure ML concepts covered:

  • Run the training job locally with Azure ML and run TensorBoard locally. Start (and stop) an Azure ML TensorBoard object to stream and view the logs
  • Run the training job on a remote DSVM and stream the logs to TensorBoard
  • Run the training job on a remote Batch AI cluster and stream the logs to TensorBoard
  • Start a Tensorboard instance that displays the logs from all three above runs in one

• 08.export-run-history-to-tensorboard

  • Start an Azure ML Experiment and log metrics to Run history
  • Export the Run history logs to TensorBoard logs
  • View the logs in TensorBoard