Release Notes

Version 0.21

Version 0.21.0

Release Date: March 27, 2023

Breaking Changes

• Cluster: K80 GPUs are no longer supported.

• API: Remove all old PATCH endpoints under /agents*, including the APIs for enabling and disabling slots. Users should use the new APIs under /api/v1/agents.

• API: The on_validation_step_start and on_validation_step_end callbacks on PyTorchTrial and DeepSpeedTrial were deprecated in 0.12.12 (Jul 2020) and have been removed. Please use on_validation_start and on_validation_end instead.

• Trial API: records_per_epoch has been dropped from PyTorch code paths. We were previously using this value internally to estimate epoch lengths. We are now using the chief worker’s epoch length as the epoch length.

• API: average_training_metrics is no longer configurable. This value previously defaulted to false and was dropped to simplify the training API. We always average training metrics now.

• API: The unused latest_training field has been removed from the GetTrial and GetExperimentTrials APIs due to slow performance.

Bug Fixes

• CLI: Fix an issue where det user change-password would return an authentication error when trying to change the current user’s password.

Improvements

• CLI: Command-line deployments will now default to provisioning Nvidia T4 GPU instances instead of

  K80 instances. This change is intended to improve the performance/cost and driver support of the default deployment.

• Kubernetes: Ease permission requirements in Kubernetes so master no longer requires access to all Kubernetes namespaces. This only affects custom modified Helm chart configurations.

• Checkpoints: Improve performance of checkpoint insertion and deletion.

New Feature

• API: Deprecate TorchWriter and add a PyTorch SummaryWriter object to PyTorchTrialContext and DeepSpeedTrialContext that we manage on behalf of users. See get_tensorboard_writer() for details.

• API: Introduce Trainer, a high-level training API for PyTorchTrial that allows for Python-side training loop customizations and includes support for off-cluster local training.

Removed Features

• The following methods of Checkpoint, Model, and ModelVersion were deprecated in 0.17.9 (Feb 2022) and are now removed:

  • Checkpoint.load()

  • Checkpoint.load_from_path()

  • Checkpoint.parse_metadata()

  • Checkpoint.get_type()

  • Checkpoint.from_json()

  • Model.from_json()

  • ModelVersion.from_json()

Version 0.20

Version 0.20.1

Release Date: March 15, 2023

Breaking Changes

• Database: Several unused columns have been dropped from the raw_steps, raw_validations, and raw_checkpoints database tables. The database migration will involve a sequential scan for these tables, and it may take a significant amount of time, depending on the database size and performance.

New Features

• Allow tasks and experiments to expose arbitrary ports that can be tunneled to the using the CLI. See Exposing Custom Ports for more details or an example at examples/features/ports.

• Container Images: Add maintained images for PyTorch-only environments.

  • Current environment images contain both PyTorch and TensorFlow which results in large image sizes. The new images are appropriate for users who do not require TensorFlow but may still require TensorBoard.

Removed Features

• API: Remove internal ExpCompareMetricNames and ExpCompareTrialsSample endpoints, which have been unused and deprecated since 0.19.5.

Known Issue

• For multi-trial experiments, training metrics do not start appearing unless there has been at least one validation.

Version 0.20.0

Release Date: February 28, 2023

Breaking Changes

• Cluster: The resources.agent_label task option and label agent config option are no longer supported and will be ignored. If you are not explicitly using these options, or only use a single empty or non-empty label value per resource pool, no changes are necessary. Otherwise, cluster admins should create a resource pool for each existing resource_pool/agent_label combination and reconfigure agents to use these new pools. Cluster users should update their tasks to use the new resource pool names.

Bug Fixes

• Model Registry: Fix an issue where a model with versions from multiple workspaces could have its versions modified by a user with edit access to only a single one of those workspaces.

• WebUI: Patch an issue where logging out would not properly redirect to the login page.

• WebUI: Fix a bug where the cluster’s job queue page could crash in certain cases.

Improvements

• Agents: The master configuration agent_reattach_enabled is always enabled and agents will now always reattach containers on restart.

• Kubernetes: The cluster information page now takes resource quotas into account if there are any on relevant namespaces.

• RBAC: Model registry models and commands that are inaccessible to the user will appear as uneditable. Previously, users could attempt the action and would encounter a permission denied error.

• CLI: When listing TensorBoards, show workspaceName instead of workspaceId for better readability and prevent N/A values from appearing.

New Features

• RBAC: Following on the initial RBAC support added in 0.19.7, the enterprise edition of Determined (HPE Machine Learning Development Environment) has added support for role-based access control (RBAC) over new entities:

  • Notebooks, TensorBoards, shells, and commands are now housed under workspaces. Access to these tasks can now be restricted by role.

  • Model Registry: Models are now associated with workspaces. Models can be moved between workspaces and access to them can be restricted by role.

  These changes allow for more granular control over who can access what resources. See RBAC for more information.

Version 0.19

Version 0.19.11

Release Date: February 17, 2023

Bug Fixes

• Kubernetes: Fix an issue where environment variables with an equals character in the value, such as func=f(x)=x, were processed incorrectly in Kubernetes.

• Agent: Fix a bug where if agent reattach was enabled and the master was down while an active task’s Docker container failed, the task could get stuck in an unkillable running state.

• det deploy aws: Update CloudFormation permissions to allow checkpoint downloads through master.

• Tasks: Fix a bug where in rare cases tasks could take an extra 30 seconds to complete.

Improvements

• Container Images: Publish multi-arch master and agent container image manifests with AMD64, ARM64, and PPC64 architectures.

• Experiments: If an experiment with no checkpoints is deleted, a checkpoint GC task will no longer be launched. Launching a checkpoint GC task could prevent experiments with certain incorrect configuration from being deleted.

• Cluster: Capability added for checkpoint downloads from Google Cloud Storage via a master instance.

• Installation: .deb and .rpm Linux packages will now install master and agent binaries into /usr/bin/ instead of /usr/local/bin/, to be more in line with the Filesystem Hierarchy Standard.

• Kubernetes: Empty environment variables can now be specified in Kubernetes, while before they would throw an error.

• Kubernetes: Zero-slot tasks on GPU clusters will not request nvidia.com/gpu: 0 resources any more, allowing them to be scheduled on CPU-only nodes.

• Installation: Add experimental Homebrew (macOS) package.

• Scheduler: The scheduler can be configured to find fits for distributed jobs against agents of different sizes.

New Features

• CLI: Add a --add-tag flag to AWS det deploy aws up, which specifies tags to add to the underlying CloudFormation stack.

  • New tags will not replace automatically added tags such as deployment-type or managed-by.

  • Any added tags that should persist across updates should be always be included when using det deploy aws up – if the argument is missing, any previously added tags would be removed.

Version 0.19.10

Release Date: January 20, 2023

Breaking Changes

• Kubernetes: Add the kubernetes_namespace config field for resource pools, specifying a Kubernetes namespace that tasks will be launched into.

• The name of the resource pool in Kubernetes has changed from "kubernetes" to "default". Forked experiments will need to have their configurations manually modified to update the resource pool name.

New Features

• Cluster: Add support for experiment tag propagation.

  • The enterprise edition of Determined (HPE Machine Learning Development) now allows for experiment tags to be propagated as labels to the associated jobs on the HPC cluster. A number of labeling schemes are supported, controlled by the configuration item resource_manager.job_project_source.

• Cluster: Add support for launcher-provided resource pools.

  • The enterprise edition of Determined (HPE Machine Learning Development) now allows for custom resource pools to be defined that submit work to an underlying Slurm/PBS partition on an HPC cluster with different submission options.

• Cluster: Determined Enterprise Edition now supports the NVIDIA Enroot container platform as an alternative to Apptainer/Singularity/PodMan.

Improvements

• Notebooks: The default idle notebook termination timeout can now be set via the notebook_timeout master config option.

• Trials: Trials can now be killed when in the STOPPING_CANCELED state. Previously, if a trial did not implement preemption correctly and was canceled, the trial did not stop and was unkillable until the preemption timeout of an hour.

Bug Fixes

• Fix a bug where notebooks, TensorBoards, shells, and commands restored after a master restart would have a submission time of when the master restarted rather than the original job submission time.

• det deploy aws: Fix reliability issue in efs deployment type, fix broken fsx deployment type.

• Job queue: Fix an issue where the CLI command det job list would ignore the argument --resource-pool.

• Distributed training: Fix a bug where a distributed training trial that called context.set_stop_requested would cause the trial to error and prevent it from completing successfully.

Removed Features

• The data layer feature, which was deprecated in 0.18.0 (May 2022), has been removed. A migration guide to use the underlying yogadl library directly may be found here. Affected users are encouraged to follow the migration guide before upgrading to avoid downtime.

Version 0.19.9

Release Date: December 20, 2022

New Features

• WebUI: Display total checkpoint size for experiments.

• WebUI: Add links from forked experiments and continued trials to their parents.

• API: Add structured fields to task log objects.

• Cluster: Add support for launcher-provided resource pools. Determined Enterprise Edition now allows for custom resource pools to be defined that submit work to an underlying Slurm/PBS partition on an HPC cluster with different submission options.

• Cluster: Determined Enterprise Edition now supports the NVIDIA Enroot container platform as an alternative to Apptainer/Singularity/PodMan.

Version 0.19.8

Release Date: December 02, 2022

Breaking Changes

• API: The GetModelVersion, PatchModelVersion, and DeleteModelVersion APIs now take a sequential model version number model_version_num instead of a surrogate key model_version_id.

Bug Fixes

• Experiment: Fix an issue where experiments created before version 0.16.0 could have issues loading.

• Python SDK: Fix an issue where the Model Registry call model.get_version(version) did not work when a specific version was passed.

Improvements

• Kubernetes: If a pod exits and Determined cannot get the exit code, the code will be set to 1025 instead of 137 to avoid confusion with potential out-of-memory issues.

• API: Patching a user will no longer make partial updates if an error occurs.

• Kubernetes: Specifying tensorboardTimeout in Helm will cause the specified timeout to be applied.

• AWS: det deploy aws will use IMDSv2 for improved security.

New Features

• Experiment: Determined Enterprise Edition now allows control of the GPU type within a Slurm GRES expression. If you have partitions with mixed GPU types, you may now specify the desired type using the slurm.gpu_type attribute of the experiment configuration.

Version 0.19.7

Release Date: November 14, 2022

New Features

• WebUI: Adds support for creating and managing webhooks to enable receiving updates regarding experiment state changes.

• Checkpoint storage can now be configured at a workspace level. Experiments created in projects will now inherit checkpoint storage configuration from the project’s workspace if set. Experiment configuration can override the workspace level checkpoint storage configuration.

• Example: Textual Inversion training and generation using Stable Diffusion with Core API and Hugging Face’s Diffusers.

• Python SDK now supports reading logs from trials, via the new logs() method. Additionally, the Python SDK also supports a new blocking call on an experiment to get the first trial created for an experiment via the await_first_trial() method. Users who have been writing automation around the det e create --follow-first-trial CLI command may now use the python SDK instead, by combining .await_first_trial() and .logs().

• RBAC: the enterprise edition of Determined (HPE Machine Learning Development Environment) has added preliminary support for Role-Based Access Control. Administrators can now configure which users or user groups can administer users, create or configure workspaces, run or view experiments in particular workspaces, or perform other actions. See RBAC for more information.

Bug Fixes

• Master: Correctly handle pending allocations in historical resource allocation aggregation.

Version 0.19.6

Release Date: October 28, 2022

Breaking Changes

• API: Remove the legacy endpoint /tasks/:task_id due to it always incorrectly returning a missing parameter.

• Experiment: Additional Slurm options formerly specified in the experiment environment section are now part of a new slurm section of the experiment configuration. For example, what was formerly written as

  environment:
  ...
    slurm:
      - --mem-per-cpu=10
      - --exclusive
  

  is now specified as

  environment:
  ...
  slurm:
    sbatch_args:
      - --mem-per-cpu=10
      - --exclusive
  

Improvements

• CLI: Add the ls abbreviation for list to all applicable CLI commands.

• CLI: Support a new -i/--include option in task-starting CLI commands. The context option (--context) is useful for copying a directory of files into the task container, but it may only be provided once, and it can be clunky if you only care about one or two files. The --include option also copies files into the task container, but:

  • The directory name is preserved, so -i my_data/ would result in a directory named my_data/ appearing in the working directory of the task container.

  • It may point to a file, so -i my_data.csv will place my_data.csv into the working directory.

  • It may be specified multiple times to include multiple files and/or directories.

• Breaking Change: det deploy aws by default now configures agent instances to automatically shut down if they lose their connection to the master. The --no-shut-down-agents-on-connection-loss option can be used to turn off this behavior.

New Features

• Custom Searcher: users can now define their own logic to coordinate across multiple trials within an experiment. Examples of use cases are custom hyperparameter searching algorithms, ensembling, active learning, neural architecture search, reinforcement learning. See Custom Search Methods for more information.

• Cluster: The enterprise edition of HPE Machine Learning Development Environment can now be deployed on a PBS cluster. When using PBS scheduler, HPE Machine Learning Development Environment delegates all job scheduling and prioritization to the PBS workload manager. This integration enables existing PBS workloads and HPE Machine Learning Development Environment workloads to coexist and access all of the advanced capabilities of the PBS workload manager. You can use either Singularity or Podman for the container runtime.

Version 0.19.5

Release Date: October 10, 2022

Improvements

• Added the ability to set what Unix user and group tasks will run as on the agent at the workspace level. The setting takes precedence over users’ individual user and group settings.

• CLI: The det workspace edit command now accepts a new workspace name as an optional --name flag, e.g., det workspace edit OLD_WORKSPACE_NAME --name NEW_WORKSPACE_NAME.

Bug Fixes

• Agent: Fixed a bug where in certain cases of the master restarting with active tasks, the agent resource manager could prevent other tasks from running.

• Kubernetes: When a TensorBoard inherits its images from an experiment configuration, it now also inherits the environment.pod_spec.spec.imagePullSecrets value.

Version 0.19.4

Release Date: September 22, 2022

Breaking Changes

• det deploy aws: Remove --deployment-type=vpc option. Please use efs or fsx deployment types instead.

API Changes

• The STATE_ACTIVE state for experiments and trials is now divided into four sub-states: STATE_QUEUED, STATE_PULLING, STATE_STARTING, and STATE_RUNNING. Queries to GetExperimentsRequest that filter by state continue to use STATE_ACTIVE.

• The possible states of tasks have been adjusted to match those of experiments and trials. The previous STATE_PENDING and STATE_ASSIGNED are now STATE_QUEUED.

Bug Fixes

• Checkpoints: Fixed a bug where operations that listed checkpoints could sometimes return the same checkpoint multiple times.

Version 0.19.3

Release Date: September 09, 2022

Improvements

• Slurm: Singularity containers may now use AMD ROCm GPUs.

• Slurm: PodMan V4.0+ is now supported in conjunction with the Slurm job scheduler.

• Kubernetes: The UID and GID of Fluent Bit logging sidecars may now be configured on a cluster-wide basis.

New Features

• Example: Allow training of models that do not fit into GPU memory using DeepSpeed ZeRO Stage 3 with CPU offloading.

• Kubernetes: Allow the UID and GID of Fluent Bit logging sidecars to be configured on a cluster-wide basis.

Version 0.19.2

Release Date: August 26, 2022

Breaking Changes

• API: Response format for metrics has been standardized to return aggregated and per-batch metrics in a uniform way. GetTrialWorkloads, GetTrials API response format has changed. ReportTrialTrainingMetrics, ReportTrialValidationMetrics API request format has changed as well.

• API: GetJobs request format for pagination object has changed. Instead of being contained in a nested pagination object, these are now top level options, in line with the other paginatable API requests.

• CLI: det trial describe --json output format has changed. Fixed a bug where det trial describe --json --metrics would fail for trials with a very large number of steps.

• CLI: det job list will now return all jobs by default instead of a single API results page. Use --pages=1 option for the old behavior.

• The /api/v1/trials/:id endpoint no longer returns the workloads attribute. Workloads should instead be retrieved from the paginated /api/v1/trials/:id/workloads endpoint.

Bug Fixes

• Kubernetes: Fixed an issue where restoring a job in a Kubernetes set up could crash the resource manager.

• CLI: Fixed a bug where det e set gc-policy would fail when deserializing an api response because it wasn’t adjusted for the new format.

• Distributed training: Previously, experiments launched with determined.launch.torch_distributed were wrongly skipping torch.distributed.run for single-slot trials and invoking training scripts directly. As a result, functions such as torch.distributed.init_process_group() would fail, but only inside single-slot trials. Now, determined.launch.torch_distributed will conform to the intended behavior as a wrapper around torch.distributed.run and will invoke torch.distributed.run on all training scripts.

• Experiments with a single trial are now considered canceled when their trial is canceled or killed.

Improvements

• API: GetTrialWorkloads can now optionally include per-batch metrics when includeBatchMetrics query parameter is set.

New Features

• Cluster: The enterprise edition of Determined (HPE Machine Learning Development), can now be deployed on a Slurm cluster. When using Slurm, Determined delegates all job scheduling and prioritization to the Slurm workload manager. This integration enables existing Slurm workloads and Determined workloads to coexist and access all of the advanced capabilities of the Slurm workload manager. The Determined Slurm integration can use either Singularity or Podman for the container runtime.

Version 0.19.1

Release Date: August 11, 2022

Fixes

• Fix the Python SDK with Determined 0.19.0. An important endpoint broke in the 0.19.0 release, causing several Python SDK methods to break. Additional tests have been added to prevent similar breakages in the future.

Improvements

• API: new on_training_workload_end and on_checkpoint_upload_end PyTorchCallback methods available for use with PyTorchTrial and DeepSpeedTrial.

• API: PyTorchTrial and DeepSpeedTrial callback `on_checkpoint_end deprecated in favor of on_checkpoint_write_end, re-named for clarity.

New Features

• Web: Add a button to start a hyperparameter search experiment based on an experiment or trial. The button brings up a form allowing users to change searcher settings and hyperparameter ranges.

Version 0.19.0

Release Date: July 29, 2022

New Features

• Introduce a file system cache for model definition files, configured via cache.cache_dir in the master configuration. The default path is /var/cache/determined. Note that the master will crash on startup if the directory does not exist and cannot be created.

Improvements

• Security: Setting registry_auth.serveraddress will now only send credentials to the server configured. Not setting registry_auth.serveraddress is now deprecated when registry_auth is set. In the future, serveraddress will be required whenever registry_auth is set.

• Agent: Users may now run docker login on agent host machines to authenticate with Docker registries. Note that if the agent is running inside a Docker container then ~/.docker/config.json will need to be mounted to $HOME/.docker/config.json (by default /root/.docker/config.json) inside the container.

• CLI: The Determined CLI now supports reading a username and password from the DET_USER and DET_PASS environment variables to avoid the need to run det user login, allowing for easier use of the CLI in scripts. det user login is still the preferred mechanism for most use cases of the CLI.

Breaking Changes

• Experiment: The default value for the average_training_metrics experiment configuration option has been changed to true. This change only affects distributed training. The previous default of false leads to only the chief worker’s training metrics being reported. Setting this configuration to true instead reports the true average of all workers’ training metrics at the cost of increased communication overhead. Users who do not require accurate training metrics may explicitly set the value to false as an optimization.

• API: The /projects/:id/experiments endpoint has been removed and replaced with a project_id parameter on the /experiments endpoint.

• API: The config attribute in the response of the /experiments/:id endpoint has been moved into the experiment object. The config attribute is now also available for experiments returned from the /experiments endpoint.

Bug Fixes

• When creating a test experiment, the container storage path was not being set correctly.

• Notebooks: Fix a bug where notebooks would ignore the --template CLI argument.

• Notebooks: Fix a bug where running det notebook start --preview would launch a notebook instead of just displaying the configuration.

• Kubernetes: Fix an issue where zero-slot tasks would use the GPU image instead of the CPU image.

• Kubernetes: Fix an issue where zero-slot tasks would incorrectly be exposed to all GPUs.

• Kubernetes: Fix an issue where the Helm option defaultPassword caused the deployment to hang.

• Ensure an allocation’s recorded end time is always valid, even on restoration failures. Invalid end times could cause historical reporting rollups to fail. If there were any failures, they will be fixed by database migrations this update.

Security Fixes

• Breaking Change PyTorch Lightning is no longer a part of Determined environments. When needed, it should be installed as part of startup hooks.

Version 0.18

Version 0.18.4

Release Date: July 14, 2022

New Features

• Configuration: Add support for task_container_defaults.environment_variables in the master config, which allows users to specify a list of environment variables that will be set in the default task container environment.

• Web: Most user settings and preferences, like filters, are now persisted to the database. Users will now be able to retain their settings across devices.

Bug Fixes

• Since 0.17.7, det experiment download-model-def $ID has been saving the downloaded tarballs as just $ID. This release corrects that behavior and names them experiment_$ID_model_def.tgz instead.

• Kubernetes: Fix a bug where following the link to live TensorBoards would redirect to the Uncategorized page.

• Ensure an allocation’s recorded end time is always valid, even on restoration failures. Invalid end times could cause historical reporting rollups to fail. Previous failures, if any, will be fixed by database migrations this update.

Improvements

• Add the resource pool field when listing experiments or commands in Kubernetes, where it was previously left blank.

Version 0.18.3

Release Date: July 07, 2022

Breaking Changes

• WebUI: Remove previously unlisted cluster page. This page has been replaced by a new version available through the navigation bar.

New Features

• Workspaces & Projects: Teams can now organize related experiments into projects and workspaces. See video for a walkthrough.

• Logging: Master configuration now supports logging.additional_fluent_outputs allowing advanced users to specify custom integrations for task logs.

• Kubernetes: Task init containers no longer require root privileges.

• API: Trial API now uploads profiling data to the checkpoint storage from all workers. Core API users can now pass a new optional argument, tensorboard_mode, to core.init(). The default value is AUTO. In AUTO mode, TensorBoard metrics are written on the chief, and metrics as well as profiling data are uploaded to checkpoint storage from the chief only. In MANUAL mode, the user is responsible for writing TensorBoard metrics and uploading profiling data. In order to make that possible, two new methods are introduced on TrainContext: get_tensorboard_path() returns the path to the directory where metrics can be written and upload_tensorboard_files() uploads metrics and other files, such as profiling data, to checkpoint storage.

• Add support for recovering live commands, notebooks, TensorBoards, and shells on master restart. This is an extension of live trial recovery, available since version 0.18.1.

Bug Fixes

• WebUI: Fix a bug where a previous resource pool selection would not update when a new resource pool is selected for viewing associated jobs.

• API: Fix a bug where /api/v1/tasks/{taskId} would often return incorrect allocation states.

• Since 0.17.15, there was a bug where task_container_defaults.registry_auth was not correctly passed to tasks, resulting in tasks being unable to pull images.

Improvements

• CLI: Add new flag --agent-config-path to det deploy local agent-up allowing custom agent configs to be used.

• CLI: Add det (notebook|shell|tensorboard) list --json option, allowing user to get JSON-formatted notebook, shell or tensorboard task list.

• Configuration: Experiment configuration resources.shm_size now supports passing in a unit like 4.5 G or 128MiB.

Version 0.18.2

Release Date: June 14, 2022

Bug Fixes

• Web: Update task cards to only truncate task UUIDs and leave experiment IDs alone.

• CLI: Fix an issue for det task logs where trial task IDs and checkpoint GC task IDs could not be used.

• Agent: Fix being unable to use control-C to cancel the agent when it is connecting to master.

• Trial: Fix a bug where the rendezvous timeout warning could be printed erroneously.

• Commands: Fix an issue for commands where setting an environment variable as FOO instead of FOO=bar in environment.environment_variables causes the agent to panic.

Fixes

• Prevent certain hangs when using one of Determined’s built-in launchers, which begin in release 0.18.0. These hangs were caused by wrapper processes seeing SIGTERM but not passing it to their child process.

• Supports running in containers that do not have a /bin/which path, such as python-slim. The error was caused by accidentally hardcoding /bin/which instead of letting the shell find which on the path.

• Automatically add a determined_version key to the metadata of checkpoints created by any of the Trial APIs. This automatic key was accidentally dropped in release 0.18.0. Note that Core API checkpoints have full control over their checkpoint metadata and so are unaffected.

Improvements

• Scheduler: Tasks now release resources as they become free instead of holding them until all resources are free.

• CLI: det deploy aws up, det deploy aws down, and det deploy gcp down now take --yes to skip prompts asking for confirmation. --no-prompt is still usable.

• Experiments: When attempting to delete an experiment, if the delete fails it is now retryable.

• Agents: Improve behavior and observability when agents lose WebSocket messages due to network failures.

• Trials: Trial logs will report some system events such as when a trial gets canceled, paused, killed, or preempted.

New Features

• Kubernetes: Specifying observability.enable_prometheus in Helm will now correctly enable Prometheus monitoring routes.

• Kubernetes: Users may now specify a checkpointStorage.prefix in the Determined Helm chart if using S3 buckets for checkpoint storage. Checkpoints will now be uploaded with the path prefix whereas before it was ignored.

• CLI: Add new command det experiment logs <experiment-id> to get logs of the first trial of an experiment. Flags from det trial logs are supported.

• Configuration: Add support for checkpointStorage.prefix in master and experiment configuration for Google Cloud Storage (gcs).

Security Fixes

• API: Endpoints under /debug/pprof now require authentication.

Version 0.18.1

Release Date: May 24, 2022

New Features

• Web: Themes have been introduced and styles have been adjusted to support various themes. Theme switching is currently limited to dark/light mode and is set first through OS-level preferences, then through browser-level preference. In-app controllers will be coming soon.

• Add experimental support for recovering live trials on master process restart. Users can restart the master (with updated configuration options or an upgraded software version), and the current running trials will continue running using the original configuration and harness versions. This requires the agent to reconnect within a configurable agent_reconnect_wait period. This is only available for the agent resource manager, and can be enabled for resource pools using the agent_reattach_enabled flag. May only be available for patch-level releases.

• Web: A trial restart counter has been added to the experiment detail header for single-trial experiments. For multi-trial experiments, trial restart counts are shown in a new Restarts column in the Trials table.

  

Improvements

• Security: Improved security by requiring admin privileges for the following actions.

  • Reading master config.

  • Enabling or disabling an agent.

  • Enabling or disabling a slot.

• Logging: Ensure logs for very short tasks are not truncated in Kubernetes.

• Web: Centralize sidebar options Cluster, Job Queues, and Cluster Logs into Cluster page for a simplified layout.

• Web: In order to provide a more precise view of resource pools, new fields like accelerator and warm slots have been added.

• Web: Clicking on resource pool cards will lead to a detail page, which also includes a Stats tab showing average queued time by day.

Breaking Changes

• Security: The following routes and CLI commands now need admin privileges.

  • /config

  • /api/v1/master/config

  • /api/v1/agents/:agent_id/enable

  • /api/v1/agents/:agent_id/disable

  • /agents/:agent_id/slots/:slot_id

  • /api/v1/agents/:agent_id/slots/:slot_id/enable

  • /api/v1/agents/:agent_id/slots/:slot_id/disable

  • det master config

  • det agent enable

  • det agent disable

  • det slot enable

  • det slot disable

• Logging: The default Fluent Bit version in all deployment modes is now 1.9.3, changed from 1.6.

Bug Fixes

• Web: Fix the user filtering for migrating from Determined 0.17.15 to Determined 0.18.0.

• API: Fix an issue where the POST /users endpoint always returned an error instead of the user’s information, even when the user was created successfully.

Version 0.18.0

Release Date: May 09, 2022

New Features

• Add the Core API. The Core API is the first API offered by Determined that allows users to fully integrate arbitrary models and training loops into the Determined platform. All of the features offered by the higher-level Trial APIs, such as reporting metrics, pausing and reactivating, hyperparameter search, and distributed training, are now available to arbitrary models, frameworks, and training loops, with only light code changes.

• Breaking Change: Checkpoints: The Python SDK’s Checkpoint.download() method now writes a differently formatted metadata.json file into the checkpoint directory. Previously, the JSON content in the file contained many system-defined fields, plus a metadata field that contained the user-defined metadata for the checkpoint, which was also available as a Python object as Checkpoint.metadata. Now, metadata.json contains only the user-defined metadata, and those metadata appear as top-level keys. Some of the fields which were previously system-defined are now considered user-defined, even though they are uploaded automatically in Trial-based training. This decision is in line with the Trial APIs now being optional—that is, part of userspace—after the release of the Core API.

• Job queue: Add support for dynamic job modification on Kubernetes using the job queue. Users can now use the WebUI or CLI to change the priority and queue position of jobs in k8s. To update jobs through the WebUI, go to the Job Queue section, find the target job, and click on the Manage Job option. To update jobs in the CLI, use the det job update command. Run det job update --help for more information.

Bug Fixes

• CLI: API requests executed through the Python bindings have been erroneously using the SSL “noverify” option since version 0.17.6, making them potentially insecure. The option is now disabled.

Deprecated Features

• The Determined Data Layer has been deprecated and will be removed in a future version. New code should not begin using it, but we will assist existing users to migrate to using YogaDL directly before removing the feature.

Removed Features

• Python API: The old experimental namespace methods for custom reducers in both PyTorchTrial and EstimatorTrial have been removed. The experimental names were deprecated in 0.15.2 (April 2021) when custom reducers were promoted to general availability. Any users who have not already migrated to the non-experimental namespace for custom reducer methods must do so.

• Searcher: Remove the PBT searcher, which was deprecated in version 0.17.6 (January 2022).

• API: Remove the notebook logs endpoint in favor of the new task logs endpoint.

• Python API: Remove the remaining parts of the Native API, which was deprecated in version 0.13.5 (September 2020). The only Native API functions that still remained were det.experimental.create() and det.experimental.create_trial_instance().

• Python API: Remove the det.pytorch.reset_parameters() function, which was deprecated in 0.12.13 (August 2020).

Version 0.17

Version 0.17.15

Release Date: April 22, 2022

Breaking Changes

• API: Endpoints for getting or updating a user now accept a userId instead of username as the path parameter.

Bug Fixes

• Fix an issue where deleted experiments would get stuck in a DELETING state indefinitely due to their checkpoint GC tasks not completing.

• API: Fix an issue where a reported job state could be stale due to a faulty caching mechanism. This could have resulted in an experiment showing in queued or scheduled state, either in CLI or WebUI, when it was in the other state.

New Features

• Add a translation of DeepSpeed’s DCGAN example using the new DeepSpeedTrial API.

Version 0.17.14

Release Date: April 13, 2022

Bug Fixes

• Resource Pool: Fix a bug that causes the resource pool and resource manager to crash after submitting a command with a non-default priority. We recommend that all users on 0.17.12 and 0.17.13 update to 0.17.14 or later.

Version 0.17.13

Release Date: April 07, 2022

New Features

• Support DeepSpeed with a new DeepSpeedTrial API.

  DeepSpeed is a powerful library for training large scale models. With the new DeepSpeedTrial you can combine all the benefits of Determined with the features available in DeepSpeed like the Zero Redundancy Optimizer and pipeline parallel training. We also provide an example based on Eleuther AI’s GPT-NeoX repo to help you get started training state-of-the-art language models.

• CLI: Allow the CLI to accept any unique prefix of a task UUID to refer to the task, rather than requiring the entire UUID. In some places, Determined only displays the first few characters of a UUID.

Improvements

• Model Hub: add support for panoptic segmentation.

  • Model Hub mmdetection now supports panoptic segmentation task in addition to object detection. Previously, the associated docker image lacked dependencies for panoptic segmentation. Users can now use mmdetection configs under panoptic_fpn and also the coco_panoptic dataset base config.

• Collect data for agent/instance start time and end time in order to track unused GPUs. Two new kinds (agent and instance) added to CSV report at Cluster page.

API Changes

• The model registry API now accepts either the ID or model name in /api/v1/models/:id or /api/v1/models/:name. This applies to all API routes for models and model versions.

• The ID can be used in the API and the WebUI (/det/models/:id) as a permanent link to the model.

Breaking Changes

• Changed the message body of PatchModelRequest and PatchModelVersionRequest such that the POST-ed body is the PatchModel or PatchModelVersion object, instead of being wrapped in { "model": PatchModel }.

• Updated typing hints on other Model Registry API endpoints to make it clear which fields will be returned in API responses.

Bug Fixes

• Fix an issue where the originally requested page to redirect to after a previously successful authentication flow was not remembered.

• Fix an issue where trial logs may display timestamps twice.

Version 0.17.12

Release Date: March 28, 2022

New Features

• Job queue: Add support for dynamic job modification using the job queue. Users can use the WebUI or CLI to change the priority, weight, resource pool, and queue position of jobs without having to cancel and resubmit them. This feature is currently available for the fair share and priority schedulers. To update jobs through the WebUI, go to the Job Queue section and find the Manage Job option for a job. To update jobs using the CLI, use the det job update command. Run det job update --help for more information.

Breaking Changes

• API: Remove these legacy endpoints:

  • /:experiment_id

  • /:experiment_id/checkpoints

  • /:experiment_id/config

  • /:experiment_id/summary

  • /:experiment_id/metrics/summary

  • /:trial_id/details

  • /:trial_id/metrics

  The data from those endpoints are still available through the new REST API endpoints under the /api/v1/experiments/:experiment_id and /api/v1/trials/:trialᵢd prefixes.

Improvements

• Images: Update default environment images to PyTorch 1.10.2, TensorFlow 2.8, and Horovod 0.24.2.

Bug Fixes

• Database migrations: Ensure that migrations run in transactions. The lack of transactional migrations surfaced as a bug where, if the master was restarted during a migration, it would attempt to rerun the migration when it was already partially or wholly applied (but not marked as complete), resulting in various SQL errors on non-idempotent DDL statements.

• Distributed training: Allow multiple ranks within a distributed training job to report invalid hyperparameter exits. Previously, if more than one report was received, the experiment would fail.

Version 0.17.11

Release Date: March 14, 2022

New Features

• Add on_trial_startup() and on_trial_shutdown() methods to PyTorchCallback. Whenever on_trial_startup() is called, on_trial_shutdown() is always called before the trial container shuts down. These callbacks make it possible to do reliable resource management in a training container, such as if you wish to start a background thread or process for data loading and shut it down before the process exits.

Version 0.17.10

Release Date: March 03, 2022

Breaking Change

• API: PyTorch Lightning has been updated from 1.3.5 to 1.5.9 to address a security vulnerability. Experiments using PyTorch Lightning Adapter with v1.3.5 are no longer supported.

New Features

• Added PyTorch example using Bootstrap Your Own Latent (BYOL) to do self-supervised, no labels, image classification.

• PyTorchTrial and TFKerasTrial now automatically log the number of batches and number of records in every training and validation workload, as well as the duration of the workload and the calculated batches per second and records per second to make tracking progress easier.

• All (non-experiment) task logs are now persisted. Task logs can be retrieved through the new det task logs CLI command, or the WebUI or REST API. Task logs are now accessible even after a master restart, or 72 hours post completion.

• Support specifying root certificates for the DB via the Determined Helm chart. This allows Determined to use SSL to connect to the DB without having to replace the master config manually. To use this feature, save the certificate in a configmap or secret and set the following values: sslMode, sslRootCert, resourceType, and certResourceName. Additional details can be found in the default values.yaml file.

Version 0.17.9

Release Date: February 11, 2022

New Features

• Python API: Add new framework-specific methods for loading checkpoints:

  • determined.pytorch.load_trial_from_checkpoint_path()

  • determined.keras.load_model_from_checkpoint_path()

  • determined.estimator.load_estimator_from_checkpoint_path()

  These new methods are part of a larger effort to support more frameworks.

• Python API: Add on_training_epoch_end() method to PyTorchCallback. Add epoch_idx argument to on_training_epoch_start(). Overriding on_training_epoch_start without the epoch_idx argument is still supported for backward compatibility, but doing so is discouraged.

• Web: Add a column picker to the experiment list page to allow users to choose which table columns to display.

  

• Notebooks: Add a config field notebook_idle_type that changes how the idleness of a notebook is determined for the idle timeout feature. If the value is different from the default, users do not need to manually shut down kernels to allow the idle timeout to take effect.

• Web: Use the Page Visiblity API to detect changes in page visibility and avoid unnecessary polling, which can be expensive. While the user is not actively focused on the page, all polling is stopped; if the page becomes visible again, any previously active polling is restarted.

Improvements

• Breaking Change: CLI: The det master config command now takes the --json and --yaml options to configure its output format, rather than -o <output>.

• Breaking Change: API: The /api/v1/preview-hp-search endpoint no longer includes units (epochs/records/batches) in its response.

• API: The PATCH /api/v1/experiments/:id route no longer uses a field mask. When you include a field in the body (e.g., notes or labels) that field will be updated, if it is excluded then it will remain unchanged.

• API: When an experiment successfully completes, its progress value will be set to 100% instead of 0% or null; when an experiment fails, its progress value will stay the same instead of being reset to 0% or null.

• API: Calls to /api/v1/experiments and /api/v1/experiments/:id will return a progress value of null instead of 0 in cases where the progress has not been recorded or was reset to null.

Deprecations

• Python API: Checkpoint.load is deprecated. It should be replaced by determined.experimental.client.Checkpoint.download() along with the appropriate one of the new framework-specific functions for loading checkpoints.

• Python API: The following methods on objects in determined.experimental.client are formally deprecated (even though they were not technically public methods previously):

  • Model.from_json

  • Checkpoint.from_json

  • Checkpoint.parse_metadata

  • Checkpoint.get_type

  These methods will be removed in a future version.

Removed Features

• API: Remove /searcher/preview, /checkpoints, and /checkpoints/:checkpoint_id/* endpoints from the legacy API. These functions were already replaced by the gRPC API (/api/v1/preview-hp-search and /api/v1/checkpoints) in the web UI, CLI, and tests.

Version 0.17.8

Release Date: February 3, 2022

Bug Fixes

• Distributed Training: Fix a bug that shows experiments in a COMPLETED state even if they errored out. We recommend that users of distributed training update to 0.17.8 or later.

Version 0.17.7

Release Date: January 26, 2022

Breaking Changes

• API: Routes with /api/v1/models/:id/* are replaced by /api/v1/models/:name/*. Spaces and special characters in a name must be URI-encoded. You can get a model by ID with /api/v1/models?id=<id>.

• API: On the list of models (/api/v1/models) the optional name parameter is now a case-sensitive match, unless you add the parameter name_case_insensitive=true.

• Python API: determined.experimental.client.Determined.get_model() now takes a name rather than an ID. Use determined.experimental.client.Determined.get_model_by_id() to get a model from its ID.

• Model Registry: New model names must not be blank, have a slash, have multiple spaces, only numbers, or be case-insensitive matches to an existing model name.

• Model Registry: Model names with a forward slash will replace the slash in the name with ‘–’.

Bug Fixes

• Master: Fix a bug in the priority scheduler where jobs with equal priority would be scheduled or preempted in an order not correctly respecting job submission time.

Removed Features

• API: remove /experiment-list, /experiment-summaries, and /:experiment_id/kill endpoints from the legacy API. These functions are now replaced by the gRPC API (/api/v1/experiments) in the web UI, CLI, and tests.

Version 0.17.6

Release Date: January 20, 2022

New Features

• Master: Add support for systemd socket activation to the master.

• Scheduling/CLI: Add support for adjusting job priority and weight through the WebUI and CLI.

• Add experimental ROCm support. In the environment config for images and environment variables, the rocm key configures ROCm support. The gpu key has been renamed to cuda; gpu is still supported for backward compatibility, but its use is discouraged.

Improvement

• Docs: Improve many pages to address onboarding gaps.

Bug Fixes

• Master: Fix an issue where an update to an experiment’s name wouldn’t be reflected in its job representation until a master restart.

• Agent: Fix displayed CPU core count for CPU slots.

• WebUI: Fix an issue where the JupyterLab modal didn’t pass the full config.

• WebUI: Fix the issue of the profiler filter UI not triggering updates.

Improvements

• Logging: Decrease the volume of Docker image pull logs that are rendered into trial logs, and make the overall image pull progress more understandable by combining all layers’ progress into a single progress bar.

Deprecated Features

• Searcher: The Population Based Training searcher (pbt in the searcher config) will be removed in the next release.

• Model Registry: The API and Python interface will be returning to primarily identifying models based on their names, rather than their numeric IDs, in the next release.

Removed Features

• Remove support for Python 3.6, which has reached end-of-life.

Version 0.17.5

Release Date: December 10, 2021

New Features

• Add reporting of job queue state. The ordering of jobs in the queue and their status can be viewed through Determined WebUI, and CLI.

• WebUI: Add buttons to the WebUI to create new models in the Model Registry, as well as add checkpoints as versions to existing models. The Register Checkpoint modal can be accessed through the Checkpoint modal. The New Model modal can be accessed through the Register Checkpoint modal or on the Model Registry page.

  

• API: Add a method for listing trials within an experiment.

Improvements

• Agent: Improve handling of master connection failures.

Bug Fixes

• Deploy: Fix a bug where GCP clusters created with --no-filestore still had unused filestores created.

Version 0.17.4

Release Date: November 30, 2021

New Features

• WebUI: Add the model registry as a new top-level navigation option, allowing for viewing, editing, and deleting existing models created using the CLI.

• Add experimental support for Multi-Instance GPUs (MIGs) to agent-based setups, in parity with the experimental support for MIGs in Kubernetes-based setups. Static agents and Kubernetes clusters may be able to use MIG instances for some workloads. Distributed training is not supported, and all MIG instances and nodes within a resource pool must still be homogeneous.

Improvements

• Breaking Change: Model Registry: The names of models in the model registry must now be unique. If multiple models were previously created with the same name in the registry, the names will change.

• Model Registry CLI: Allow models to be referred to by their now-unique names, not only by ID.

• Tasks: Historical usage over users now properly accounts for all task types (commands, notebooks, etc.), not just trials.

• Images: Add environment images for TF 2.7.

• Agent: The environment.force_pull_image: true option no longer deletes the environment image before re-pulling it. Now, it will only fetch updated layers, which is much less wasteful of network resources and execution time.

Bug Fixes

• Master: Fix a bug where deleting experiments with trial restarts always failed, and then failed to be marked as failed.

Version 0.17.3

Release Date: November 12, 2021

Improvements

• Model Registry APIs: Add PATCH and DELETE endpoints to update the attributes of models and model versions.

• Model Registry: Allow models to be deleted only by the user who created them.

• Security and Logging: When a job is run on Kubernetes as a non-root user, the corresponding Fluent Bit sidecar will also run as a non-root user.

• Deploy: det deploy will now confirm potentially destructive updates on AWS unless --no-prompt is specified.

Bug Fixes

• Model Registry APIs: Change the /models/{}/versions/{} to accept model ID as an int.

Version 0.17.2

Release Date: October 29, 2021

New Features

• Model Registry APIs: Add new APIs to create a model with labels and to update the labels of an existing model.

Improvements

• Breaking Change: Deploy: det deploy now uses cloud images that use the NVIDIA Container Toolkit on agent hosts instead of relying on an older NVIDIA runtime, and custom images should be updated to do the same. Determined will no longer override the default container runtime according to the workload.

• Breaking Change: Model Registry APIs: Require name in the body rather than the URL for the post_model endpoint.

• Breaking Change: Model Registry APIs: Use model ID (integer) instead of name (string) as the lookup parameter for the get_model and get_model_versions endpoints.

• Docs: Switch to the Furo Sphinx theme, which fixes searching in the docs.

Bug Fixes

• Model Registry APIs: Sort models by name, description, and other attributes.

• Harness: Represent infinite and NaN metric values as strings in JSON.

• WebUI: Convert infinite and NaN value strings to numeric metrics.

• WebUI: Report login failures caused by the cluster being unreachable.

Version 0.17.1

Release Date: October 18, 2021

New Features

• WebUI: Add a “Notes” tab allowing for the input and viewing of free-form Markdown text on experiment pages. This works for both single-trial experiments and trials within a multi-trial experiment.

  

Improvements

• Docs: reorganize documents to be more user-friendly.

  • Merge some how-to guides, topic guides, and reference guides. Users should now need to read very few documents to understand what they need to do in Determined rather than having to jump around between documents.

  • Merge most information on best practices into how-to guides so that users find out about best practices as soon as they learn how to use something.

  • Decompose the top-level FAQ document and move different parts of it to relevant pages so that users can develop a better expectation of what common issues they might hit.

• Profiler: samples_per_second in PyTorch now reflects samples across all workers.

• Database migrations: Run upgrades in transactions to improve stability.

Bug Fixes

• Deploy: Fix an issue where the default checkpoint storage directory was not created for some users.

Version 0.17.0

Release Date: September 28, 2021

Breaking Changes

• Deploy: Remove --auto-bind-mount support from det deploy local. The new --auto-work-dir feature should be a strictly better experience. Users who depended on the shared_fs directory created by --auto-bind-mount can implement the same behavior by calling det deploy local cluster_up with a --master-config-path pointing to a master.yaml file containing the following text:

  task_container_defaults:
    bind_mounts:
      container_path: ./shared_fs
      host_path: /path/to/your/HOME/dir
  

• Deploy: This version of det deploy will not be able to deploy previous versions of Determined. If you need to deploy an older version, please use a matching version of the determined package.

• Experiment: Include maxval in int-type hyperparameter ranges. Previously, the docs said that the endpoints of the hyperparameter were both inclusive, but in reality the upper limit maxval was never actually selected.

  The reproducibility of hyperparameter selection may differ between Determined v0.16.5 and v0.17.0 for hyperparameter searches containing int-type hyperparameters as a result of this fix. However, the reproducibility of model training for any given set of hyperparameters should be unaffected.

• API: Endpoints no longer return the start times of workloads (training, validation, and checkpoints). This is part of a longer move to model metrics and workloads separately as part of the upcoming generic API.

• CLI: det master config now outputs YAML instead of JSON by default. To obtain the old behavior, run det master config -o json.

New Features

• Notebooks/TensorBoards: Support a configurable timeout field idle_timeout that will cause notebook and TensorBoard instances to automatically shut down after a period of idleness. A notebook is considered to be idle if no kernels or terminals are running and there is no network traffic going to the server. A TensorBoard is considered to be idle if there is no network traffic going to the server. Note that if you open a notebook file it might open a kernel for you, and the kernels and the terminals will not be shut down automatically. You need to manually shut down the kernels to make the idle timeout effective.

• Deploy: Add a new --auto-work-dir feature to det deploy local. Setting --auto-work-dir /some/path will have two effects: first, /some/path will be bind-mounted into the container (still as /some/path); second, interactive jobs (notebooks, shells, and commands) will run in the provided working directory by default. Note that containers run as the root user by default, so you may want to configure your user with det user such that interactive jobs run as your regular user.

• Commands/shells/notebooks: Support configuring the working directory using the work_dir configuration field for commands, shells, and notebooks. You can also optionally set it in the task_container_defaults.work_dir field of the master configuration. The value set in the master configuration will be ignored when a context directory is submitted.

• WebUI: Allow experiment owners to delete their own experiments, singly or in batches.

  

• WebUI: Display the latest log entry available for a trial at the bottom of the trial’s page. This works for both single-trial experiments and trials within a multi-trial experiment.

  

• WebUI: Add support for displaying NaN and Infinity metric values.

• Model Hub: Support the MMDetection library to easily train object detection models. MMDetection provides efficient implementations of popular object detection methods like Mask R-CNN, Faster R-CNN, and DETR on par with Detectron2. In addition, cutting-edge approaches from academia are regularly added to the library.

• Deploy: Add the ability to use customizable master configuration templates in det deploy aws|gcp.

• Images: Add an environment image for CPU-only TensorFlow 2.5 and 2.6.

Improvements

• API: The aggregated historical resource allocation APIs /api/v1/resources/allocation/aggregated and /allocation/aggregated now account for all resources, not just those allocated to experiments.

• Images: Add CPU-only images for TF 2.5 and 2.6.

• Images: Upgrade JupyterLab to version 3.1.

• Images: TF 2.5 and 2.6 images will no longer include PyTorch builds. For PyTorch 1.9, please use the combined TF 2.4/PyTorch 1.9 image.

• Images: TF 2.4, 2.5, 2.6, and PyTorch 1.9 images will now use Python 3.8. The legacy TF 1.15/PyTorch 1.7 image will continue to use Python 3.7.

Changes

• WebUI: Change the task list page to open new tabs when user clicks on task links.

• WebUI: The trial detail page will no longer show workload-based start time information, including training time, validation time, and checkpoint time.

Bug Fixes

• WebUI: Fix continuing trials with nested hyperparameters.

Version 0.16

Version 0.16.5

Release Date: September 3, 2021

New Features

• Support custom PyTorch data loaders with PyTorchTrial. You may now call context.experimental.disable_dataset_reproducibility_checks() in your trial’s __init__() method, which will allow you to return arbitrary DataLoader objects from build_training_data_loader() and build_validation_data_loader(). This is desirable when your data loader is not compatible with Determined’s det.pytorch.DataLoader. The usual dataset reproducibility that det.pytorch.DataLoader provides is still possible to achieve, but it is your responsibility. You may find the Sampler classes in determined.pytorch.samplers to be helpful.

Improvements

• Add the ability to disable agents while allowing currently running tasks to finish using det agent disable --drain AGENT_ID.

Bug Fixes

• WebUI: Show metrics with a value of 0 in graphs.

• Properly load very old (pre-0.13.8) checkpoints with TFKerasTrial.

Version 0.16.4

Release Date: August 23, 2021

New Features

• WebUI: Add a trial comparison modal, allowing comparison of information, metrics, and hyperparameters between specific trials within an experiment. This is available from the experiment trials and experiment visualization pages.

• Scheduling/CLI: Support changing task priorities using the det experiment/command/notebook/shell/tensorboard set priority commands.

• CLI: Allow command-line config overrides in experiment creation, e.g., det e create const.yaml . --config key=value.

• WebUI: Allow cluster admins to delete individual experiments.

Bug Fixes

• Cluster: Fix breakage in trial fault tolerance caused by not sending enough state snapshots.

• WebUI: Prevent logs from potentially introducing harmful HTML/JS injections via Unicode.

• WebUI: Change y-axis of the profiler timing metrics chart from milliseconds to seconds.

• WebUI: Prevent the zoom from resetting when chart data series are added.

• WebUI: Fix the issue of learning curves not resizing properly.

Version 0.16.3

Release Date: July 22, 2021

New Features

• Add the ability to use Azure Blob Storage for checkpoint storage.

• Add support for Azure Kubernetes Service, including updating Helm to support Azure Blob Storage and adding additional docs for AKS.

• WebUI: Add support for nested hyperparameters in experiment config, trial hyperparameters, and hyperparameter visualization.

• WebUI: Add the ability to view trial logs and open TensorBoards directly from the trial list view.

• WebUI: Enable sorting and filtering trials by state on an experiment’s trials page.

• WebUI: Add a server availability check on load.

Bug Fixes

• Fix a bug where experiments with model definitions exceeding 50% of the maximum allowable size would cause trials to never start.

• WebUI: Prevent hyperparameter visualization from getting stuck showing a spinner after clicking through all the different tabs.

• WebUI: Fix the issue of experiments showing incorrect data if they were forked from another experiment or continued from a trial.

Version 0.16.2

Release Date: July 9, 2021

New Features

• Make det deploy aws up automatically bind-mount FSx and EFS directories into task containers when available.

• Make det deploy local bind-mount the user’s home directory into task containers. The mounted directory can be changed with the --auto-bind-mount=<path> option and mounting can be disabled entirely with --no-auto-bind-mount.

Improvements

• PyTorchTrial: Improve support for custom batches in PyTorch, e.g., as used in pytorch_geometric. See get_batch_length() or examples/graphs/proteins_pytorch_geometric for further details.

Bug Fixes

• WebUI: Avoid waiting for an extra polling cycle to load trial data when loading single-trial experiments.

• WebUI: Fix an issue with boolean hyperparameter values not being rendered in learning curve tables.

Version 0.16.1

Release Date: June 28, 2021

New Features

• Add support for CPU-based training. This makes it possible to run Determined on clusters without GPUs, including on-prem, AWS, GCP, and Kubernetes-based (default scheduler only) configurations.

• Support spinning up and down a Filestore instance when running det deploy gcp up/down. The Filestore instance will automatically be mounted to agents and bind-mounted into task containers. You can also use a pre-existing Filestore instance.

Improvements

• Breaking Change: REST API: Rename gpu and cpu fields in ResourcePool object to compute and aux.

• Breaking Change: Deploy: In det deploy gcp and det deploy aws, rename the default compute pool from gpu-pool to compute-pool. When upgrading a cluster from a previous version, existing pending experiments may error out and need to be resubmitted.

Bug Fixes

• Support using Docker images with EXPOSE commands as images for notebooks/shells/TensorBoards. Previously, the EXPOSE command could break proxying through the Determined master.

Version 0.16.0

Release Date: June 14, 2021

New Features

• Python SDK: Extend the Checkpoint Export API into a Python SDK capable of launching and controlling experiments on the cluster directly from Python. See the documentation and examples in the client module.

• Trials: Add new support for profiling model code. For all frameworks, collecting system metrics, such as GPU utilization and memory, is supported. For PyTorch, additional profiling for timing is available. To quickly try out profiling, set profiling.enabled = true in the experiment configuration.

• Experiments: Add new notes and name fields to experiments.

• REST API: Add new parameters to /api/v1/experiments to filter and sort experiments by name.

• Master configuration: Support bind_mounts in task_container_defaults in the master configuration. The configured directories will be mounted for experiments, notebooks, commands, shells, and TensorBoards.

• Images: Add an environment image containing TensorFlow 2.5 and CUDA 11.2.

Improvements

• Breaking change: JupyterLab: Upgrade the JupyterLab version to 3.0.16. JupyterLab will no longer work with previously released images. Custom image users should upgrade to JupyterLab 3.0 or higher.

• Scheduling: Support backfilling in the priority scheduler. If there are slots that cannot be filled with high-priority tasks, low-priority tasks will be scheduled onto them. This requires preemption to be enabled in the master configuration.

• WebUI: Improve task list filtering by moving column filters to the table header.

• REST API: Change filtering experiments by description to be case-insensitive when using the /api/v1/experiments endpoint.

Bug Fixes

• Fix a bug where InvalidHP exceptions raised in the trial __init__() caused the trial to restart.

• WebUI: Fix an issue with representing some hyperparameter values as text.

• Kubernetes: Prevent Determined from sometimes crashing when handling concurrent job submissions.

• Master configuration: Fix a bug that was triggered when the master configuration had S3 secrets explicitly configured in checkpoint_storage. Experiments that did not override the master-provided checkpoint storage would fail.

Deprecated Features

• The method create_trial_instance() is now deprecated. Users should instead use the more flexible TrialContext.from_config(), which is described in How to Debug Models.

Removed Features

• The methods det.experimental.keras.init() and det.experimental.estimator.init() have been removed. They were deprecated in 0.13.5.

Version 0.15

Version 0.15.6

Release Date: June 2, 2021

New Features

• Add PyTorch’s word-level language modeling RNN example as a Determined example.

• Support using the Determined shell as a remote host inside Visual Studio Code and PyCharm IDEs.

Improvements

• Deploy: Add support for terraform 0.15 when using det deploy gcp.

• REST API: Add a preview parameter to the Notebook launch API (POST /api/v1/notebooks). If set, this API will return a full configuration that is populated with the template and Notebook configuration.

• WebUI: Improve Experiment list filtering by moving column filters to the table header.

• WebUI: Improve the Trial details page by moving hyperparameters, workloads, and logs into separate “tabs” on the Trial detail page.

Bug Fixes

• PyTorchTrial: Fix an issue where a DataLoader iterator that uses multiprocessing could cause a hang when exiting.

• WebUI: Prevent TQDM log lines from generating large quantities of whitespace when rendering logs.

Version 0.15.5

Release Date: May 18, 2021

Bug Fixes

• Fix an issue where the master would attempt to schedule onto agents that had previously disconnected.

Deprecated Features

• Deprecate the scheduler and provisioner fields in the master configuration in favor of resource_manager and resource_pools. They will be removed in the next minor release, Determined 0.16.0.

Version 0.15.4

Release Date: May 12, 2021

New Features

• Model Hub: Publish Determined’s Model Hub library to make it easy to train models from supported third-party libraries with a Determined cluster. The first library supported in Model Hub is the HuggingFace transformers library for NLP.

Minor Changes

• API: Remove redundant APIs for commands, shells, TensorBoards, and notebooks. The CLI now uses updated versions of these endpoints; related CLI commands on versions 0.15.4 and beyond are not backward-compatible with previous versions of Determined clusters.

• API: Update the trial detail endpoint (GET /api/v1/trials/:id), dropping prior_batches_processed and num_inputs in favor of total_batches.

Version 0.15.3

Release Date: May 5, 2021

Bug Fixes

• Images: Fix GPU support in CUDA 10.2 + TensorFlow 1.15 images.

• Trials: Update to match websockets>= 9.0 library API change.

• Trials: Fix a bug that caused trials to panic upon receiving too many rendezvous addresses.

Version 0.15.2

Release Date: April 29, 2021

New Features

• Kubernetes: Support priority scheduling with preemption. The preemption scheduler is able to preempt experiments when higher priority ones are submitted.

• APIs: Promote the custom metric reducer APIs for both pytorch and estimators from experimental status to general availability.

• Resource pools: Support configuring distinct task_container_defaults for each resource pool configured on the cluster. This can allow different resource pools which may have very different hardware to configure tasks in each pool with the correct settings.

Improvements

• Agent: Support configuring the name of the Fluent Bit logging container via the --fluent-container-name option.

• Docker: Support specifying the --devices, --cap-add, and --cap-drop arguments to the docker run command. These are configured in an experiment or command/notebook config via resources.devices, environment.add_capabilities, and environment.drop_capabilities. These settings can combine to allow an experiment to take advantage of cluster hardware not previously available to training or notebook task. These configurations are only honored by resource managers of type agent, and are ignored by resource managers of type kubernetes.

Bug Fixes

• Agent: Support the --fluent-port option.

• PyTorchTrial: Fix learning rate scheduler behavior when used with gradient aggregation.

• PyTorchTrial’s to_device() no longer throws errors on non-numeric Numpy-like data. As PyTorch is still unable to move such data to the GPU, non-numeric arrays will simply remain on the CPU. This is especially useful to NLP practitioners who wish to make use of Numpy’s string manipulations anywhere in their data pipelines.

• TFKerasTrial: Fix support for TensorFlow v2.2.x.

• WebUI: Fix the issue of the WebUI crashing when user selects a row in experiment list page then changes the user filter.

• WebUI: Fix the issue of agents overview on the Cluster page not updating properly when agents shutdown.

• WebUI: Fix the issue of trial logs not rendering properly on Safari 14.

Version 0.15.1

Release Date: April 16, 2021

Bug Fixes

• Trials: Fix TFKerasTrial on TensorFlow 2 with disabled v2 behavior and/or disabled eager execution.

• Master: Fix two issues that caused experiments to not recover successfully on master crashes (after upgrading to version 0.15.0).

Version 0.15.0

Release Date: April 14, 2021

New Features

• WebUI: Provide historical allocation data on the Cluster page. This page breaks down GPU hours by user, label, and resource pool.

• WebUI: Add a gallery mode to the hyperparameter scatter plot and heatmap visualizaztions to allow users to inspect each scatter plot in full detail.

Improvements

• Breaking Change CLI: Consolidate det and det-deploy executables into the determined package, which now includes all Determined libraries and tools. The determined-cli, determined-deploy, and determined-common packages are now deprecated.

  • When upgrading from older versions, det command may break for some users because of pip limitations. Please uninstall outdated packages, and then reinstall Determined.

• PyTorch: Remove cloudpickle as a dependency for PyTorch checkpoints. This does not affect compatibility of existing checkpoints. This change will improve portability across Python versions.

• Deploy: Move default storage location for checkpoint data in local clusters deployed via det deploy local to an OS-specific user data directory (e.g. $XDG_DATA_HOME/determined or ~/.local/share/determined on Linux, and ~/Library/Application Support/determined on macOS). Previously, /tmp was used. This location can be changed using the --storage-host-path command line flag of det deploy local. If users provide their own custom master.yaml via --master-config-path, the configured checkpoint_storage in master.yaml will take precedence.

• Searcher: Remove support for adaptive and adaptive_simple searchers which were deprecated in Determined 0.13.7.

Bug Fixes

• WebUI: Fix an issue where the metric value occasionally had the word “undefined” prepended.

Version 0.14

Version 0.14.6

Release Date: April 1, 2021

New Features

• REST API: Add a new endpoint to delete experiments. This endpoint is only enabled for admin users and deletes all resources associated with an experiment. This includes checkpoint storage, TensorBoards, trial logs from all backends and metadata such as history and metrics, stored in PostgreSQL.

• REST API: Add a new endpoint to fetch aggregated historical resource allocation information.

• CLI: Add new commands det resources raw and det resources aggregated to access resource allocation information.

• PyTorch Lightning: Add an adapter to support LightningModule from PyTorch Lightning in the PyTorchTrial API.

Improvements

• Images: The default environment images have been updated to CUDA 10.2, PyTorch 1.8, and TensorFlow 1.15.5 with Python 3.7. Previous images are still available but must be specified in the experiment or command configuration. It is recommended to validate the performance of models when changing CUDA versions as some models can experience significant changes in training time, etc.

• WebUI: Improve the hyperparameter scatter plot and heat map visualizations by adding support for showing categorical hyperparameters.

Bug Fixes

• WebUI: Fix the hyperparameter visualization page crashing when viewing single trial or PBT experiments, both of which are intentionally unsupported for hyperparameter visualizations.

Version 0.14.5

Release Date: March 18, 2021

New Features

• REST API: Add a REST API endpoint exposing historical cluster resource allocation. Currently, information about experiment workloads (training, checkpoints, and validations) is included.

• Hyperparameter Search: Introduce a stopping-based variant of Adaptive (Asynchronous) Method that will continue training trials by default unless stopped by the algorithm. Compared to the default promotions-based algorithm, the stopping variant will promote to higher rungs faster and does not require fault tolerance since it will not resume stopped trials.

• PyTorch: Add an option to LRScheduler to accept a frequency option alongside batch and epoch step modes.

• Kubernetes: Add support for priority scheduling, with gang-scheduling for distributed training, on Kubernetes.

Improvements

• WebUI: Add a margin of comparison to hyperparameter visualizations to enable better grouping of trials with a similar but not identical number of batches processed.

Bug Fixes

• Correct model code uploaded to checkpoints so it now matches the model code provided during experiment creation. Previously, it may have included additional files that had been bind-mounted with a container_path that was either relative or was a subdirectory of /run/determined/workdir.

• Fix an unauthorized access issue when attempting to use the Determined CLI within a notebook.

Version 0.14.3

Release Date: March 4, 2021

New Features

• Examples: Add the Deformable DETR model for object detection in Determined. Check out our example in the Determined repository.

• Searcher: Support programmatic rejection of certain hyperparameters to further optimize your hyperparameter search.

• WebUI: Add additional hyperparameter visualizations to multi-trial experiments. The new parallel coordinate, scatter plot, and heat map visualizations will allow you to better explore relationships between hyperparameters and model performance.

Improvements

• WebUI: Use anchor tags instead of click event listeners across all table rows. This increases accessibility and improves keyboard navigation support.

Bug Fixes

• Keras: Ensure that keras.utils.Sequence objects receive their on_epoch_end() calls after validation is completed.

• WebUI: Fix the order of batches to be numeric instead of alphanumeric.

Version 0.14.2

Release Date: February 17, 2021

New Features

• Support CUDA 11. New Docker images are available for experiments and commands to support CUDA 11, as well as some updated versions of frameworks on CUDA 10.1. It is recommended to validate the performance of models when changing CUDA versions as some models can experience significant changes in training time, etc.

• Support startup-hook.sh for notebooks and shells. This is the same mechanism supported by experiments.

Improvements

• Improve local test mode for experiment creation, det experiment create, to test with only a single batch.

• Invoke python as python3 rather than as python3.6. This makes it possible to use custom images containing higher versions of Python with Determined (3.6 is still the minimum required version).

  If the desired python cannot be found as python3, it is now possible to customize this invocation by setting the environment variable DET_PYTHON_EXECUTABLE=/path/to/python3, for experiments, notebooks, and shells.

Bug Fixes

• Kubernetes: Fix a bug that caused the Cluster page to not render when using a Kubernetes cluster.

Version 0.14.1

Release Date: February 9, 2021

Bug Fixes

• Trial: Fix a bug that prevented trial logs created before 0.13.8 from loading correctly.

Version 0.14.0

Release Date: February 4, 2021

New Features

• Add resource pools, which allows for different types of tasks to be scheduled onto different types of agents.

• det-deploy will now create clusters with two resource pools, one that uses GPU instances and one that uses CPU instances for tasks that only require CPUs.

• WebUI: Revamp cluster page with information about configured resource pools.

Removed Features

• Trial API: Remove the old PyTorch APIs, including:

  • the build_model, optimizer, and create_lr_scheduler methods in PyTorchTrial;

  • the callback on_before_optimizer_step;

  • the field optimizations.mixed_precision in the experiment configuration;

  • the model arguments to train_batch(), evaluate_batch(), and evaluate_full_dataset().

  Model code that uses these APIs will no longer run in Determined 0.14.0 or later. However, model checkpoints produced by old experiments that used these APIs will still be supported.

Improvements

• Breaking Change REST API: The trial and checkpoint API endpoints can now return non-scalar metric values, which are represented as JSON objects or protobuf structs.

• Documentation: Add a topic guide on debugging models. The new guide will walk you step-by-step through solving problems with a model in Determined, with a focus on testing features incrementally until the model is fully working. It may also be useful when porting new models to Determined.

• Documentation: Add a topic guide on commands and shells. It describes how to use Determined’s support for managing GPU-powered batch commands and interactive shells.

• REST API: Improve the performance of the experiments API.

Bug Fixes

• Database: Migrate public.trial_logs.id to be an int8 in Postgres, instead of an int4. This avoids issues for customers with extremely large amounts of trial logs. Note: This migration will be more time-consuming than usual for deployments with large amounts of trial logs.

• REST API: Fix an issue where requesting checkpoint or trial details of a trial that had non-scalar metric values associated with it would fail.

• Trial: Fix an issue where the trial was not deallocating resources when it failed to write to the DB.

• WebUI: Show better messaging for different learning curve edge cases.

• WebUI: Fix sorting on the experiment trials table within the experiment detail page.

• WebUI: Fix issue of incorrect trial log order when viewing oldest logs first.

• WebUI: Update Cancel confirm button label to show Confirm to avoid double Cancel buttons.

• WebUI: Improve the sorting behavior for numeric table columns.

Version 0.13

Version 0.13.13

Release Date: January 25, 2021

New Features

• Update experiment details pages to include a learning curve visualization. This will enable a comparison of hyperparameter performance among many different trials within an experiment.

• Support Elasticsearch as an alternative backend for logging. Read more about Elasticsearch-backed logging to see if it’s appropriate for your Determined deployment.

Improvements

• Breaking Change: REST API: Update trial logs API to return string IDs.

• WebUI: Enable filtering of trial logs by agent, container, rank, log level, and timestamp.

• WebUI: Improve section contrast on all pages.

• Deployment: Add the command det-deploy aws list, which shows all the CloudFormation stacks that are managed by det-deploy aws (using the tag managed-by: determined). This only applies to new deployments since this version, not previous deployments.

• Update examples to use the new PyTorch APIs.

Deprecated Features

• The old PyTorch API was deprecated in 0.12.13 and will be removed in the next release. See the PyTorch migration guide for details on updating your PyTorch model code to use the new API.

Version 0.13.12

Release Date: January 11, 2021

Bug Fixes

• WebUI: Fix the Okta sign-in workflow.

• WebUI: Fix an issue with unexpected hyperparameter types in experiment configuration.

• WebUI: Fix trial metric workload duration reporting in the trial detail page.

Version 0.13.11

Release Date: January 6, 2021

Improvements

• Trials: Add experimental support for custom metric reducers with PyTorchTrial. This enables calculating advanced metrics like F1 score or mean IOU; returning multiple metrics from a single reducer is also supported. See determined.pytorch.PyTorchExperimentalContext.wrap_reducer() for detailed documentation and code snippets.

  See determined/examples/features/custom_reducers_mnist_pytorch for a complete example of how to use custom reducers. The example emits a per-class F1 score using the new custom reducer API.

• Trials: Support more than 1 backward pass per optimizer step for distributed training in PyTorchTrial.

• Logging: Allow the trial logging backend to be configured in Kubernetes-based deployments of Determined.

• Agents: Add support for labels when starting agents with det-deploy.

Bug Fixes

• WebUI: Update the Trial Information Table to be usable on mobile devices.

• HP Search: Fix a bug where adaptive_asha could run with more maximum concurrent trials than intended.

• Scheduling: Fix a bug where command priority was not respected.

Version 0.13.10

Release Date: December 10, 2020

New Features

• WebUI: Add support for mobile and tablet devices. Check your experiment results on the go!

• Scheduler: Update the priority scheduler to support specifying priorities and preemption.

Improvements

• Improve the scheduling and scaling behavior of CPU tasks, and allow the maximum number of CPU tasks per agent to be configured via the Step 7 - Configure the Cluster.

• Add custom tagging support to AWS dynamic agents. Thank you to sean-adler for contributing this improvement!

• Support validation_steps in TFKerasTrial’s context.configure_fit(). validation_steps means the same thing in Determined as it does in model.fit(), and has the same limitation (in that it only applies when validation_data is of type tf.data.Dataset).

• Kubernetes: Support a default user password for Kubernetes deployments. This affects the admin and determined default user accounts.

• Kubernetes: Release version 0.3.1 of the Determined Helm chart.

Bug Fixes

• Fix a bug in --local --test mode where all GPUs were being passed to the training loop despite the distributed training code paths being disabled.

• Fix a bug causing active trials that have failed to not be restored properly on a master restart when max_restarts is greater than 0.

• Allow configurations with a . character in the keys for map fields in the Master Configuration Reference (e.g. task_container_defaults.cpu_pod_spec.metadata.labels).

• Fix a bug where restoring a large number of experiments after a failure could lead to deadlock.

• Fix an issue where templates with user-specified bind mounts would merge incorrectly. Thank you to zjorgensenbits for reporting this issue!

Deprecated Features

• The previous version of the priority scheduler is now deprecated. It will remain available as the round_robin scheduler for a limited period of time.

Version 0.13.9

Release Date: November 20, 2020

Improvements

• Commands: Support configuring shmSize for commands (e.g., notebooks, shells, TensorBoards) in command configurations.

Bug Fixes

• API: Fix a bug that caused the WebUI’s log viewer to fail to render previous pages of trial logs.

• WebUI: Fix a bug in opening TensorBoards from the experiment list page via batch selection.

Version 0.13.8

Release Date: November 17, 2020

New Features

• API: Add support for models that subclass tf.keras.Model when using the Determined TFKerasTrial API. This is a new feature that became available starting in TensorFlow 2.2, allowing user to further customize their training process.

• Deployment: When using the simple deployment type with det-deploy aws, you can now use the --agent-subnet-id flag to specify which existing subnet to launch agents in. As each subnet is associated with a single availability zone, this allows users to explicitly choose an availability zone that has GPU instances (there is no public information about which availability zones have GPU instances so trial and error is the suggested approach).

• Logs: Support filtering trial logs by individual fields in the CLI. Log entries for trials can now be filtered by container ID, agent ID, log level, and other fields.

• Security: Allow the master to use a TLS certificate that is valid for a different name than the agents use to connect to it. This ability is useful in situations where the master is accessed using multiple different addresses (e.g., private and public IP addresses of a cloud instance). The agent now accepts a --security-tls-master-cert-name option to override the expected name in the master’s TLS certificate. The CLI uses the DET_MASTER_CERT_NAME environment variable for the same purpose.”

Improvements

• Breaking Change: API: Perform salting and hashing on server-side for the password change endpoint. This makes this endpoint consistent with the new login endpoint described at https://docs.determined.ai/latest/rest-api/ .

• Breaking Change: Logging: Start using Fluent Bit for handling trial logs internally. The agent machines now need to have access to the fluent/fluent-bit:1.6 Docker image. If the Determined agent machines are able to connect to Docker Hub, they will pull it automatically and no changes are required; if not, the image must be manually made available beforehand. The Determined agent accepts a --fluent-logging-image option to specify an alternate name for the image. This change is part of an effort to improve the handling of trial logs by increasing scalability and allowing more options for log storage.

• Agent: Support configurable slot types for agents. Previously, Determined only supported auto-detecting the slot type for agents. If Determined did not detect any GPUs, the agents would fall back to mapping one slot to all the CPUs. With this change, this behavior can be configured to one of auto, gpu, and none in the field slot_type of the agent configuration agent.yaml. Dynamic agents having GPUs will be configured to gpu while those agents having no GPUs will be configured to none. For static agents this field defaults to auto.

• API: Add self.context.wrap_optimizer() to the Determined TFKerasTrial API.

• API: Add tf.keras DCGAN example that subclasses tf.keras.Model.

• API: Add self.context.configure_fit() to the Determined TFKerasTrial API. Many parameters which would be passed to model.fit(), such as class_weight, verbose, or workers, can now be passed to configure_fit() and will be honored by TFKerasTrial.

• Kubernetes: Add option to configure the service type of the Determined deployed database in the Determined Helm chart. This is useful if your cluster does not support ClusterIP, which is the service type that is used by default.

• WebUI: Make the page/tab title more descriptive.

• WebUI: Add navigation sidebar, breadcrumb, and back buttons to log view pages.

• WebUI: Update the trial and master log buttons to open in the same page by default, with the option to open in a new tab.

• WebUI: Update trial details URL to include the experiment id.

Bug Fixes

• API: Fix support for Keras Callbacks.

  • Previously, stateful Keras Callbacks (EarlyStopping and ReduceLROnPlateau) did not work in Determined across pause/activate boundaries. We have introduced Determined-friendly implementations, determined.keras.callbacks.EarlyStopping and determined.keras.callbacks.ReduceLROnPlateau, which address this shortcoming. User-defined callbacks may subclass determined.keras.callbacks.Callback (and define get_state and load_state methods) to also benefit from this and other new features.

  • Previously, Keras Callbacks which relied on on_epoch_end in Determined would see their on_epoch_end called every scheduling_unit batches by default. Now, on_epoch_end will be reliably called at the end of each epoch, as defined by the records_per_epoch setting in the experiment config. As before, on_epoch_end will not contain validation metrics, as the validation data is not always fresh at epoch boundaries. Therefore, the Determined implementations of EarlyStopping and ReduceLROnPlateau are both based on on_test_end, which can be tuned using min_validation_period.

• API: Fix issue that occasionally made TFKerasTrial hang for multi-GPU training during COMPUTE_VALIDATION_STEP.

• Kubernetes: Gracefully handle cases where the Kubernetes API server responds with unexpected object types.

• Scheduler: Fix not being able to find resource pools for experiments.

• Scheduler: Fix not being able to disable slots.

• WebUI: Prevent navigation item tooltips from showing up when hovering outside of the navigation bar.

• WebUI: Fix an issue where the experiment archive action button was out of sync.

• WebUI: Fix experiment actions to not display a loading spinner.

Deprecated Features

• API: Deprecate the name det.keras.TFKerasTensorBoard in favor of det.keras.callbacks.TensorBoard. The old name will be removed eventually, and user code should be updated accordingly.

• API: Deprecated the old det.keras.SequenceAdapter. SequenceAdapter will be removed in a future version. Users should use self.context.configure_fit() instead, which is both more capable and more similar to the normal tf.keras APIs.

Version 0.13.7

Release Date: October 29, 2020

New Features

• Add support for running workloads on spot instances on AWS. Spot instances can be up to 70% cheaper than on-demand instances. If a spot instance is terminated, Determined’s built-in fault tolerance means that model training will continue on a different agent automatically. Spot instances can be enabled by setting spot: true in the Step 7 - Configure the Cluster.

• Support MMDetection, a popular library for object detection, in Determined. MMDetection allows users to easily train state-of-the-art object detection models; with Determined, users can take things one step further with cutting-edge distributed training and hyperparameter tuning to further boost performance. See the Determined implementation of MMDetection for more information on how to get started.

• WebUI: Allow the experiments list page to be filtered by labels. Selecting more than one label will filter experiments by the intersection of the selected labels.

Deprecated Features

• Deprecate the simple and advanced adaptive hyperparameter search algorithms. They will be removed in a future release. Both algorithms have been replaced with Adaptive (Asynchronous) Method, which has state-of-the-art performance, as well as better scalability and resource-efficiency.

Improvements

• Documentation: Add a guide for Set up and Manage an AWS Kubernetes (EKS) Cluster.

• Master: Support a minimum instance count for dynamic agents. The master will attempt to scale the cluster to at least the configured value at all times. This is configurable via provisioner.min_instances in the Step 7 - Configure the Cluster. This will increase responsiveness to workload demand because agent(s) will be ready even when the cluster is idle.

• Kubernetes: Improve the performance of the /agents endpoint for Kubernetes deployments. This will improve the performance of the cluster page in the WebUI, as well as when using det slot list and det task list via the CLI.

• Kubernetes: Release version 0.3.0 of the Determined Helm chart.

• WebUI: Improve metric selection on the trial detail page. This should improve filtering for trials with many metrics.

• WebUI: Use scientific notation when appropriate for floating point metric values.

• WebUI: Show both experiment and trial TensorBoard sources when applicable.

Bug Fixes

• WebUI: Fix an issue where TensorBoard sources did not display properly for TensorBoards started via the CLI.

• WebUI: Fix an issue with rendering boolean hyperparameters in the WebUI.

• CLI: Fix an issue where trial IDs were occasionally not displayed when running det task list or det slot list in the CLI.

• Master: Fix the default value for the fit field if the scheduler is set in the Step 7 - Configure the Cluster.

Version 0.13.6

Release Date: October 14, 2020

Improvements

• Agent: The boot_disk_source_image field for GCP dynamic agents and image_id field for AWS dynamic agents are now optional. If omitted, the default value is the Determined agent image that matches the Determined master being used.

• Documentation: Ship Swagger UI with Determined documentation. The /swagger-ui endpoint has been renamed to /docs/rest-api.

• Documentation: Add a guide on configuring TLS in Determined.

• Kubernetes: Add support for configuring memory and CPU requirements for the Determined database when installing via the Determined Helm chart.

• Kubernetes: Add support for configuring the storageClass that is used when deploying a database using the Determined Helm chart.

Bug Fixes

• Harness: Do not require the master to present a full TLS certificate chain when the certificate is signed by a well-known Certificate Authority.

• Harness: Fix a bug which affected TFKerasTrial using TensorFlow 2 with gradient_aggregation > 1.

• Master: Fix a bug where the master instance would fail if an experiment could not be read from the database.

• WebUI: Preserve the colors used for multiple metrics on the metric chart.

• WebUI: Fix the ability to cancel a batch of experiments.

• WebUI: Fix a bug which caused the Experiment Details page to not render when the latest validation metric is not available.

Version 0.13.5

Release Date: September 30, 2020

Improvements

• Security: Use one TCP port for all incoming connections to the master and use TLS for all connections if configured.

  • Breaking Change: The http_port and https_port options in the master configuration have been replaced by the single port option. The security.http option is no longer accepted; the master can no longer be configured to listen over HTTP and HTTPS simultaneously.

• Security: Support configuring TLS encryption when deploying Determined on Kubernetes.

• Agent: Increase default max agent starting and idle timeouts to 20 minutes and increase max disconnected period from 5 to 10 minutes.

• Deployment: Add support for det-deploy aws in the following new regions: ap-northeast-1, eu-central-1, eu-west-1, us-east-2.

• Docker: Publish new Docker task containers that upgrade TensorFlow versions from 1.15.0 to 1.15.4, and 2.2.0 to 2.2.1.

• Documentation: Add extra documentation and reorganize examples by use case.

• Documentation: Add a tf.layers-in-Estimator example.

• Kubernetes: Add support for users to specify initContainers and containers as part of their custom pod specs.

• Kubernetes: Publish version 0.2.0 of the Determined Helm chart.

• Native API: Deprecate Native API. Removed related examples and docs.

• Trials: Remove support for TensorpackTrial.

• WebUI: Improve polling behavior for experiment and trial details pages to avoid hanging indefinitely for very large experiments/trials.

Bug Fixes

• Trials: Fix a bug where if only a subset of workers on a machine executed the on_trial_close() EstimatorTrial callback, the container would terminate as soon as one worker exited.

• Trials: Fix a bug where det e create --test would succeed when there were checkpointing failures.

• WebUI: Fix the issue of multiple selected rows dissappearing after a successful table batch action.

• WebUI: Remove unused TensorBoard sources column from the task list page.

• WebUI: Fix rendering metrics with the same name on the metric chart.

• WebUI: Make several fixes to improve select appearance and user experience.

• WebUI: Fix the issue of agent and cluster info not loading on slow connections.

• WebUI: Fix the issue where the chart in the Experiment page does not have the metric name in the legend.

Version 0.13.4

Release Date: September 16, 2020

Improvements

• Support configuring default values for the task image, Docker pull policy, and Docker registry credentials via the Master Configuration Reference and the Helm Chart Configuration Reference. In previous versions of Determined, these values had to be specified on a per-task basis (e.g., in the experiment configuration). Per-task configuration is still supported and will overwrite the default value (if any).

• Add connection checks for dynamic agents. A dynamically provisioned agent will be terminated if it is not actively connected to the master for at least five minutes.

• Emit a warning if DistributeConfig is specified for an Estimator. Configuring an Estimator via tf.distribute.Strategy can conflict with how Determined performs distributed training. With this change, Determined will attempt to catch this problem and surface an error message in the experiment logs. An Estimator can still be configured with an empty DistributeConfig without issue.

• Remove support for dataflow_to_tf_dataset in EstimatorTrial. Dataflows should be wrapped using wrap_dataset(shard=False) instead.

• WebUI: Add middle mouse button click detection on tables to open in a new tab/page.

• WebUI: Improve the trial detail metrics view.

  • Support metrics with non-numeric values.

  • Default to showing only the searcher metric on initial page load.

  • Add search capability to the metric select filter. This should improve the experience when there are many metrics.

  • Add support for displaying multiple metrics on the metric chart.

• WebUI: Move TensorBoard sources from a table column into a separate modal.

• WebUI: Optimize loading of active TensorBoards and notebooks.

Bug Fixes

• Improve handling of certain corner cases where distributed training jobs could hang indefinitely.

• Fix an issue where detecting GPU availability in TensorFlow code would cause EstimatorTrial models to OOM.

• Fix an issue where accessing logs could create a memory leak.

• Fix an issue that prevents resuming from checkpoints that contain a large number of files.

• WebUI: Fix an issue where table page sizes were not saved between page loads.

• WebUI: Fix an issue where opening a TensorBoard on an experiment would not direct the user to an already running TensorBoard, but instead create a new one.

• WebUI: Fix an issue where batch actions on the experiments table would cause rows to disappear.

Known Issues

• WebUI: In the trial detail metrics view, experiments that have both a training metric and a validation metric of the same name will not be displayed correctly on the metrics chart.

Version 0.13.3

Release Date: September 8, 2020

Bug Fixes

• Deployment: Fix a bug where det-deploy local cluster-up was failing.

• WebUI: Fix a bug where experiment labels were not displayed on the experiment list page.

• WebUI: Fix a bug with decoding API responses because of unexpected non-numeric metric values.

Version 0.13.2

Release Date: September 3, 2020

New Features

• Support deploying Determined on Kubernetes.

  • Determined workloads run as a collection of pods, which allows standard Kubernetes tools for logging, metrics, and tracing to be used. Determined is compatible with Kubernetes >= 1.15, including managed Kubernetes services such as Google Kubernetes Engine (GKE) and AWS Elastic Kubernetes Service (EKS).

  • When using Determined with Kubernetes, we currently do not support fair-share scheduling, priority scheduling, per-experiment weights, or gang-scheduling for distributed training experiments; workloads will be scheduled according the behavior of the default Kubernetes scheduler.

  • Users can configure the behavior of the pods that are launched for Determined workloads by specifying a custom pod spec. A default pod spec can be configured when installing Kubernetes, but a custom pod spec can also be specified on a per-task basis (e.g., via the environment.pod_spec field in the experiment configuration file).

• Support running multiple distributed training jobs on a single agent.

  • In previous versions of Determined, a distributed training job could only be scheduled on an agent if it was configured to use all of the GPUs on that agent. In this release, that restriction has been lifted: for example, an agent with 8 GPUs can now be used to run two 4-GPU distributed training jobs. This feature is particularly useful as a way to improve utilization and fair resource allocation for smaller clusters.

Improvements

• WebUI: Update primary navigation. The primary navigation is all to one side, and is now collapsible to maximize content space.

• WebUI: Trial details improvements:

  • Update metrics selector to show the number of metrics selected to improve readability.

  • Add the “Has Checkpoint or Validation” filter.

  • Persist the “Has Checkpoint or Validation” filter setting across all trials, and persist the “Metrics” filter on trials of the same experiment.

• WebUI: Improve table pagination behavior. This will improve performance on Determined instances with many experiments.

• WebUI: Persist the sort order and sort column for the experiments, tasks, and trials tables to local storage.

• WebUI: Improve the default axes’ ranges for metrics charts. Also, update the range as new data points arrive.

• Add a warning when the PyTorch LR scheduler incorrectly uses an unwrapped optimizer. When using PyTorch with Determined, LR schedulers should be constructed using an optimizer that has been wrapped via the wrap_optimizer() method.

• Add a reminder to remove sys.exit() if SystemExit exception is caught.

Bug Fixes

• WebUI: Fix an issue where the recent task list did not apply the limit filter properly.

• Fix Keras and Estimator wrapping functions not returning the original objects when exporting checkpoints.

• Fix progress reporting for adaptive_asha searches that contain failed trials.

• Fix an issue that was causing OOM errors for some distributed EstimatorTrial experiments.

Version 0.13.1

Release Date: August 31, 2020

Bug Fixes

• Database migration: Fix a bug with a database migration in Determined version 0.13.0 which caused it to run slow and backfill incorrect values. Users on Determined versions 0.12.13 or earlier are recommended to upgrade to version 0.13.1. Users already on version 0.13.0 should upgrade to version 0.13.1 as usual.

• TensorBoard: Fix a bug that prevents TensorBoards from experiments with old experiment configuration versions from being loaded.

• WebUI: Fix an API response decoding issue on React where a null checkpoint resource was unhandled and could prevent trial detail page from rendering.

• WebUI: Fix an issue where terminated TensorBoard and notebook tasks were rendered as openable.

Version 0.13.0

Release Date: August 20, 2020

This release of Determined introduces several significant new features and modifications to existing features. When upgrading from a prior release of Determined, users should pay particular attention to the following changes:

• The concept of “steps” has been removed from the CLI, WebUI, APIs, and configuration files. Before upgrading, terminate all active and paused experiments (e.g., via det experiment cancel or det experiment kill). The format of the experiment config file has changed – configuration files that worked with previous versions of Determined will need to be updated to work with Determined >= 0.13.0.

• The WebUI has been partially rewritten, moving several components that were implemented in Elm to now being written in React and TypeScript. As part of this change, many improvements to the performance, appearance, and usability of the WebUI have been made. For more details, see the list of changes below. Please notify the Determined team of any regressions in functionality.

• The usability of the det shell feature has been significantly enhanced. As part of this change, the way in which arguments to det shell are parsed has changed; see details below.

We recommend taking a backup of the database before upgrading Determined.

New Features

• Allow trial containers to connect to the master using TLS.

• Allow agent’s TLS verification to skip verification or use a custom certificate for the master.

• For TFKerasTrial and EstimatorTrial, add support for disabling automatic sharding of the training dataset when doing distributed training. When wrapping a dataset via context.wrap_dataset, users can now pass shard_dataset=False. If this is done, users are responsible for splitting their dataset in such a manner that every GPU (rank) sees unique data.

Improvements

• Remove Steps from the UX: Remove the concept of a “step” from the CLI, WebUI, and configuration files. Add new configuration settings to allow settings previously in terms of steps to be configured instead in terms of records, batches or epochs..

  • Many configuration settings can now be set in terms of records, batches or epochs. For example, a single searcher can be configured to run for 100 records by setting max_length: {records: 100}, 100 batches by setting max_length: {batches: 100}, or 100 epochs by setting records_per_epoch at the root of the config and max_length: {epochs: 100}.

  • A new configuration setting, records_per_epoch, is added that must be specified when any quantity is configured in terms of epochs.

  • Breaking Change: For single, random and grid searchers searcher.max_steps has been replaced by searcher.max_length

  • Breaking Change: For ASHA based searchers, searcher.target_trial_steps and searcher.step_budget has been replaced by searcher.max_length and searcher.budget, respectively.

  • Breaking Change: For PBT, searcher.steps_per_round has been replaced by searcher.length_per_round.

  • Breaking Change: For all experiments, the names for min_validation_period and min_checkpoint_period are unchanged but they are now configured in terms of records, batches or epochs.

• Shell Mode Improvements: Determined supports launching GPU-attached terminal sessions via det shell. This release includes several changes to improve the usability of this feature, including:

  • The determined and determined-cli Python packages are now automatically installed inside containers launched by det shell. Any user-defined environment variables for the task image will be passed into the ssh sessions opened via det shell start or det shell open.

  • det shell should now work correctly in “host” networking mode.

  • det shell should now work correctly with dynamic agents and in cloud environments.

  • Breaking Change: Change how additional arguments to ssh are passed through det shell start and det shell open. Previously they were passed as a single string, like det shell open SHELL_ID --ssh-opt '-X -Y -o SomeSetting="some string"', but now the --ssh-opt has been removed and all extra positional arguments are passed through without requiring double-layers of quoting, like det shell open SHELL_ID -- -X -Y -o SomeSetting="some string" (note the use of -- to indicate all following arguments are positional arguments).

• WebUI changes

  • Tasks List: /det/tasks

    • Consolidate notebooks, tensorboards, shells, commands into single list page.

    • Add type filter to control which task types to display. By default all task types are shown when none of the types are selected.

    • Add type column with iconography to train users to familiarize task types with visual indicators.

    • Convert State filter from multi-select to single-select.

    • Convert actions from expanded buttons to overflow menu (triple vertical dots).

    • Move notebook launch buttons to task list from notebook list page.

    • Add pagination support that auto turns on when entries extend beyond 10 entries.

    • Add list of TensorBoard sources in a table Source column.

  • Experiment List: /det/experiments

    • State filter converted from multi-select to single-select.

    • Convert actions from expanded buttons to overflow menu (triple vertical dots).

    • Batch operation logic change to available if the action can be applied to any of the selected experiments

    • Add pagination support that auto turns on when entries extend beyond 10 entries.

  • Experiment Detail: /det/experiments/<id>

    • Implement charting with Plotly with zooming capability.

    • Trial table paginates on the WebUI side in preparation for API pagination in the near future.

    • Convert steps to batches in trials table and metric chart.

    • Update continue trial flow to use batches, epochs or records.

    • Use Monaco editor for the experiment config with YAML syntax highlighting.

    • Add links to source for Checkpoint modal view, allowing users to navigate to the corresponding experiment or trial for the checkpoint.

  • Trial Detail: /det/trials/<id>

    • Add trial information table.

    • Add trial metrics chart.

    • Implement charting with Plotly with zooming capability.

    • Trial info table paginates on the WebUI side in preparation for API pagination in the near future.

    • Add support for batches, records and epochs for experiment config.

    • Convert metric chart to show batches.

    • Convert steps table to batches table.

  • Master Logs: /det/logs, Trial Logs: /det/trials/<id>/logs, Task Logs: /det/<tasktype>/<id>/logs

    • Limit logs to 1000 lines for initial load and load an additional 1000 for each subsequent fetch of older logs.

    • Use new log viewer optimized for efficient rendering.

    • Introduce log line numbers.

    • Add ANSI color support.

    • Add error, warning, and debug visual icons and colors.

    • Add tailing button to enable tailing log behavior.

    • Add scroll to top button to load older logs out

    • Fix back and forth scrolling behavior on log viewer.

  • Cluster: /det/cluster

    • Separate out GPU from CPU resources.

    • Show resource availability and resource count (per type).

    • Render each resource as a donut chart.

  • Navigation

    • Update sidebar navigation for new task and experiment list pages.

    • Add link to new swagger API documentation.

    • Hide pagination controls for tables with less than 10 entries.

Bug Fixes

• Configuration: Do not load the entire experiment configuration when trying to check if an experiment is valid to be archived or unarchived.

• Configuration: Improve the master to validation hyperparameter configurations when experiments are submitted. Currently, the master checks whether global_batch_size has been specified and if it is numeric.

• Logs: Fix issue of not detecting newlines in the log messages, particularly Kubernetes log messages.

• Logs: Add intermediate step to trial log download to alert user that the CLI is the recommended action, especially for large logs.

• Searchers: Fix a bug in the SHA searcher caused by the promotion of already-exited trials.

• Security: Apply user authentication to streaming endpoints.

• Tasks: Allow the master certificate file to be readable even for a non-root task.

• TensorBoard: Fix issue affecting TensorBoards on AWS in us-east-1 region.

• TensorBoard: Recursively search for tfevents files in subdirectories, not just the top level log directory.

• WebUI: Fix scrolling issue that occurs when older logs are loaded, the tailing behavior is enabled, and the view is scrolled up.

• WebUI: Fix colors used for different states in the cluster resources chart.

• WebUI: Correct the numbers in the Batches column on the experiment list page.

• WebUI: Fix cluster and dashboard reporting for disabled slots.

• WebUI: Fix issue of archive/unarchive not showing up properly under the task actions.

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.

• 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 PyTorch 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(). Gradient clipping should now be done 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 to 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.

Improvements

• Allow proxy environment variables to be set in the agent config.

• 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.

• Add support for Nvidia T4 GPUs.

• det-deploy: Add support for g4 instance types on AWS.

• Upgrade Nvidia drivers on the default AWS and GCP images from 410.104 to 450.51.05.

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.

• 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 Checkpoints by including checkpoint metadata management.

• Reorganize examples by splitting the Trial 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 0.12.9

Release Date: June 16, 2020

• Retry ConnectionError and ProtocolError types for uploads to Google Cloud Storage.

• Fix a bug where the CLI was unable to make secure websocket connections to the master.

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

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 the Python SDK.

• 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 0.12.6

Release Date: June 5, 2020

• Add end of training callback to EstimatorTrial.

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.

Enterprise:

• Add support for using SCIM (System for Cross-domain Identity Management) to provision users.

• Add support for using OAuth2 to secure Determined’s SCIM integration.

• Add support for users to sign-on through an external IdP with SAML.

Version 0.12.4

Release Date: May 14, 2020

• Breaking Change: Users are no longer automatically logged in as the “determined” user.

• 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 (since deprecated).

• 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().

• 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 file system 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.