package grpc.trainer.v2

Get desktop application:
View/edit binary Protocol Buffers messages

• rpc CreateTrainingJob (CreateRequest, CreateResponse)

  trainer.proto:25

  Creates a new deep learning training job for a given model definition.

  message CreateRequest

  trainer.proto:83

  • string user_id = 1

  • optional ModelDefinition model_definition = 2

  • optional Training training = 3

  • repeated Datastore datastores = 4

  • optional EMExtractionSpec evaluation_metrics = 5

    EMExtractionSpec allows the caller to specify evaluation metrics extraction.

  message CreateResponse

  trainer.proto:248

  • string training_id = 1

• rpc GetTrainingJob (GetRequest, GetResponse)

  trainer.proto:29

  Reads a training job with a given ID.

  message GetResponse

  trainer.proto:270

  • optional Job job = 1

• rpc GetAllTrainingsJobs (GetAllRequest, GetAllResponse)

  trainer.proto:33

  Returns all training jobs for a given user.

  message GetAllRequest

  trainer.proto:290

  • string user_id = 1

  message GetAllResponse

  trainer.proto:294

  • repeated Job jobs = 1

• rpc DeleteTrainingJob (DeleteRequest, DeleteResponse)

  trainer.proto:37

  Deletes a training job with a given ID.

  message DeleteRequest

  trainer.proto:320

  • string training_id = 1

  • string user_id = 2

  message DeleteResponse

  trainer.proto:325

  • string training_id = 1

• rpc HaltTrainingJob (HaltRequest, HaltResponse)

  trainer.proto:41

  Halts the training with a given ID without discarding the result.

  message HaltRequest

  trainer.proto:298

  • string training_id = 1

  • string user_id = 2

  message HaltResponse

  trainer.proto:303

  • string training_id = 1

  • string user_id = 2

  • Status status = 3

• rpc GetModelDefinition (ModelDefinitionRequest, stream ZippedDataChunk)

  trainer.proto:45

  Returns the model definition that was used for training as application/zip.

  message ModelDefinitionRequest

  trainer.proto:459

  • string training_id = 1

  • string user_id = 2

• rpc GetTrainedModel (TrainedModelRequest, stream ZippedDataChunk)

  trainer.proto:49

  Returns the trained model as application/zip.

  message TrainedModelRequest

  trainer.proto:464

  • string training_id = 1

  • string user_id = 2

• rpc GetTrainedModelLogs (TrainedModelLogRequest, stream ByteStreamResponse)

  trainer.proto:54

  Returns the logs from the ZIP file stored in the object store. Deprecated

  message TrainedModelLogRequest

  trainer.proto:469

  • string training_id = 1

  • string user_id = 2

  • bool follow = 5

  • bool is_metrics = 3

    These are for internal use only, and will be eventually removed!

  • bool is_summary = 4

  message ByteStreamResponse

  trainer.proto:506

  • bytes data = 1

• rpc GetTrainingLogs (Query, stream LogLine)

  trainer.proto:57

  message LogLine

  trainer.proto:185

  LogLine represents one line of log information, returned by training data endpoints.

  • optional MetaInfo meta = 1

    For MetaInfo, at the minimum training_id and user_id must be specified.

  • string line = 2

    Raw line from the logs.

• rpc GetTrainingEMetrics (Query, stream EMetrics)

  trainer.proto:60

  Get evaluation metrics records, based on query

  message EMetrics

  trainer.proto:209

  EMetrics specifies an evaluation metrics record from the training data.

  • optional MetaInfo meta = 1

    For MetaInfo, at the minimum training_id and user_id must be specified.

  • map<string, Any> etimes = 2

    Map of temporal keys, normally values for the x-axis on a graph. Example: {"iteration": 209}

  • string grouplabel = 3

    Group label, such as test, train, or validate

  • map<string, Any> values = 4

    Map of value keys, normally values for the y-axis on a graph. / Example: {"cross_entropy": 0.4430539906024933, "accuracy": 0.8999999761581421}

• rpc GetVersions (GetVersionsRequest, Frameworks)

  trainer.proto:62

  message GetVersionsRequest

  trainer.proto:514

  (message has no fields)

  message Frameworks

  trainer.proto:519

  Contains a list of all frameworks currently supported along with the versions of that framework and whether a specific framework version can be used by anyone or only for internal usage.

  • map<string, FrameworkDetailList> frameworks = 1

• rpc GetTrainingStatusID (GetRequest, GetStatusIDResponse)

  trainer.proto:68

  For internal use only!

  message GetStatusIDResponse

  trainer.proto:278

  • Status status = 1

• rpc UpdateTrainingJob (UpdateRequest, UpdateResponse)

  trainer.proto:74

  For internal use only! Updates an existing training status TODO we should not have this but until we fix the status update handling properly, we have no other choice.

  message UpdateRequest

  trainer.proto:252

  • string training_id = 1

  • string user_id = 2

  • Status status = 3

  • string status_message = 4

  • string error_code = 5

  • string timestamp = 6

  message UpdateResponse

  trainer.proto:261

  • string training_id = 1

• rpc ResumeTrainingJob (ResumeRequest, ResumeResponse)

  trainer.proto:78

  Not implemented, to be removed (for GA)

  message ResumeRequest

  trainer.proto:309

  • string training_id = 1

  • string user_id = 2

  message ResumeResponse

  trainer.proto:314

  • string training_id = 1

  • string user_id = 2

  • Status status = 3

Any represents a typed value used with the evaluation metrics record.

Used in: EMetrics

• Any.DataType type = 1

  Datatype of the value.

• string value = 2

  String representation of a value

Used in: Any

• STRING = 0

• JSONSTRING = 1

• INT = 2

• FLOAT = 3

Used in: CreateRequest, Job

• string id = 1

• string type = 2

• map<string, string> fields = 3

  additional fields for the given Datastore type

• map<string, string> connection = 4

  connection information for the given Datastore type

Typed value for EMExtractionSpec. The data type here can't be an enum, due to internal issues with YAML conversion.

Used in: EMGroup

• string type = 1

  one of: STRING, JSONSTRING, INT, FLOAT

• string value = 2

  String represenation of the value.

EMExtractionSpec represents the specification for extracting structured evaluation metrics from training jobs. It is used across all log collectors, so some fields may not be relevent for all log collectors. Note: Don't use enums with this, as need to do untyped YAML convert to string and back Refer to https://github.ibm.com/deep-learning-platform/dlaas-training-metrics-service for complete documentation.

Used in: CreateRequest

• string type = 1

  Loosly typed string representing what kind of log-collector to use. For Logs-only, specify `type: logger` For the Regex_extractor log-collector, specify `type: regex_extractor` For Tensorboard, specify `type: tensorboard` To invoke the emetrics_file_extractor, you can specify the following synonyms `type: optivist` || `type: emetrics_file` || `type: file`.

• string image_tag = 6

  Dev only.

• string in = 2

  The filename of the logfile. (Normally this should be left to default).

• int32 line_lookahead = 3

  For the regex_extractor, number of lines to keep in the buffer for regex matching.

• repeated string eventTypes = 4

  (Eventual) Available event types: 'images', 'distributions', 'histograms', 'images' 'audio', 'scalars', 'tensors', 'graph', 'meta_graph', 'run_metadata'. For now only scalars are supported.

• map<string, EMGroup> groups = 5

  For the regex_extractor, the `EMExtractionSpec` should contain a `groups` section, which should contain templates for groups such as `test` and `train`, which group names should be be the keys of this map.

EMGroup represents a group, such as `test` or `train`, that acts as a template for structured evaluation metrics, and which allows the specification of a regular expression (regex) that contains named bindings with sub-expressions, which can then be used as references to specify structured time-related (x-axis) and value-related (y-axis) values.

Used in: EMExtractionSpec

• string regex = 1

  Python regular expressions, which use the named group feature `(?P<name>...)`, to specify a name of a matching expression, which can then be used to specify the value that is used in the template for the `EMetrics` record. To help with verbosity, the regex_extractor allows the following macros: GLOG_STAMP, TIMESTAMP, FLOAT, INT, INT_ANY, and HEX. (See dlaas-training-metrics-service README for more details.

• optional EMMeta meta = 2

  Allows the caller to specify a binding for the time field of the meta structure.

• map<string, EMAny> values = 3

  Map of keys and regex references for value-related (y-axis) values.

• map<string, EMAny> etimes = 4

  Map of keys and regex references for time-related (x-axis) values.

Allows the user to bind an extracted value to the time field of the evaluation metrics.

Used in: EMGroup

• string time = 3

  Time that the metric occured: representing the number of millisecond since midnight January 1, 1970. (ref, for instance $timestamp). Value will be extracted from timestamps

• string subid = 5

  Optional subid

Used in: ModelDefinition

• string name = 1

• string version = 2

• string image_tag = 3

  Optional: tag used for learner testing

• optional ImageLocation image_location = 4

  Optional: non-standard location for learner image

Used in: Frameworks

• repeated FrameworkDetails versions = 1

Used in: FrameworkDetailList

• string version = 1

• bool external = 2

  If true, the image can be used by any user. If false, the image is only available for internal usage.

• string training_id = 1

• string user_id = 2

• string training_id = 1

• string user_id = 2

• optional Metrics metrics = 3

• string metrics = 1

Used as request type in: Trainer.GetTrainingJob, Trainer.GetTrainingStatusID

• string training_id = 1

• string user_id = 2

• optional TrainingStatus status = 1

• string test = 1

Used in: Framework

• string registry = 1

  the server name for the docker registry

• string namespace = 2

  namespace within the registry

• string access_token = 3

  Token used to access images stored in the registry+namespace

• string email = 4

  Email address associated with the account

Used in: GetAllResponse, GetResponse

• string training_id = 1

• string user_id = 2

• optional ModelDefinition model_definition = 3

• optional Training training = 4

• optional TrainingStatus status = 5

• repeated Datastore datastores = 6

• string job_id = 7

• optional Metrics metrics = 8

MetaInfo represents data shared with both log lines and evaluation metrics.

Used in: EMetrics, LogLine, Query

• string training_id = 1

  Unique id identifying the training job

• string user_id = 2

  Unique id identifying the user

• int64 time = 3

  Time that the metric occured: representing the number of millisecond since midnight January 1, 1970.

• int64 rindex = 4

  Sequential index, 1-based

• string subid = 5

  Optional subid

Used in: GetLatestMetricsResponse, Job, UpdateTrainedModelMetricsRequest

• string timestamp = 1

• string type = 2

• int32 iteration = 3

• map<string, string> values = 4

Used in: CreateRequest, Job

• string name = 1

• string description = 2

• bytes content = 3

  Optional: application/zip as bytes containing the model definition. If not present field location needs to be set.

• string location = 4

  Optional: data store location where the model definition (code) is located

• optional Framework framework = 5

Query specifies the input query for logs and evaluation metrics.

Used as request type in: Trainer.GetTrainingEMetrics, Trainer.GetTrainingLogs

• Query.SearchType searchType = 1

  At this time, the SearchType value should normally always be TERM.

• optional MetaInfo meta = 3

  At the minimum, the training_data and user_data must be specified in the meta substructure.

• string since = 4

  representing the number of milliseconds since midnight January 1, 1970, exclusive with pos.

• int32 pagesize = 5

  Only get this many records

• int64 pos = 6

  The starting position. If positive or zero, count from beginning, if negative, count from end, exclusive with since.

Used in: Query

• TERM = 0

• NESTED = 1

• MATCH = 2

• ALL = 3

Used to specify resource requirements of a training job

Used in: Training

• float cpus = 1

  Number of CPU cores

• float gpus = 2

  Number of GPUs

• float memory = 3

  RAM

• SizeUnit memory_unit = 4

• float storage = 5

• SizeUnit storage_unit = 6

• int32 learners = 7

  Number of learners

• string schedpolicy = 8

• string topology = 9

• string architecture = 10

• string gpu_type = 11

  Optional. If not specified, job will be scheduled ONLY on nvidia-TeslaK80 Constraint strictly enforced. If e.g., a nvidia-TeslaP100 is requested, job will NOT start until a nvidia-TeslaP100 is available Can only be nvidia-TeslaK80, nvidia-TeslaP100 or nvidia-TeslaV100

Used in: ResourceRequirements

• MB = 0

• MiB = 1

• GB = 2

• GiB = 3

Used in: GetStatusIDResponse, HaltResponse, ResumeResponse, TrainingStatus, UpdateRequest

• NOT_STARTED = 0

• PENDING = 1

• HALTED = 5

• FAILED = 10

• DEPLOY = 20

• DOWNLOADING = 30

• PROCESSING = 40

• STORING = 50

• COMPLETED = 60

• QUEUED = 70

• string training_id = 1

• string user_id = 2

• bool follow = 3

Used in: CreateRequest, Job

• string command = 1

  Command to execute during training

• optional ResourceRequirements resources = 2

  Resource requirements for the training

• repeated string input_data = 3

  Input and output data as data store references

• repeated string output_data = 4

• bool profiling = 5

  whether we want to enable detailed profiling during the training

Used in: GetStatusResponse, Job

• Status status = 1

• string submission_timestamp = 3

• string completion_timestamp = 4

• string download_start_timestamp = 5

• string process_start_timestamp = 6

• string store_start_timestamp = 7

• string status_message = 8

• string error_code = 9

• string training_id = 1

• string user_id = 2

• optional Metrics new_metrics = 3

• string training_id = 1

Used as response type in: Trainer.GetModelDefinition, Trainer.GetTrainedModel

• bytes data = 1