package determined.trial.v1

Get desktop application:
View/edit binary Protocol Buffers messages

CheckpointWorkload is an artifact created by a trial during training.

Used in: api.v1.WorkloadContainer, run.v1.Run, Trial

• string uuid = 1

  UUID of the checkpoint.

• optional google.protobuf.Timestamp end_time = 3

  The time the workload finished or was stopped.

• checkpoint.v1.State state = 4

  The state of the checkpoint.

• map<string, int64> resources = 5

  Dictionary of file paths to file sizes in bytes of all files in the checkpoint.

• int32 total_batches = 8

  Total number of batches as of this workload's completion.

• optional google.protobuf.Struct metadata = 9

  User defined metadata associated with the checkpoint.

Metrics report.

Used in: api.v1.GetMetricsResponse, api.v1.GetTrainingMetricsResponse, api.v1.GetTrialMetricsByCheckpointResponse, api.v1.GetTrialMetricsByModelVersionResponse, api.v1.GetValidationMetricsResponse

• int32 trial_id = 1

  ID of the trial.

• optional google.protobuf.Timestamp end_time = 2

  End time of when metric was reported.

• optional google.protobuf.Struct metrics = 3

  Struct of the reported metrics.

• int32 total_batches = 4

  batches completed in the report.

• bool archived = 5

  If metric is archived.

• int32 id = 6

  ID of metric in table.

• int32 trial_run_id = 7

  Run ID of trial when metric was reported.

• string group = 8

  Name of the Metric Group ("training", "validation", anything else)

MetricsWorkload is a workload generating metrics.

Used in: api.v1.WorkloadContainer, run.v1.Run, Trial

• optional google.protobuf.Timestamp end_time = 2

  The time the workload finished or was stopped.

• optional common.v1.Metrics metrics = 40

  Metrics.

• int32 num_inputs = 5

  Number of inputs processed.

• int32 total_batches = 8

  Total number of batches as of this workload's completion.

The rendezvous info for the trial to rendezvous with sibling containers.

Used in: api.v1.AllocationRendezvousInfoResponse

• repeated string addresses = 1

  The rendezvous addresses of the other containers.

• int32 rank = 2

  The container rank.

• repeated int32 slots = 3

  The slots for each address, respectively.

The current state of the trial. see \dT+ trial_state in db

Used in: api.v1.PatchTrialRequest, run.v1.FlatRun, run.v1.Run, Trial

• STATE_UNSPECIFIED = 0

  The trial is in an unspecified state.

• STATE_ACTIVE = 1

  The trial is in an active state.

• STATE_PAUSED = 2

  The trial is in a paused state

• STATE_STOPPING_CANCELED = 3

  The trial is canceled and is shutting down.

• STATE_STOPPING_KILLED = 4

  The trial is killed and is shutting down.

• STATE_STOPPING_COMPLETED = 5

  The trial is completed and is shutting down.

• STATE_STOPPING_ERROR = 6

  The trial is errored and is shutting down.

• STATE_CANCELED = 7

  The trial is canceled and is shut down.

• STATE_COMPLETED = 8

  The trial is completed and is shut down.

• STATE_ERROR = 9

  The trial is errored and is shut down.

• STATE_QUEUED = 10

  The trial is queued (waiting to be run, or job state is still queued). Queued is a substate of the Active state.

• STATE_PULLING = 11

  The trial is pulling the image. Pulling is a substate of the Active state.

• STATE_STARTING = 12

  The trial is preparing the environment after finishing pulling the image. Starting is a substate of the Active state.

• STATE_RUNNING = 13

  The trial's allocation is actively running. Running is a substate of the Active state.

Trial is a set of workloads and are exploring a determined set of hyperparameters.

Used in: api.v1.ComparableTrial, api.v1.CreateTrialResponse, api.v1.GetExperimentTrialsResponse, api.v1.GetTrialByExternalIDResponse, api.v1.GetTrialResponse, api.v1.PatchTrialResponse, api.v1.PutTrialResponse, api.v1.SearchExperimentExperiment

• int32 id = 1

  The id of the trial.

• int32 experiment_id = 2

  The id of the parent experiment.

• optional google.protobuf.Timestamp start_time = 3

  The time the trial was started.

• optional google.protobuf.Timestamp end_time = 4

  The time the trial ended if the trial is stopped.

• State state = 5

  The current state of the trial.

• int32 restarts = 17

  Number times the trial restarted.

• optional google.protobuf.Struct hparams = 6

  Trial hyperparameters.

• int32 total_batches_processed = 7

  The current processed batches.

• optional MetricsWorkload best_validation = 8

  Best validation.

• optional MetricsWorkload latest_validation = 9

  Latest validation.

• optional CheckpointWorkload best_checkpoint = 10

  Best checkpoint.

• string runner_state = 11

  The last reported state of the trial runner (harness code).

• double wall_clock_time = 12

  The wall clock time is all active time of the cluster for the trial, inclusive of everything (restarts, initiailization, etc), in seconds.

• string warm_start_checkpoint_uuid = 13

  UUID of checkpoint that this trial started from.

• string task_id = 14

  Id of the first task associated with this trial. This field is deprecated since trials can have multiple tasks.

• uint64 total_checkpoint_size = 15

  The sum of sizes of all resources in all checkpoints for the trial.

• int32 checkpoint_count = 18

  The count of checkpoints.

• optional google.protobuf.Struct summary_metrics = 19

  summary metrics

• repeated string task_ids = 20

  Task IDs of tasks associated with this trial. Length of task_ids will always be greater or equal to one when TaskID is sent. For example CompareTrial we will send a reduced Trial object, without TaskID or TaskIDs fileld in. The first element of task_ids will be the same as task_id. task_ids is sorted ascending by task_run_id.

• double searcher_metric_value = 21

  Signed searcher metrics value.

• optional int32 log_retention_days = 22

  Number of days to retain logs for.

• optional google.protobuf.Struct metadata = 23

  metadata associated with the trial (based off the metadata stored in the run).

• optional string log_policy_matched = 24

  Log Policy Matched.

Signals to the experiment the trial early exited.

Used in: api.v1.ReportTrialSearcherEarlyExitRequest

• TrialEarlyExit.ExitedReason reason = 1

  The reason for the exit.

The reason for an early exit.

Used in: TrialEarlyExit

• EXITED_REASON_UNSPECIFIED = 0

  Zero-value (not allowed).

• EXITED_REASON_INVALID_HP = 1

  Indicates the trial exited due to an invalid hyperparameter.

• EXITED_REASON_INIT_INVALID_HP = 3

  Indicates the trial exited due to an invalid hyperparameter in the trial init.

Metrics from the trial some duration of training.

Used in: api.v1.ReportTrialMetricsRequest, api.v1.ReportTrialTrainingMetricsRequest, api.v1.ReportTrialValidationMetricsRequest

• int32 trial_id = 1

  The trial associated with these metrics.

• int32 trial_run_id = 2

  The trial run associated with these metrics.

• optional int32 steps_completed = 3

  The number of batches trained on when these metrics were reported.

• optional google.protobuf.Timestamp report_time = 4

  The client-reported time associated with these metrics.

• optional common.v1.Metrics metrics = 9

  The metrics for this bit of training, including: - avg_metrics: metrics reduced over the reporting period). - batch_metrics: (optional) per-batch metrics.

TrialProfilerMetricLabels are the labels for a single series, where a series is a defined as all metrics sharing a distinct set of labels

Used in: api.v1.GetTrialProfilerAvailableSeriesResponse, api.v1.GetTrialProfilerMetricsRequest, TrialProfilerMetricsBatch

• int32 trial_id = 1

  The ID of the trial.

• string name = 2

  The name of the metric.

• string agent_id = 3

  The agent ID associated with the metric.

• string gpu_uuid = 4

  The GPU UUID associated with the metric.

• TrialProfilerMetricLabels.ProfilerMetricType metric_type = 5

  The type of the metric.

To distinguish the 2 different categories of metrics.

Used in: TrialProfilerMetricLabels

• PROFILER_METRIC_TYPE_UNSPECIFIED = 0

  Zero-value (not allowed).

• PROFILER_METRIC_TYPE_SYSTEM = 1

  For systems metrics, like GPU utilization or memory.

• PROFILER_METRIC_TYPE_TIMING = 2

  For timing metrics, like how long a backwards pass or getting a batch from the dataloader took.

• PROFILER_METRIC_TYPE_MISC = 3

  For other miscellaneous metrics.

TrialProfilerMetricsBatch is a batch of trial profiler metrics. A batch will contain metrics pertaining to a single series. The fields values, batches and timestamps will be equal length arrays with each index corresponding to a reading.

Used in: api.v1.GetTrialProfilerMetricsResponse, api.v1.PostTrialProfilerMetricsBatchRequest

• repeated float values = 1

  The measurement for a reading, repeated for the batch of metrics.

• repeated int32 batches = 2

  The batch at which a reading occurred, repeated for the batch of metrics.

• repeated google.protobuf.Timestamp timestamps = 3

  The timestamp at which a reading occurred, repeated for the batch of metrics.

• optional TrialProfilerMetricLabels labels = 4

  The labels for this series.

The metadata pertaining to the current running task for a trial.

Used in: api.v1.PostTrialRunnerMetadataRequest

• string state = 1

  The state of the trial runner.

Denotes a connection between a given trial and a checkpoint or model_version

Used in: api.v1.ReportTrialSourceInfoRequest

• int32 trial_id = 1

  ID of the trial.

• string checkpoint_uuid = 2

  UUID of the checkpoint.

• optional int32 model_id = 3

  Source `id`` for the model which generated the checkpoint (if applicable)

• optional int32 model_version = 4

  Source `version` in the model_version version field which generated the checkpoint (if applicable)

• TrialSourceInfoType trial_source_info_type = 5

  Type for this trial_source_info

TrialSourceInfoType is the type of the TrialSourceInfo, which serves as a link between a trial and a checkpoint or model version

Used in: api.v1.GetTrialMetricsByCheckpointRequest, api.v1.GetTrialMetricsByModelVersionRequest, TrialSourceInfo

• TRIAL_SOURCE_INFO_TYPE_UNSPECIFIED = 0

  The type is unspecified

• TRIAL_SOURCE_INFO_TYPE_INFERENCE = 1

  "Inference" Trial Source Info Type, used for batch inference

• TRIAL_SOURCE_INFO_TYPE_FINE_TUNING = 2

  "Fine Tuning" Trial Source Info Type, used in model hub