Shortcuts

Release Notes

Version 0.12

Version 0.12.13

Release Date: August 6, 2020

New Features

  • Model Registry: Determined now includes a built-in model registry, which makes it easy to organize trained models by providing versioning and labeling tools. See Organizing Models in the Model Registry to get started.

  • New PyTorch API: Add a new version of the PyTorch API that is more flexible and supports deep learning experiments that use multiple models, optimizers, and LR schedulers. The old API is still supported but is now deprecated and will be removed in a future release. See the migration guide for details on updating your PyTorch model code. Deprecated methods will be supported until at least the next minor release.

    • The new API supports PyTorch code that uses multiple models, optimizers, and LR schedulers. In your trial class, you should instantiate those objects and wrap them with wrap_model(), wrap_optimizer(), and wrap_lr_scheduler() in the constructor of your PyTorch trial class. The previous API methods build_model, optimizer, and create_lr_scheduler in PyTorchTrial are now deprecated.

    • Support customizing forward and backward passes in train_batch(). Support gradient clipping by passing a function to the clip_grads argument of step_optimizer(). The callback on_before_optimizer_step is now deprecated.

    • Configuring automatic mixed precision (AMP) in PyTorch should now be done by calling configure_apex_amp() in the constructor of your PyTorch trial class. The optimizations.mixed_precision experiment configuration key is now deprecated.

    • The model arguments in train_batch(), evaluate_batch(), and evaluate_full_dataset() are now deprecated.

  • More Efficient Hyperparameter Search: This release introduces a new hyperparameter search method, adaptive_asha. This is based on an asynchronous version of the adaptive algorithm, and should enable large searches to find high-quality hyperparameter configurations more quickly. See the documentation and the associated paper for more information.

Improvements

  • Allow proxy environment variables to be set in the agent config. See Environment Variables for more information.

  • Preserve random state for PyTorch experiments when checkpointing and restoring.

  • Remove determined.pytorch.reset_parameters(). This should have no effect except when using highly customized nn.Module implementations.

  • WebUI: Show total number of resources in the cluster resource charts.

Bug Fixes

  • Fix an issue with the SHA searcher that could cause searches to stop making progress without finishing.

  • Fix an issue where $HOME was not properly set in notebooks running in nonroot containers.

  • Fix an issue where killed experiments had their state reset to the latest checkpoint.

  • Randomize the notebook listening port to avoid port binding issues in host mode.

Version 0.12.12

Release Date: July 22, 2020

Improvements

  • Remove support for on_train_step_begin and on_train_step_end, deprecate on_validation_step_end, and introduce new callback on_validation_end with same functionality. Add helper methods is_epoch_start and is_epoch_end to PyTorch context.

  • Add a new API to support custom reducers in EstimatorTrial. See :ref:estimator-trial for details.

  • CLI: Add the register_version command for registering a new version of a model.

  • CLI: Add a --head option when printing trial logs.

  • WebUI: Make it possible to launch TensorBoard from experiment dashboard cards.

Bug Fixes

  • Fix distributed training and Determined shell with non-root containers. The default task environments now include a user plugin to support running containers with arbitrary non-root users. Custom images based on the latest default task environments should also work.

  • Fix convergence issue for TF 2 multi-GPU models. Change default TF1 version from 1.14 to 1.15.

  • Fix issue affecting TensorFlow TensorBoard outputs.

  • Use local log line IDs for trial logs.

  • CLI: Improve the CLI’s custom TLS certificate handling with non-self-signed certs.

  • WebUI: Fix a parsing problem with task start times.

  • WebUI: Fix log viewer timestamp copy/paste.

Known Issues

  • WebUI: Older trial logs are not loaded by scrolling to the top of the page.

Version 0.12.11

Release Date: July 8, 2020

  • Add logging to console in test mode for the Native API when using determined.experimental.create.

  • Improve reliability of saving checkpoints to GCS in the presence of transient network errors.

  • Add an example using TensorFlow’s Image Segmentation via UNet tutorial.

  • WebUI: Improve trial log rendering performance.

  • WebUI: Fix an issue where cluster utilization was displayed incorrectly.

  • WebUI: Fix an issue where active experiments and commands would not appear on the dashboard.

  • WebUI: Fix an issue where having telemetry enabled with an invalid key would cause the WebUI to render incorrectly.

Version 0.12.10

Release Date: June 26, 2020

Improvements

  • WebUI: Add a dedicated page for master logs at /det/logs.

  • WebUI: Provide a Swagger UI for exploring the Determined REST API. This can be accessed via the API link on the WebUI.

  • WebUI: Default the Experiments view list length to 25 entries. More entries can be shown as needed.

  • WebUI: Improve detection of situations where the WebUI version doesn’t match the master version as a result of browser caching.

  • CLI: Improve performance when retrieving trial logs.

  • CLI: Add the det user rename command for administrators to change the username of existing users.

  • Expand documentation on Using Checkpoints by including checkpoint metadata management.

  • Reorganize examples by splitting Native API and Trial API examples into separate folders.

Bug Fixes

  • Allow det-deploy local agent-up to work with remote masters.

  • Ensure network failures during checkpoint upload do not unrecoverably break the associated trial.

  • Ensure shared_fs checkpoint storage is usable for non-root containers for some host_path values.

  • Fix a timeout issue that affected large (40+ machines) distributed experiments.

  • Ensure the CLI can make secure connections to the master.

  • Fix an issue that affected multi-GPU in PyTorchTrial with mixed precision enabled.

  • Add a timeout to trial containers to ensure they are terminated promptly.

Version 0.12.7

Release Date: June 11, 2020

  • Breaking Change: Gradient clipping for PyTorchTrial should now be specified via determined.pytorch.PyTorchCallback via the on_before_optimizer_step() method instead of being specified via the experiment configuration. Determined provides two built-in callbacks for gradient clipping: determined.pytorch.ClipGradsL2Norm and determined.pytorch.ClipGradsL2Value.

  • Add a metadata field to checkpoints. Checkpoints can now have arbitrary key-value pairs associated with them. Metadata can be added, queried, and removed via a Python API. See the documentation for details.

  • Add support for Keras callbacks that stop training early, including the official EarlyStopping callback. When a stop is requested, Determined will finish the training (or validation) step we are in, checkpoint, and terminate the trial.

  • Add support for Estimator callbacks that stop training early, including the official stop_if_no_decrease_hook. When a stop is requested, Determined will finish the training (or validation) step we are in, checkpoint, and terminate the trial.

  • Add support for model code that stops training of a trial programmatically.

    • We recommend using the official Keras callbacks or Estimator hooks if you are using those frameworks. For PyTorch, you can request that training be stopped by calling set_stop_requested() from a PyTorch callback. When a stop is requested, Determined will finish the current training or validation step, checkpoint, and terminate the trial. Trials that are stopped early are considered to be “completed” (e.g., in the WebUI and CLI).

  • More robust error handling for hyperparameter searches where one of the trials in the search encounters a persistent error.

    • Determined will automatically restart the execution of trials that fail within an experiment, up to max_restart failures. After this point, any trials that fail are marked as “errored” but the hyperparameter search itself is allowed to continue running. This is particularly useful when some parts of the hyperparameter space result in models that cannot be trained successfully (e.g., the search explores a range of batch sizes and some of those batch sizes cause GPU OOM errors). An experiment can complete successfully as long as at least one of the trials within it completes successfully.

  • Support multi-GPU training for TensorFlow 2 models that use IndexedSlices for model parameters.

  • NaN values in training and validation metrics are now treated as errors.

    • This will result in restarting the trial from the most recently checkpoint if it has been restarted fewer than max_restarts times. Previously, NaN values were converted to the maximum floating point value.

  • Preserve the last used user name on the log-in page.

  • Add on_trial_close method to determined.estimator.RunHook. Use this for post-trial cleanup.

  • Finalize gradient communication prior to applying gradient clipping in PyTorchTrial when perfoming multi-GPU training.

  • WebUI: Add pause, activate, and cancel actions to dashboard tasks.

  • Add a det-nobody user (with UID 65533) to default images. This provides an out-of-the-box option for running non-privileged containers with a working home directory.

Version 0.12.5

Release Date: May 27, 2020

  • Breaking Change: Alter command-line options for controlling test mode and local mode. Test experiments on the cluster were previously created with det e create --test-mode ... but now should be created with det e create --test .... Local testing is started with det e create --test --local .... Fully local training (meaning --local without --test) is not yet supported.

  • Add support for TensorFlow 2.2.

  • Add support for post-checkpoint callbacks in PyTorchTrial.

  • Add support for checkpoint hooks in EstimatorTrial.

  • Add support for TensorBoard backed by S3-compliant APIs that are not AWS S3.

  • Add generic callback support for PyTorch.

  • TensorBoards now shut down after 10 minutes if metrics are unavailable.

  • Update to NCCL 2.6.4 for distributed training.

  • Update minimum required task environment version to 0.4.0.

  • Fix Native API training one step rather than one batch when using TensorFlow Keras and Estimator.

  • CLI: Add support for producing CSV and JSON output to det slot list and det agent list.

  • CLI: Include the number of containers on each agent in the output of det agent list.

Version 0.12.4

Release Date: May 14, 2020

  • Breaking Change: Users are no longer automatically logged in as the “determined” user. Refer to Users for more details.

  • Support multi-slot notebooks. The number of slots per notebook cannot exceed the size of the largest available agent. The number of slots to use for a notebook task can be configured when the notebook is launched: det notebook start --config resources.slots=2

  • Support fetching the configuration of a running master via the CLI (det master config).

  • Authentication sessions now expire after 7 days.

  • Improve log messages for tf.keras trial callbacks.

  • Add nvidia-container-toolkit support.

  • Fix an error in the experimental bert_glue_pytorch example.

  • The tf.keras examples for the Native and Trial APIs now refer to the same model.

  • Add a topic guide explaining Determined’s approach to Elastic Infrastructure.

  • Add a topic guide explaining the Native API.

  • UI: The Determined favicon acquires a small dot when any slots are in use.

  • UI: Fix an issue with command sorting in the WebUI.

  • UI: Fix an issue with badges appearing as the wrong color.

Version 0.12.3

Release Date: April 27, 2020

  • Add a tutorial for the new (experimental) Native API.

  • Add support for locally testing experiments via det e create --local.

  • Add determined.experimental.Determined class for accessing ExperimentReference, TrialReference, and Checkpoint objects.

  • TensorBoard logs now appear under the storage_path for shared_fs checkpoint configurations.

  • Allow commands, notebooks, shells, and TensorBoards to be killed before they are scheduled.

  • Print container exit reason in trial logs.

  • Choose a better default for the --tail option of command logs.

  • Add REST API endpoints for trials.

  • Support the execution of a startup script inside the agent docker container

  • Master and agent Docker containers will have the ‘unless-stopped’ restart policy by default when using det-deploy local.

  • Prevent the det trial logs -f command from waiting for too long after the trial being watched reaches a terminal state.

  • Fix bug where logs disappear when an image is pulled.

  • Fix bug that affected the use of LRScheduler in PyTorchTrial for multi-GPU training.

  • Fix bug after master restart where some errored experiments would show progress indicators.

  • Fix ordering of steps from det trial describe --json.

  • Docs: Added topic guide for effective distributed training.

  • Docs: Reorganize install documentation.

  • UI: Move the authenticated user to the top of the users list filter on the dashboard, right after “All”.

Version 0.12.2

Release Date: April 21, 2020

Breaking Changes

  • Rename PEDL to Determined. The canonical way to import it is via import determined as det.

  • Reorganize source code. The frameworks module was removed, and each framework’s submodules were collapsed into the main framework module. For example:

    • det.frameworks.pytorch.pytorch_trial.PyTorchTrial is now det.pytorch.PyTorchTrial

    • det.frameworks.pytorch.data.DataLoader is now det.pytorch.DataLoader

    • det.frameworks.pytorch.checkpoint.load is now det.pytorch.load

    • det.frameworks.pytorch.util.reset_parameters is now det.pytorch.reset_parameters

    • det.frameworks.keras.tf_keras_trial.TFKerasTrial is now det.keras.TFKerasTrial

    • det.frameworks.tensorflow.estimator_trial.EstimatorTrial is now det.estimator.EstimatorTrial

    • det.frameworks.tensorpack.tensorpack_trial is now det.tensorpack.TensorpackTrial

    • det.frameworks.util and det.frameworks.pytorch.util have been removed entirely

  • Unify all plugin functions under the Trial class. make_data_loaders has been moved to two functions that should be implemented as part of the Trial class. For example, PyTorchTrial data loaders should now be implemented in build_training_data_loader() and build_validation_data_loader() in the trial definition. Please see updated examples and documentation for changes in each framework.

  • Trial classes are now required to define a constructor function. The signature of the constructor function is:

    def __init__(self, context) -> None:
    

    where context is an instance of the new det.TrialContext class. This new object is the primary mechanism for querying information about the system. Some of its methods include:

    • get_hparam(name): get a hyperparameter by name

    • get_trial_id(): get the trial ID being trained

    • get_experiment_config(): get the experiment config for this experiment

    • get_per_slot_batch_size(): get the batch size appropriate for training (which will be adjusted from the global_batch_size hyperparameter in distributed training experiments)

    • get_global_batch_size(): get the effective batch size (which differs from per-slot batch size in distributed training experiments)

    • distributed.get_rank(): get the unique process rank (one process per slot)

    • distributed.get_local_rank(): get a unique process rank within the agent

    • distributed.get_size(): get the number of slots

    • distributed.get_num_agents: get the number of agents (machines) being used

  • The global_batch_size hyperparameter is required (that is, a hyperparameter with this name must be specified in the configuration of every experiment). Previously, the hyperparameter batch_size was required and was manipulated automatically for distributed training. Now global_batch_size will not be manipulated; users should train based on context.get_per_slot_batch_size(). See Distributed Training for more context.

  • Remove download_data(). If users wish to download data at runtime, they should make sure that each process (one process per slot) downloads to a unique location. This can be accomplished by appending context.get_rank() to the download path.

  • Remove det.trial_controller.util.get_rank() and det.trial_controller.util.get_container_gpus(). Use context.distributed.get_rank() and context.distributed.get_num_agents() instead.

General Improvements

  • tf.data.Dataset is now supported as input for all versions of TensorFlow (1.14, 1.15, 2.0, 2.1) for TFKerasTrial and EstimatorTrial. Please note that Determined currently does not support checkpointing tf.data.Dataset inputs. Therefore, when resuming training, it resumes from the start of the dataset. Model weights are loaded correctly as always.

  • TFKerasTrial now supports five different types of inputs:

    1. A tuple (x_train, y_train) of NumPy arrays. x_train must be a NumPy array (or array-like), a list of arrays (in case the model has multiple inputs), or a dict mapping input names to the corresponding array, if the model has named inputs. y_train should be a NumPy array.

    2. A tuple (x_train, y_train, sample_weights) of NumPy arrays.

    3. A tf.data.Dataset returning a tuple of either (inputs, targets) or (inputs, targets, sample_weights).

    4. A keras.utils.Sequence returning a tuple of either (inputs, targets) or (inputs, targets, sample weights).

    5. A det.keras.SequenceAdapter returning a tuple of either (inputs, targets) or (inputs, targets, sample weights).

  • PyTorch trial checkpoints no longer save in MLflow’s MLmodel format.

  • The det trial download command now accepts -o to save a checkpoint to a specific path. PyTorch checkpoints can then be loaded from a specified local filesystem path.

  • Allow the agent to read configuration values from a YAML file.

  • Include experiment ID in the downloaded trial logs.

  • Display checkpoint storage location in the checkpoint info modal for trials and experiments.

  • Preserve recent tasks’ filter preferences in the WebUI.

  • Add task name to det slot list command output.

  • Model definitions are now downloaded as compressed tarfiles (.tar.gz) instead of zipfiles (.zip).

  • startup-hook.sh is now executed in the same directory as the model definition.

  • Rename projects to examples in the Determined repository.

  • Improve documentation:

    • Add documentation page on the lifecycle of an experiment.

    • Add how-to and topic guides for multi-GPU (both for single-machine parallel and multi-machine) training.

    • Add a topic guide on best practices for writing model definitions.

  • Fix bug that occasionally caused multi-machine training to hang on initialization.

  • Fix bug that prevented TensorpackTrial from successfully loading checkpoints.

  • Fix a bug in TFKerasTrial where runtime errors could cause the trial to hang or would silently drop the stack trace produced by Keras.

  • Fix trial lifecycle bugs for containers that exit during the pulling phase.

  • Fix bug that led to some distributed trials timing out.

  • Fix bug that caused tf.keras trials to fail in the multi-GPU setting when using an optimizer specified by its name.

  • Fix bug in the CLI for downloading model definitions.

  • Fix performance issues for experiments with very large numbers of trials.

  • Optimize performance for scheduling large hyperparameter searches.

  • Add configuration for telemetry in master.yaml.

  • Add a utility function for initializing a trial class for development (det.create_trial_instance)

  • Add security.txt.

  • Add det.estimator.load() to load TensorFlow Estimator saved_model checkpoints into memory.

  • Ensure AWS EC2 keypair exists in account before creating the CloudFormation stack.

  • Add support for gradient aggregation in Keras trials for TensorFlow 2.1.

  • Add TrialReference and Checkpoint experimental APIs for exporting and loading checkpoints.

  • Improve performance when starting many tasks simultaneously.

Web Improvements

  • Improve discoverability of dashboard actions.

  • Add dropdown action menu for killing and archiving recent tasks on the dashboard.

  • Add telemetry for web interactions.

  • Fix an issue around cluster utilization status showing as “No Agent” for a brief moment during initial load.

  • Add Ace editor to attributions list.

  • Set UI preferences based on the logged-in user.

  • Fix an issue where the indicated user filter was not applied to the displayed tasks.

  • Improve error messaging for failed actions.