HuggingfaceClient.Hub.TensorBoard (huggingface_client v0.1.0)

HuggingFace TensorBoard Logger — push TensorBoard logs to the Hub.

Allows you to track experiment metrics and push TensorBoard logs to a HuggingFace repository, making them viewable in the Hub's embedded TensorBoard.

See: https://huggingface.co/docs/huggingface_hub/package_reference/tensorboard

Example

# Create a logger
logger = HuggingfaceClient.TensorBoard.new("my-org/my-model",
  logdir: "./runs/exp1",
  access_token: "hf_..."
)

# Log scalars (call during training)
HuggingfaceClient.TensorBoard.log_scalar(logger, "train/loss", 0.245, step: 100)
HuggingfaceClient.TensorBoard.log_scalar(logger, "eval/accuracy", 0.923, step: 100)

# Push logs to Hub
{:ok, url} = HuggingfaceClient.TensorBoard.push(logger)
IO.puts("TensorBoard available at: #{url}")

Summary

Types

t()

Functions

log_scalar(logger, tag, value, opts \\ [])

Logs a scalar metric value. Writes to local logdir.

new(repo_id, opts \\ [])

Creates a new TensorBoard logger connected to a Hub repository.

push(logger, opts \\ [])

Pushes all TensorBoard log files from logdir to the Hub repository.

read_scalars(logger)

Reads logged scalars from the local logdir.

upload_logs(logger, commit_message \\ "Update TensorBoard logs")

Uploads TensorBoard log files from the logdir to the Hub.

Types

t()

@type t() :: %HuggingfaceClient.Hub.TensorBoard{
  logdir: String.t(),
  path_in_repo: String.t(),
  private: boolean(),
  repo_id: String.t(),
  repo_type: atom(),
  token: String.t() | nil
}

Functions

log_scalar(logger, tag, value, opts \\ [])

@spec log_scalar(t(), String.t(), float(), keyword()) :: :ok

Logs a scalar metric value. Writes to local logdir.

In production, combine with a real TensorBoard library (e.g. :tensorflow). This creates a simple JSON log as a fallback.

Example

HuggingfaceClient.TensorBoard.log_scalar(logger, "train/loss", 0.245, step: 100)
HuggingfaceClient.TensorBoard.log_scalar(logger, "eval/f1", 0.89, step: 100)

new(repo_id, opts \\ [])

@spec new(
  String.t(),
  keyword()
) :: t()

Creates a new TensorBoard logger connected to a Hub repository.

Options

:logdir — local directory for TensorBoard logs (default: "./runs")
:access_token — HF API token
:repo_type — :model | :dataset | :space (default: :model)
:path_in_repo — directory in the repo (default: "runs")
:private — create repo as private if it doesn't exist (default: false)

Example

logger = HuggingfaceClient.TensorBoard.new("my-org/my-training-run",
  logdir: "./tb_logs",
  access_token: token
)

push(logger, opts \\ [])

@spec push(
  t(),
  keyword()
) :: {:ok, String.t()} | {:error, Exception.t()}

Pushes all TensorBoard log files from logdir to the Hub repository.

Creates the repository if it doesn't exist.

Example

{:ok, url} = HuggingfaceClient.TensorBoard.push(logger,
  commit_message: "Epoch 10 training logs"
)
IO.puts("View at: #{url}")

read_scalars(logger)

@spec read_scalars(t()) :: [map()]

Reads logged scalars from the local logdir.

Example

scalars = HuggingfaceClient.TensorBoard.read_scalars(logger)
loss_values = Enum.filter(scalars, fn s -> s["tag"] == "train/loss" end)

upload_logs(logger, commit_message \\ "Update TensorBoard logs")

@spec upload_logs(t(), String.t()) :: {:ok, String.t()} | {:error, Exception.t()}

Uploads TensorBoard log files from the logdir to the Hub.

Example

{:ok, url} = HuggingfaceClient.TensorBoard.upload_logs(logger)