package tensorboard.service
Get desktop application:
View/edit binary Protocol Buffers messages
Service for exporting data from TensorBoard.dev.
rpc StreamExperiments (, stream )
export_service.proto:14Stream the experiment_id of all the experiments owned by the caller.
message
export_service.proto:26Request to stream the experiment_id of all the experiments owned by the caller from TensorBoard.dev.
optional read_timestamp = 1
Timestamp to get a consistent snapshot of the data in the database. This is useful when making multiple read RPCs and needing the data to be consistent across the read calls.
string user_id = 2
User ID defaults to the caller, but may be set to a different user for internal Takeout processes operating on behalf of a user.
int64 limit = 3
Limits the number of experiment IDs returned. This is useful to check if user might have any data by setting limit=1. Also useful to preview the list of experiments. TODO(@karthikv2k): Support pagination.
optional experiments_mask = 4
Field mask for what experiment data to return via the `experiments` field on the response. If not specified, this should be interpreted the same as an empty message: i.e., only the experiment ID should be returned. Other fields of `Experiment` will be populated if their corresponding bits in the `ExperimentMask` are set. The server may choose to populate fields that are not explicitly requested.
message
export_service.proto:48Streams experiment metadata (ID, creation time, etc.) from TensorBoard.dev.
repeated string experiment_ids = 1
Deprecated in favor of `experiments`. If a response has `experiments` set, clients should ignore `experiment_ids` entirely. Otherwise, clients should treat `experiment_ids` as a list of `experiments` for which only the `experiment_id` field is set, with the understanding that the other fields were not populated regardless of the requested field mask. For example, the following responses should be treated the same: # Response 1 experiment_ids: "123" experiment_ids: "456" # Response 2 experiments { experiment_id: "123" } experiments { experiment_id: "456" } # Response 3 experiment_ids: "789" experiments { experiment_id: "123" } experiments { experiment_id: "456" } See documentation on `experiments` for batching semantics.
repeated experiments = 2
List of experiments owned by the user. The entire list of experiments owned by the user is streamed in batches and each batch contains a list of experiments. A consumer of this stream needs to concatenate all these lists to get the full response. The order of experiments in the stream is not defined. Every response will contain at least one experiment. These messages may be partially populated, in accordance with the field mask given in the request.
rpc StreamExperimentData (, stream )
export_service.proto:17Stream scalars for all the runs and tags in an experiment.
message
export_service.proto:84Request to stream scalars from all the runs and tags in an experiment.
string experiment_id = 1
The permanent ID of the experiment whose data need to be streamed.
optional read_timestamp = 2
Timestamp to get a consistent snapshot of the data in the database. This is useful when making multiple read RPCs and needing the data to be consistent across the read calls. Should be the same as the read timestamp used for the corresponding `StreamExperimentsRequest` for consistency.
message
export_service.proto:103Streams data from all the runs and tags in an experiment. Each stream result only contains data for a single tag from a single run. For example if there are five runs and each run had two tags, the RPC will return a stream of at least ten `StreamExperimentDataResponse`s, each one having the either scalars or tensors for one tag. The values from a single tag may be split among multiple responses. Users need to aggregate information from entire stream to get data for the entire experiment. Empty experiments will have zero stream results. Empty runs that doesn't have any tags need not be supported by a hosted service.
string tag_name = 1
Name of the tag whose data is contained in this response.
string run_name = 2
Name of the run that contains the tag `tag_name`.
optional tag_metadata = 3
The metadata of the tag `tag_name`.
oneof data
Data to store for the tag `tag_name.
points = 4
Scalar data to store.
tensors = 5
Tensor data to store.
blob_sequences = 6
Blob sequences to store.
rpc StreamBlobData (, stream )
export_service.proto:20Stream blob as chunks for a given blob_id.
message
export_service.proto:159string blob_id = 1
message
export_service.proto:163bytes data = 1
The bytes in this chunk.
int64 offset = 2
The position in the blob where this chunk begins. This must equal the sum of the sizes of the chunks sent so far. Ignored if no data is provided.
bool final_chunk = 4
Indicates that this is the last chunk of the stream.
fixed32 final_crc32c = 5
CRC32C of the entire blob. This should be set when final_chunk=True, to protect against data corruption.
Service for writing data to TensorBoard.dev.
Request for a new location to write TensorBoard readable events.
message
write_service.proto:50This is currently empty on purpose. No information is necessary to request a URL, except. authorization of course, which doesn't come within the proto.
string name = 1
User provided name of the experiment.
string description = 2
User provided description of the experiment, in markdown source format.
message
write_service.proto:60Carries all information necessary to: 1. Inform the user where to navigate to see their TensorBoard. 2. Subsequently load (Scalars, Tensors, etc.) to the specified location.
string experiment_id = 1
Service-wide unique identifier of an uploaded log dir. eg: "1r9d0kQkh2laODSZcQXWP"
string url = 2
Url the user should navigate to to see their TensorBoard eg: "https://example.com/public/1r9d0kQkh2laODSZcQXWP"
Request to mutate metadata associated with an experiment.
message
write_service.proto:70Request to change the metadata of one experiment.
optional experiment = 1
Description of the data to set. The experiment_id field must match an experiment_id in the database. The remaining fields should be set to the desired metadata to be written. Only those fields marked True in the experiment_mask will be written. The service may deny modification of some metadata used for internal bookkeeping, such as num_scalars, etc.
optional experiment_mask = 2
Field mask for what experiment data to set. The service may deny requests to set some metatadata.
message
write_service.proto:84Response for setting experiment metadata.
This is empty on purpose.
(message has no fields)
Request that an experiment be deleted, along with all tags and scalars that it contains. This call may only be made by the original owner of the experiment.
message
write_service.proto:88string experiment_id = 1
Service-wide unique identifier of an uploaded log dir. eg: "1r9d0kQkh2laODSZcQXWP"
message
write_service.proto:94This is empty on purpose.
(message has no fields)
rpc PurgeData (, )
write_service.proto:27Request that unreachable data be purged. Used only for testing; disabled in production.
message
write_service.proto:99Only used for testing; corresponding RPC is disabled in prod.
int32 batch_limit = 1
Maximum number of entities of a given kind to purge at once (e.g., maximum number of tags to purge). Required; must be positive.
message
write_service.proto:106Only used for testing; corresponding RPC is disabled in prod.
optional purge_stats = 1
Stats about how many elements where purged. Compare to the batch limit specified in the request to estimate whether the backlog has any more items.
rpc WriteScalar (, )
write_service.proto:29Request additional scalar data be stored in TensorBoard.dev.
message
write_service.proto:125Carries all that is needed to add additional run data to the hosted service.
string experiment_id = 1
Which experiment to write to - corresponding to one hosted TensorBoard URL. The requester must have authorization to write to this location.
repeated runs = 2
Data to append to the existing storage at the experiment_id.
message
write_service.proto:159Everything the caller needs to know about how the writing went. (Currently empty)
This is empty on purpose.
(message has no fields)
rpc WriteTensor (, )
write_service.proto:31Request additional tensor data be stored in TensorBoard.dev.
message
write_service.proto:164Carries all that is needed to add tensor data to the hosted service.
string experiment_id = 1
Which experiment to write to - corresponding to one hosted TensorBoard URL. The requester must have authorization to write to this location.
repeated runs = 2
Data to append to the existing storage at the experiment_id.
message
write_service.proto:198Everything the caller needs to know about how the writing went. (Currently empty)
This is empty on purpose.
(message has no fields)
Request to obtain a specific BlobSequence entry, creating it if needed, to be subsequently populated with blobs.
message
write_service.proto:267Obtain a unique ID for a blob sequence, given the composite key (experiment_id, run, tag, step). If such a blob sequence already exists, return its ID. If not, create it first, and return the new ID.
string experiment_id = 1
Service-wide unique identifier of an uploaded log dir. eg: "1r9d0kQkh2laODSZcQXWP"
string run = 2
The name of the run to which the blob sequence belongs, for example "/some/path/mnist_experiments/run1/".
string tag = 3
The name of the tag to which the blob sequence belongs, for example "loss".
int64 step = 4
Step index within the run.
optional wall_time = 5
Timestamp of the creation of this blob sequence.
int64 final_sequence_length = 6
The total number of elements expected in the sequence. This effectively delares a number of initially empty 'upload slots', to be filled with subsequent WriteBlob RPCs.
optional metadata = 7
Note that metadata.plugin_data.content does not carry the payload.
message
write_service.proto:288string blob_sequence_id = 1
A unique ID for the the requested blob sequence.
Request the current status of blob data being stored in TensorBoard.dev, to support resumable uploads.
message
write_service.proto:293string blob_sequence_id = 1
The ID of the BlobSequence of which this blob is a member.
int64 index = 2
The position of this Blob within the BlobSequence
message
write_service.proto:300BlobState blob_state = 1
State of the object (still appending vs. complete).
int64 size = 2
Size of the object in bytes. In the case of a partial upload, this reflects only the data actually received so far.
fixed32 crc32c = 3
crc32c of the blob data stored so far, i.e. over the byte range [0, size).
rpc WriteBlob (stream , stream )
write_service.proto:41Request additional blob data be stored in TensorBoard.dev.
message
write_service.proto:314A single chunk of the blob write stream. Note that the WriteBlobRequest does not mirror the nested structure of WriteScalarRequest, because we only ever send one blob at a time.
string blob_sequence_id = 1
The ID of the BlobSequence of which this blob is a member.
int64 index = 2
The position of this Blob within the BlobSequence
bytes data = 3
The bytes in this chunk.
int64 offset = 4
The position in the blob where this chunk begins. This must equal the sum of the sizes of the chunks sent so far. Ignored if no data is provided.
fixed32 crc32c = 5
CRC32C of current data buffer. Clients must include the crc32c for every data buffer, to protect against data corruption. Note that for multi-shot writes, specifying the crc32c for every data buffer provides stronger protection than just providing the final_crc32c at the end of the upload.
bool finalize_object = 6
Indicates that this is the last chunk of the stream.
fixed32 final_crc32c = 7
CRC32C of the entire blob. Required, to protect against data corruption. This should be set only when finalize_object=True.
int64 blob_bytes = 8
Size in bytes of the entire blob. Required in the first request to allow quota allocation and management.
message
write_service.proto:340BlobState blob_state = 1
State of the object (still appending vs. complete).
int64 size = 2
Size of the object in bytes. This is the sum of the chunk sizes received from the stream so far. In the response to the final chunk, this size should equal the total size of the blob.
rpc DeleteOwnUser (, )
write_service.proto:44Request that the calling user and all their data be permanently deleted. Used for testing purposes.
message
write_service.proto:350Requests that the calling user and all their data be permanently deleted.
This is empty on purpose.
(message has no fields)
message
write_service.proto:355Everything the caller needs to know about how the deletion went.
This is empty on purpose.
(message has no fields)
message
server_info.proto:75Used in:
string endpoint = 1
gRPC server URI: <https://github.com/grpc/grpc/blob/master/doc/naming.md>. A scheme should always be specified, even if that is `dns`. For example: "dns:///api.tensorboard.dev:443".
message
blob.proto:16Used in:
string blob_id = 1
A non-empty ID for the blob.
BlobState state = 2
message
blob.proto:28Used in:
repeated entries = 1
message
blob.proto:22Used in:
optional blob = 1
Optional. If absent, this represents a "hole" in the sequence: there is expected to be a blob here, but upload has not started.
Used in: , ,
BLOB_STATE_UNKNOWN = 0
Object state is unknown. This value should never be used; it is present only as a proto3 best practice. See https://developers.google.com/protocol-buffers/docs/proto3#enum
BLOB_STATE_UNFINALIZED = 1
Object is being written and not yet finalized.
BLOB_STATE_CURRENT = 2
Object is finalized.
message
server_info.proto:68Used in:
CompatibilityVerdict verdict = 1
string details = 2
Human-readable message to display. When non-empty, will be displayed in all cases, even when the client may proceed.
Used in:
VERDICT_UNKNOWN = 0
VERDICT_OK = 1
All is well. The client may proceed.
VERDICT_WARN = 2
The client may proceed, but should heed the accompanying message. This may be the case if the user is on a version of TensorBoard that will soon be unsupported, or if the server is experiencing transient issues.
VERDICT_ERROR = 3
The client should cease further communication with the server and abort operation after printing the accompanying `details` message.
message
experiment.proto:8Resource message representing an Experiment.
Used in: ,
string experiment_id = 1
Permanent ID of this experiment; e.g.: "AdYd1TgeTlaLWXx6I8JUbA". Output-only.
optional create_time = 2
The time that the experiment was created. Output-only.
optional update_time = 3
The time that the experiment was last modified: i.e., the most recent time that scalars were added to the experiment. Output-only.
int64 num_scalars = 4
The number of scalars in this experiment, across all time series. Output-only.
int64 num_runs = 5
The number of distinct run names in this experiment. Output-only.
int64 num_tags = 6
The number of distinct tag names in this experiment. A tag name that appears in multiple runs will be counted only once. Output-only.
string name = 7
User provided name of the experiment.
string description = 8
User provided description of the experiment, in markdown source format.
int64 total_tensor_bytes = 9
The number of bytes used for storage of tensors in this experiment, across all time series, including estimated overhead. Output-only.
int64 total_blob_bytes = 10
The number of bytes used for storage of the contents of blobs in this experiment, across all time series, including estimated overhead. Output-only.
string owner = 11
The owner of this experiment, as an opaque user_id string. This field is ignored on upload, with owner information coming from the authentication information passed alongside the proto.
message
experiment.proto:44Field mask for `Experiment` used in get and update RPCs. The `experiment_id` field is always implicitly considered to be set.
Used in: ,
bool create_time = 2
bool update_time = 3
bool num_scalars = 4
bool num_runs = 5
bool num_tags = 6
bool name = 7
bool description = 8
bool total_tensor_bytes = 9
bool total_blob_bytes = 10
bool owner = 11
message
server_info.proto:82Used in:
string template = 1
Template string for experiment URLs. All occurrences of the value of the `id_placeholder` field in this template string should be replaced with an experiment ID. For example, if `id_placeholder` is "{{EID}}", then `template` might be "https://tensorboard.dev/experiment/{{EID}}/". Should be absolute.
string id_placeholder = 2
Placeholder string that should be replaced with an actual experiment ID. (See docs for `template` field.)
message
server_info.proto:101Used in:
repeated string allowed_plugins = 1
Plugins for which data should be uploaded. These are plugin names as stored in the the `SummaryMetadata.plugin_data.plugin_name` proto field.
message
server_info.proto:94Used in:
repeated string upload_plugins = 2
Plugins for which the client wishes to upload data. These are plugin names as stored in the the `SummaryMetadata.plugin_data.plugin_name` proto field.
message
write_service.proto:115Details about what actions were taken as a result of a purge request. These values are upper bounds; they may exceed the true values.
Used in:
int32 tags = 1
Number of tags deleted as a result of this request.
int32 experiments = 2
Number of experiments marked as purged as a result of this request.
int32 users = 3
Number of users deleted as a result of this request.
message
scalar.proto:9One point viewable on a scalar metric plot.
Used in:
int64 step = 1
Step index within the run.
optional wall_time = 2
Timestamp of the creation of this point.
double value = 3
Value of the point at this step / timestamp.
message
scalar.proto:19Metadata for the ScalarPoints stored for one (Experiment, Run, Tag).
int64 max_step = 1
Maximum step recorded for the tag.
optional max_wall_time = 2
Timestamp corresponding to the max step.
optional summary_metadata = 3
Information about the plugin which created this scalar data. Note: The period is required part of the type here due to the package name resolution logic.
message
server_info.proto:8Request sent by uploader clients at the start of an upload session. Used to determine whether the client is recent enough to communicate with the server, and to receive any metadata needed for the upload session.
string version = 1
Client-side TensorBoard version, per `tensorboard.version.VERSION`.
optional plugin_specification = 2
Information about the plugins for which the client wishes to upload data. If specified then the list of plugins will be confirmed by the server and echoed in the PluginControl.allowed_plugins field. Otherwise the server will return the default set of plugins it supports. If one of the plugins is not supported by the server then it will respond with compatibility verdict VERDICT_ERROR.
message
server_info.proto:22optional compatibility = 1
Primary bottom-line: is the server compatible with the client, can it serve its request, and is there anything that the end user should be aware of?
optional api_server = 2
Identifier for a gRPC server providing the `TensorBoardExporterService` and `TensorBoardWriterService` services (under the `tensorboard.service` proto package).
optional url_format = 3
How to generate URLs to experiment pages.
optional plugin_control = 4
Information about the plugins for which data should be uploaded. If PluginSpecification.requested_plugins is specified then that list of plugins will be confirmed by the server and echoed in the the response. Otherwise the server will return the default set of plugins it supports. The client should only upload data for the plugins in the response even if it is capable of uploading more data.
optional upload_limits = 5
Limits on the upload process that the client should honor. This may include limits on data size, chunk size, records per request, request rate, bandwidth, etc. The server may enforce such limits, in which case the values reported here should be kept in sync. Providing these limits in advance allows the client to avoid triggering server errors (both through more efficent operation and by earlier detection of real error conditions), and to print better error messages. If this field is not set, the client should choose reasonable default values to guide its behavior.
message
export_service.proto:149Used in:
repeated int64 steps = 1
Step index within the run.
repeated wall_times = 2
Timestamp of the creation of this point.
repeated values = 3
Value of the blob sequence at this step / timestamp.
message
export_service.proto:127Data for the scalars are stored in a columnar fashion to optimize it for exporting the data into textual formats like JSON. The data for the ith scalar is { steps[i], wall_times[i], values[i] }. The data here is sorted by step values in ascending order.
Used in:
repeated int64 steps = 1
Step index within the run.
repeated wall_times = 2
Timestamp of the creation of this point.
repeated double values = 3
Value of the point at this step / timestamp.
message
export_service.proto:140Data for the Tensors are stored in a columnar fashion to optimize it for exporting the data into textual formats like JSON. The data for the ith tensor is { steps[i], wall_times[i], values[i] }. The data here is sorted by step values in ascending order.
Used in:
repeated int64 steps = 1
Step index within the run.
repeated wall_times = 2
Timestamp of the creation of this point.
repeated values = 3
Value of the point at this step / timestamp.
message
tensor.proto:10One point viewable on a tensor metric plot.
Used in:
int64 step = 1
Step index within the run.
optional wall_time = 2
Timestamp of the creation of this point.
optional value = 3
Value of the point at this step / timestamp.
message
tensor.proto:20Metadata for the TensorPoints stored for one (Experiment, Run, Tag).
int64 max_step = 1
Maximum step recorded for the tag.
optional max_wall_time = 2
Timestamp corresponding to the max step.
optional summary_metadata = 3
Information about the plugin which created this tensor data. Note: The period is required part of the type here due to the package name resolution logic.
message
server_info.proto:107Used in:
int64 max_scalar_request_size = 3
The maximum allowed WriteScalar request size, in bytes. If this is 0 or unset, client should use a reasonable default value. If this is negative, no scalars should be uploaded.
int64 max_tensor_request_size = 4
The maximum allowed WriteTensor request size, in bytes. If this is 0 or unset, client should use a reasonable default value. If this is negative, no tensors should be uploaded.
int64 max_blob_request_size = 5
The maximum allowed WriteBlob request size, in bytes. If this is 0 or unset, client should use a reasonable default value. If this is negative, no blobs should be uploaded.
int64 min_scalar_request_interval = 6
The minimum interval between WriteScalar requests, in milliseconds. If this is 0 or unset, client should use a reasonable default value.
int64 min_tensor_request_interval = 7
The minimum interval between WriteTensor requests, in milliseconds. If this is 0 or unset, client should use a reasonable default value.
int64 min_blob_request_interval = 8
The minimum interval between WriteBlob requests, in milliseconds. If this is 0 or unset, client should use a reasonable default value.
int64 max_blob_size = 1
The maximum allowed size for blob uploads. If this is 0 or unset, client should use a reasonable default value. If this is negative, no blobs should be uploaded.
int64 max_tensor_point_size = 2
The maximum allowed size for tensor point uploads. If this is 0 or unset, client should use a reasonable default value. If this is negative, no blobs should be uploaded.
message
write_service.proto:131All the data to store for one Run. This data will be stored under the corresponding run in the hosted storage. WriteScalarRequest is merged into the data store for the keyed run. The tags and included scalars will be the union of the data sent across all WriteScalarRequests. Metadata by default uses a 'first write wins' approach.
Used in:
string name = 1
The name of this run. For example "/some/path/mnist_experiments/run1/"
repeated tags = 2
Data to store for this Run/Tag combination.
message
write_service.proto:141All the data to store for one Tag of one Run. This data will be stored under the corresponding run/tag in the hosted storage. A tag corresponds to a single time series.
Used in:
string name = 1
The name of this tag. For example "loss"
repeated points = 2
Data to store for this Run/Tag combination.
optional metadata = 3
The metadata of this tag.
message
write_service.proto:170All the data to store for one Run. This data will be stored under the corresponding run in the hosted storage. WriteTensorRequest is merged into the data store for the keyed run. The tags and included tensors will be the union of the data sent across all WriteTensorRequests. Metadata by default uses a 'first write wins' approach.
Used in:
string name = 1
The name of this run. For example "/some/path/mnist_experiments/run1/"
repeated tags = 2
Data to store for this Run/Tag combination.
message
write_service.proto:180All the data to store for one Tag of one Run. This data will be stored under the corresponding run/tag in the hosted storage. A tag corresponds to a single time series.
Used in:
string name = 1
The name of this tag. For example "loss"
repeated points = 2
Data to store for this Run/Tag combination.
optional metadata = 3
The metadata of this tag.