package tensorboardX.hparam

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: ListSessionGroupsRequest

• AGGREGATION_UNSET = 0

• AGGREGATION_AVG = 1

• AGGREGATION_MEDIAN = 2

• AGGREGATION_MIN = 3

• AGGREGATION_MAX = 4

Defines parmeters for a ListSessionGroupsRequest for a specific column. See the comment for "ListSessionGroupsRequest" above for more details. NEXT_TAG: 9

Used in: ListSessionGroupsRequest

• oneof name

  • MetricName metric = 1

  • string hparam = 2

• SortOrder order = 3

  Sorting. The final order of session groups in the response is defined by the sub collection of ColParams messages (out of the ListSessionGroupsRequest.col_params repeated field) whose 'order' field (below) is not ORDER_UNSPECIFIED. In each of the messages in this sub-collection, the next two fields specify the ordering of the values and missing_values in the associated column of the session group. The order of the ColParams messages themselves within the sub-collection determines the "significance" of the associated column as a sorting key: with the first being the primary sorting key, the second being the secondary sorting key, etc. Note: The 'session group name' is added as a least significant sorting key to the keys defined by the user, so the order in the response is always deterministic.

• bool missing_values_first = 4

  This field is ignored if order is ORDER_UNSPECIFIED. Otherwise, if true, missing values are ordered before every other value in the column; if false they are ordered after every other value in the column.

• oneof filter

  Filtering. The 'filter' oneof specifies a subset of the domain of the values a column may take. Only session groups with each of their column values belonging to this subset are included in the response. If this field is not specified, the subset is taken to be the entire column domain.

  • string filter_regexp = 5

    Only valid for string-valued hyperparameter columns. The subset is the set of all strings matching the regular expression stored in 'regexp' as a partial match (use '^<regexp>$' to have a full match against regexp).

  • Interval filter_interval = 6

    Only valid for numeric-valued columns. The subset is the given interval.

  • google.protobuf.ListValue filter_discrete = 7

    Valid for all data types. The subset is defined explicitly.

• bool exclude_missing_values = 8

  Specifies whether to exclude session groups whose column value is missing from the response.

Used in: HParamInfo

• DATA_TYPE_UNSET = 0

• DATA_TYPE_STRING = 1

• DATA_TYPE_BOOL = 2

• DATA_TYPE_FLOAT64 = 3

Used in: MetricInfo

• DATASET_UNKNOWN = 0

• DATASET_TRAINING = 1

• DATASET_VALIDATION = 2

Represents a single experiment. An experiment consists of multiple "sessions". Typically, in each session a model is trained for a given set of hyperparameter values. In each session a training program may generate one or more series of real numbers--each containing the evaluation of some metric on the model at different training steps. Note that Sessions can consist of multiple Tensorboard "runs", since in a distributed Tensorflow deployment, training can be accomplished using several cooporating processes, each one emitting Summary data to a different log directory or run. For example, in a single session one process could periodically compute the loss on the validation set, and another could compute the loss on the training set. NEXT_TAG: 7

Used in: HParamsPluginData

• string name = 6

  -- Experiments are scoped by a global name. Currently, Tensorboard supports displaying data for a single experiment.

• string description = 1

  A description. May contain markdown.

• string user = 2

  An id for the owning user or group.

• double time_created_secs = 3

  The time the experiment was created. In seconds since the UNIX epoch.

• repeated HParamInfo hparam_infos = 4

  Information about each hyperparameter used in the experiment.

• repeated MetricInfo metric_infos = 5

  Information about each metric used in the experiment.

Parameters for a GetExperiment API call. Each experiment is scoped by a unique global id. NEXT_TAG: 2

• string experiment_name = 1

  REQUIRED

NEXT_TAG: 7

Used in: Experiment

• string name = 1

  An id for the hyperparameter.

• string display_name = 2

  A string used to display the hyperparameter in the UI. If empty, the UI will display the 'name' field.

• string description = 3

  A description. May contain markdown.

• DataType type = 4

  The data type of this hyperparameter.

• oneof domain

  Specifies the set of values this hyperparameter can hold. The UI assumes every instance of this hyperparameter will hold a value from this set. It is used by the UI to allow filtering so that only session groups (see below) whose associated hyperparameter value "passes" the filter are displayed. If this is not populated, the domain is assumed to be the entire domain of the type of the hyperparameter.

  • google.protobuf.ListValue domain_discrete = 5

    A discrete set of the values this hyperparameter can hold.

  • Interval domain_interval = 6

    Numeric data type only. The (real) interval from which values of this hyperparameter are taken.

HParam summaries created by `tensorboard.plugins.hparams.summary` module will include `SummaryMetadata` whose `plugin_data` field has as `content` a serialized HParamsPluginData message.

• int32 version = 1

  The version of the plugin data schema.

• oneof data

  • Experiment experiment = 2

  • SessionStartInfo session_start_info = 3

  • SessionEndInfo session_end_info = 4

Represents the closed interval [min_value, max_value] of the real line. NEXT_TAG: 3

Used in: ColParams, HParamInfo

• double min_value = 1

• double max_value = 2

See ListMetricEvalsRequest in http_api.md. NEXT_TAG: 4

• string experiment_name = 3

• string session_name = 1

• optional MetricName metric_name = 2

Parameters for a ListSessionGroups API call. Computes a list of the current session groups allowing for filtering and sorting by metrics and hyperparameter values. Returns a "slice" of that list specified by start_index and slice_size. NEXT_TAG: 8

• string experiment_name = 6

• repeated Status allowed_statuses = 7

  Filters the set of sessions (from which the session groups are formed) to contain only these sessions whose status is contained in 'allowed_statuses'.

• repeated ColParams col_params = 1

  A list of ColParams messages--one for each "column" of a session group. A session group column contains either a metric evaluated at the current reported computation step or a hyperparameter value. In addition to 'regular' values, a column may take on a special 'missing-value' which denotes that the hyperparameter or metric is not available for the session group (for example, if the metric is not used in the group). The ColParams messages in the repeated field below configure filtering and sorting of the resulting collection of session groups. See the comments of the fields in the ColParam message below for more details.

• AggregationType aggregation_type = 2

  Fields controlling how to aggregate metrics across sessions within a session group. If aggregation_type is AGGREGATION_AVG, each metric value of the session group is the average of the values of the metric across the sessions. Otherwise, the session group metric values are taken directly from a "representative" session in the group, selected as a session for which primary_metric takes on its minimum, maximum, or median value, as specified by the choice of aggregation_type (for median, if the number of sessions in the group is even, a session with a lower "middle" value is chosen as the representative session).

• optional MetricName aggregation_metric = 3

  See comment for 'aggregation_type' above.

• int32 start_index = 4

  The next two parameters determine the "slice" of the full list of session groups--sorted and filtered by the parameters above--to return. The 0-based index of the first session group to return.

• int32 slice_size = 5

  The number of session groups to return starting at the session group indexed by 'start_index'. The actual number of session groups returned is min{slice_size, total_size - start_index}, where total_size is the number of session groups in the full list sorted and filtered by the parameters above (if start_index > total_size no session groups are returned).

See ListSessionGroups in http_api.md. NEXT_TAG: 4

• repeated SessionGroup session_groups = 1

• int32 total_size = 3

  Denotes the total number of session groups in the full filtered list. (Recall that this response may only be a slice). It is used by the UI to calculate total number of pages and can be set here to -1 to mean "unknown".

NEXT_TAG: 6

Used in: Experiment

• optional MetricName name = 1

• string display_name = 3

  A string used to display the metric in the UI. If empty, the UI will display the 'name' field.

• string description = 4

  A description. May contain markdown.

• DatasetType dataset_type = 5

  The dataset type (validation, training) on which the metric is computed.

NEXT_TAG: 3

Used in: ColParams, ListMetricEvalsRequest, ListSessionGroupsRequest, MetricInfo, MetricValue

• string group = 1

  An identifier for a metric. A metric is a real-valued function of the model. The UI can plot metrics for a session evaluated on the model at different training steps. We identify a metric by a (group, tag) pair of strings. The UI treats both of these as opaque strings. The only requirement is that the pair uniquely identifies a metric in the experiment. We use a pair so the UI could allow the user to group metrics for a single session by either group or tag to be displayed in the same chart. For instance, one can set the metric group to correspond to the dataset on which the model was evaluated, and the UI can then display different metrics describing the same underlying computation and using different datasets, on the same chart. When exporting summaries from Tensorflow, in a typical setup, a training session exports evaluations of metrics at different training steps as Scalar-plugin summaries--each having a run of the form "<session_base_log_dir>/<sub_dir>", and some associated tag. The same metric for different sessions would use the same sub_dir and tag, but would have a different session_base_log_dir. For example, a session computing two metrics: model loss on the validation set and model loss on the training set, can export these as scalar summaries with the tag "loss" and runs session_base_log_dir/validation and session_base_log_dir/training, respectively. In this setup, the 'group' field can be populated with the "sub_dir" associated with the metric, and the 'tag' field can be populated with the tag: "loss".

• string tag = 2

NEXT_TAG: 5

Used in: Session, SessionGroup

• optional MetricName name = 1

• double value = 2

• int32 training_step = 3

  The training step at which this value is computed.

• double wall_time_secs = 4

  The wall time in seconds since UNIX epoch at which this value is computed.

NEXT_TAG: 8

Used in: SessionGroup

• string name = 1

  An id for the session. Unique within an experiment (not just the group).

• double start_time_secs = 2

  In seconds since the UNIX epoch.

• double end_time_secs = 3

  In seconds since the UNIX epoch. May be 0 if unavailable or the session has not finished yet.

• Status status = 4

  May be STATUS_UNKNOWN if unavailable.

• string model_uri = 5

  A URI for a resource that will allow the user to reconstruct the model for this session. E.g., in Tensorflow this could point to a directory where the checkpoints are stored. Currently, this is treated opaquely by the UI and only displayed to the user as it is passed here.

• repeated MetricValue metric_values = 6

  Stores each metric evaluation on the model at the current training step. Unfortunately, we can't store these as a map, since proto maps can't have message keys.

• string monitor_url = 7

  An optional link to a web page monitoring the session.

Used in: HParamsPluginData

• Status status = 1

• double end_time_secs = 2

  The time the session ended in seconds since epoch.

In some experiments, the user trains a model with the same set of hyperparameters multiple times to get the distribution of metric evaluations, when the computation (such as the training algorithm, or metric evaluation) is non-deterministic. To make the UI aware of this, sessions are partitioned into groups: each group consists of all training sessions which share the same values for the hyperparameters. In experiments with no repeated executions, each group consists of exactly one session. NEXT_TAG: 6

Used in: ListSessionGroupsResponse

• string name = 1

• map<string, google.protobuf.Value> hparams = 2

  Stores the hyperparameters for sessions within this group as a mapping from the hyperparameter name to its value.

• repeated MetricValue metric_values = 3

  A list of pairs (metric, value)--one for each metric in the experiment. The value denotes the evaluation of the corresponding metric on the model aggregated across the sessions in this group. The exact method of aggregation is specified in the comments of ListSessionGroupsRequest. Unfortunately, we can't store these as a map, since proto maps can't have message keys.

• repeated Session sessions = 4

  The sessions belonging to this group.

• string monitor_url = 5

  An optional link to a web page monitoring the session group.

Used in: HParamsPluginData

• map<string, google.protobuf.Value> hparams = 1

  A map describing the hyperparameter values for the session. Maps each hyperparameter name to its value. Currently only scalars are supported.

• string model_uri = 2

  A URI for where checkpoints are saved.

• string monitor_url = 3

  An optional URL to a website monitoring the session.

• string group_name = 4

  The name of the session group containing this session. If empty, the group name is taken to be the session id (so this session is the only member of its group).

• double start_time_secs = 5

  The time the session started in seconds since epoch.

Used in: ColParams

• ORDER_UNSPECIFIED = 0

• ORDER_ASC = 1

• ORDER_DESC = 2

Represents the status of a Session.

Used in: ListSessionGroupsRequest, Session, SessionEndInfo

• STATUS_UNKNOWN = 0

• STATUS_SUCCESS = 1

• STATUS_FAILURE = 2

• STATUS_RUNNING = 3