Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/trackio_logo_old.png filter=lfs diff=lfs merge=lfs -text

__init__.py

@@ -0,0 +1,302 @@
+import hashlib
+import os
+import warnings
+import webbrowser
+from pathlib import Path
+from typing import Any
+
+from gradio.blocks import BUILT_IN_THEMES
+from gradio.themes import Default as DefaultTheme
+from gradio.themes import ThemeClass
+from gradio_client import Client
+from huggingface_hub import SpaceStorage
+
+from trackio import context_vars, deploy, utils
+from trackio.imports import import_csv, import_tf_events
+from trackio.media import TrackioImage, TrackioVideo
+from trackio.run import Run
+from trackio.sqlite_storage import SQLiteStorage
+from trackio.table import Table
+from trackio.ui.main import demo
+from trackio.utils import TRACKIO_DIR, TRACKIO_LOGO_DIR
+
+__version__ = Path(__file__).parent.joinpath("version.txt").read_text().strip()
+
+__all__ = [
+    "init",
+    "log",
+    "finish",
+    "show",
+    "import_csv",
+    "import_tf_events",
+    "Image",
+    "Video",
+    "Table",
+]
+
+Image = TrackioImage
+Video = TrackioVideo
+
+
+config = {}
+
+DEFAULT_THEME = "citrus"
+
+
+def init(
+    project: str,
+    name: str | None = None,
+    group: str | None = None,
+    space_id: str | None = None,
+    space_storage: SpaceStorage | None = None,
+    dataset_id: str | None = None,
+    config: dict | None = None,
+    resume: str = "never",
+    settings: Any = None,
+    private: bool | None = None,
+    embed: bool = True,
+) -> Run:
+    """
+    Creates a new Trackio project and returns a [`Run`] object.
+
+    Args:
+        project (`str`):
+            The name of the project (can be an existing project to continue tracking or
+            a new project to start tracking from scratch).
+        name (`str`, *optional*):
+            The name of the run (if not provided, a default name will be generated).
+        group (`str`, *optional*):
+            The name of the group which this run belongs to in order to help organize
+            related runs together. You can toggle the entire group's visibilitiy in the
+            dashboard.
+        space_id (`str`, *optional*):
+            If provided, the project will be logged to a Hugging Face Space instead of
+            a local directory. Should be a complete Space name like
+            `"username/reponame"` or `"orgname/reponame"`, or just `"reponame"` in which
+            case the Space will be created in the currently-logged-in Hugging Face
+            user's namespace. If the Space does not exist, it will be created. If the
+            Space already exists, the project will be logged to it.
+        space_storage ([`~huggingface_hub.SpaceStorage`], *optional*):
+            Choice of persistent storage tier.
+        dataset_id (`str`, *optional*):
+            If a `space_id` is provided, a persistent Hugging Face Dataset will be
+            created and the metrics will be synced to it every 5 minutes. Specify a
+            Dataset with name like `"username/datasetname"` or `"orgname/datasetname"`,
+            or `"datasetname"` (uses currently-logged-in Hugging Face user's namespace),
+            or `None` (uses the same name as the Space but with the `"_dataset"`
+            suffix). If the Dataset does not exist, it will be created. If the Dataset
+            already exists, the project will be appended to it.
+        config (`dict`, *optional*):
+            A dictionary of configuration options. Provided for compatibility with
+            `wandb.init()`.
+        resume (`str`, *optional*, defaults to `"never"`):
+            Controls how to handle resuming a run. Can be one of:
+
+            - `"must"`: Must resume the run with the given name, raises error if run
+              doesn't exist
+            - `"allow"`: Resume the run if it exists, otherwise create a new run
+            - `"never"`: Never resume a run, always create a new one
+        private (`bool`, *optional*):
+            Whether to make the Space private. If None (default), the repo will be
+            public unless the organization's default is private. This value is ignored
+            if the repo already exists.
+        settings (`Any`, *optional*):
+            Not used. Provided for compatibility with `wandb.init()`.
+        embed (`bool`, *optional*, defaults to `True`):
+            If running inside a jupyter/Colab notebook, whether the dashboard should
+            automatically be embedded in the cell when trackio.init() is called.
+
+    Returns:
+        `Run`: A [`Run`] object that can be used to log metrics and finish the run.
+    """
+    if settings is not None:
+        warnings.warn(
+            "* Warning: settings is not used. Provided for compatibility with wandb.init(). Please create an issue at: https://github.com/gradio-app/trackio/issues if you need a specific feature implemented."
+        )
+
+    if space_id is None and dataset_id is not None:
+        raise ValueError("Must provide a `space_id` when `dataset_id` is provided.")
+    space_id, dataset_id = utils.preprocess_space_and_dataset_ids(space_id, dataset_id)
+    url = context_vars.current_server.get()
+    share_url = context_vars.current_share_server.get()
+
+    if url is None:
+        if space_id is None:
+            _, url, share_url = demo.launch(
+                show_api=False,
+                inline=False,
+                quiet=True,
+                prevent_thread_lock=True,
+                show_error=True,
+                favicon_path=TRACKIO_LOGO_DIR / "trackio_logo_light.png",
+                allowed_paths=[TRACKIO_LOGO_DIR],
+            )
+        else:
+            url = space_id
+            share_url = None
+        context_vars.current_server.set(url)
+        context_vars.current_share_server.set(share_url)
+    if (
+        context_vars.current_project.get() is None
+        or context_vars.current_project.get() != project
+    ):
+        print(f"* Trackio project initialized: {project}")
+
+        if dataset_id is not None:
+            os.environ["TRACKIO_DATASET_ID"] = dataset_id
+            print(
+                f"* Trackio metrics will be synced to Hugging Face Dataset: {dataset_id}"
+            )
+        if space_id is None:
+            print(f"* Trackio metrics logged to: {TRACKIO_DIR}")
+            if utils.is_in_notebook() and embed:
+                base_url = share_url + "/" if share_url else url
+                full_url = utils.get_full_url(
+                    base_url, project=project, write_token=demo.write_token
+                )
+                utils.embed_url_in_notebook(full_url)
+            else:
+                utils.print_dashboard_instructions(project)
+        else:
+            deploy.create_space_if_not_exists(
+                space_id, space_storage, dataset_id, private
+            )
+            user_name, space_name = space_id.split("/")
+            space_url = deploy.SPACE_HOST_URL.format(
+                user_name=user_name, space_name=space_name
+            )
+            print(f"* View dashboard by going to: {space_url}")
+            if utils.is_in_notebook() and embed:
+                utils.embed_url_in_notebook(space_url)
+    context_vars.current_project.set(project)
+
+    client = None
+    if not space_id:
+        client = Client(url, verbose=False)
+
+    if resume == "must":
+        if name is None:
+            raise ValueError("Must provide a run name when resume='must'")
+        if name not in SQLiteStorage.get_runs(project):
+            raise ValueError(f"Run '{name}' does not exist in project '{project}'")
+        resumed = True
+    elif resume == "allow":
+        resumed = name is not None and name in SQLiteStorage.get_runs(project)
+    elif resume == "never":
+        if name is not None and name in SQLiteStorage.get_runs(project):
+            warnings.warn(
+                f"* Warning: resume='never' but a run '{name}' already exists in "
+                f"project '{project}'. Generating a new name and instead. If you want "
+                "to resume this run, call init() with resume='must' or resume='allow'."
+            )
+            name = None
+        resumed = False
+    else:
+        raise ValueError("resume must be one of: 'must', 'allow', or 'never'")
+
+    run = Run(
+        url=url,
+        project=project,
+        client=client,
+        name=name,
+        group=group,
+        config=config,
+        space_id=space_id,
+    )
+
+    if resumed:
+        print(f"* Resumed existing run: {run.name}")
+    else:
+        print(f"* Created new run: {run.name}")
+
+    context_vars.current_run.set(run)
+    globals()["config"] = run.config
+    return run
+
+
+def log(metrics: dict, step: int | None = None) -> None:
+    """
+    Logs metrics to the current run.
+
+    Args:
+        metrics (`dict`):
+            A dictionary of metrics to log.
+        step (`int`, *optional*):
+            The step number. If not provided, the step will be incremented
+            automatically.
+    """
+    run = context_vars.current_run.get()
+    if run is None:
+        raise RuntimeError("Call trackio.init() before trackio.log().")
+    run.log(
+        metrics=metrics,
+        step=step,
+    )
+
+
+def finish():
+    """
+    Finishes the current run.
+    """
+    run = context_vars.current_run.get()
+    if run is None:
+        raise RuntimeError("Call trackio.init() before trackio.finish().")
+    run.finish()
+
+
+def show(project: str | None = None, theme: str | ThemeClass = DEFAULT_THEME):
+    """
+    Launches the Trackio dashboard.
+
+    Args:
+        project (`str`, *optional*):
+            The name of the project whose runs to show. If not provided, all projects
+            will be shown and the user can select one.
+        theme (`str` or `ThemeClass`, *optional*, defaults to `"citrus"`):
+            A Gradio Theme to use for the dashboard instead of the default `"citrus"`,
+            can be a built-in theme (e.g. `'soft'`, `'default'`), a theme from the Hub
+            (e.g. `"gstaff/xkcd"`), or a custom Theme class.
+    """
+    if theme != DEFAULT_THEME:
+        # TODO: It's a little hacky to reproduce this theme-setting logic from Gradio Blocks,
+        # but in Gradio 6.0, the theme will be set in `launch()` instead, which means that we
+        # will be able to remove this code.
+        if isinstance(theme, str):
+            if theme.lower() in BUILT_IN_THEMES:
+                theme = BUILT_IN_THEMES[theme.lower()]
+            else:
+                try:
+                    theme = ThemeClass.from_hub(theme)
+                except Exception as e:
+                    warnings.warn(f"Cannot load {theme}. Caught Exception: {str(e)}")
+                    theme = DefaultTheme()
+        if not isinstance(theme, ThemeClass):
+            warnings.warn("Theme should be a class loaded from gradio.themes")
+            theme = DefaultTheme()
+        demo.theme: ThemeClass = theme
+        demo.theme_css = theme._get_theme_css()
+        demo.stylesheets = theme._stylesheets
+        theme_hasher = hashlib.sha256()
+        theme_hasher.update(demo.theme_css.encode("utf-8"))
+        demo.theme_hash = theme_hasher.hexdigest()
+
+    _, url, share_url = demo.launch(
+        show_api=False,
+        quiet=True,
+        inline=False,
+        prevent_thread_lock=True,
+        favicon_path=TRACKIO_LOGO_DIR / "trackio_logo_light.png",
+        allowed_paths=[TRACKIO_LOGO_DIR],
+    )
+
+    base_url = share_url + "/" if share_url else url
+    full_url = utils.get_full_url(
+        base_url, project=project, write_token=demo.write_token
+    )
+
+    if not utils.is_in_notebook():
+        print(f"* Trackio UI launched at: {full_url}")
+        webbrowser.open(full_url)
+        utils.block_main_thread_until_keyboard_interrupt()
+    else:
+        utils.embed_url_in_notebook(full_url)

cli.py

@@ -0,0 +1,32 @@
+import argparse
+
+from trackio import show
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Trackio CLI")
+    subparsers = parser.add_subparsers(dest="command")
+
+    ui_parser = subparsers.add_parser(
+        "show", help="Show the Trackio dashboard UI for a project"
+    )
+    ui_parser.add_argument(
+        "--project", required=False, help="Project name to show in the dashboard"
+    )
+    ui_parser.add_argument(
+        "--theme",
+        required=False,
+        default="citrus",
+        help="A Gradio Theme to use for the dashboard instead of the default 'citrus', can be a built-in theme (e.g. 'soft', 'default'), a theme from the Hub (e.g. 'gstaff/xkcd').",
+    )
+
+    args = parser.parse_args()
+
+    if args.command == "show":
+        show(args.project, args.theme)
+    else:
+        parser.print_help()
+
+
+if __name__ == "__main__":
+    main()

commit_scheduler.py

@@ -0,0 +1,391 @@
+# Originally copied from https://github.com/huggingface/huggingface_hub/blob/d0a948fc2a32ed6e557042a95ef3e4af97ec4a7c/src/huggingface_hub/_commit_scheduler.py
+
+import atexit
+import logging
+import os
+import time
+from concurrent.futures import Future
+from dataclasses import dataclass
+from io import SEEK_END, SEEK_SET, BytesIO
+from pathlib import Path
+from threading import Lock, Thread
+from typing import Callable, Dict, List, Union
+
+from huggingface_hub.hf_api import (
+    DEFAULT_IGNORE_PATTERNS,
+    CommitInfo,
+    CommitOperationAdd,
+    HfApi,
+)
+from huggingface_hub.utils import filter_repo_objects
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class _FileToUpload:
+    """Temporary dataclass to store info about files to upload. Not meant to be used directly."""
+
+    local_path: Path
+    path_in_repo: str
+    size_limit: int
+    last_modified: float
+
+
+class CommitScheduler:
+    """
+    Scheduler to upload a local folder to the Hub at regular intervals (e.g. push to hub every 5 minutes).
+
+    The recommended way to use the scheduler is to use it as a context manager. This ensures that the scheduler is
+    properly stopped and the last commit is triggered when the script ends. The scheduler can also be stopped manually
+    with the `stop` method. Checkout the [upload guide](https://huggingface.co/docs/huggingface_hub/guides/upload#scheduled-uploads)
+    to learn more about how to use it.
+
+    Args:
+        repo_id (`str`):
+            The id of the repo to commit to.
+        folder_path (`str` or `Path`):
+            Path to the local folder to upload regularly.
+        every (`int` or `float`, *optional*):
+            The number of minutes between each commit. Defaults to 5 minutes.
+        path_in_repo (`str`, *optional*):
+            Relative path of the directory in the repo, for example: `"checkpoints/"`. Defaults to the root folder
+            of the repository.
+        repo_type (`str`, *optional*):
+            The type of the repo to commit to. Defaults to `model`.
+        revision (`str`, *optional*):
+            The revision of the repo to commit to. Defaults to `main`.
+        private (`bool`, *optional*):
+            Whether to make the repo private. If `None` (default), the repo will be public unless the organization's default is private. This value is ignored if the repo already exists.
+        token (`str`, *optional*):
+            The token to use to commit to the repo. Defaults to the token saved on the machine.
+        allow_patterns (`List[str]` or `str`, *optional*):
+            If provided, only files matching at least one pattern are uploaded.
+        ignore_patterns (`List[str]` or `str`, *optional*):
+            If provided, files matching any of the patterns are not uploaded.
+        squash_history (`bool`, *optional*):
+            Whether to squash the history of the repo after each commit. Defaults to `False`. Squashing commits is
+            useful to avoid degraded performances on the repo when it grows too large.
+        hf_api (`HfApi`, *optional*):
+            The [`HfApi`] client to use to commit to the Hub. Can be set with custom settings (user agent, token,...).
+        on_before_commit (`Callable[[], None]`, *optional*):
+            If specified, a function that will be called before the CommitScheduler lists files to create a commit.
+
+    Example:
+    ```py
+    >>> from pathlib import Path
+    >>> from huggingface_hub import CommitScheduler
+
+    # Scheduler uploads every 10 minutes
+    >>> csv_path = Path("watched_folder/data.csv")
+    >>> CommitScheduler(repo_id="test_scheduler", repo_type="dataset", folder_path=csv_path.parent, every=10)
+
+    >>> with csv_path.open("a") as f:
+    ...     f.write("first line")
+
+    # Some time later (...)
+    >>> with csv_path.open("a") as f:
+    ...     f.write("second line")
+    ```
+
+    Example using a context manager:
+    ```py
+    >>> from pathlib import Path
+    >>> from huggingface_hub import CommitScheduler
+
+    >>> with CommitScheduler(repo_id="test_scheduler", repo_type="dataset", folder_path="watched_folder", every=10) as scheduler:
+    ...     csv_path = Path("watched_folder/data.csv")
+    ...     with csv_path.open("a") as f:
+    ...         f.write("first line")
+    ...     (...)
+    ...     with csv_path.open("a") as f:
+    ...         f.write("second line")
+
+    # Scheduler is now stopped and last commit have been triggered
+    ```
+    """
+
+    def __init__(
+        self,
+        *,
+        repo_id: str,
+        folder_path: Union[str, Path],
+        every: Union[int, float] = 5,
+        path_in_repo: str | None = None,
+        repo_type: str | None = None,
+        revision: str | None = None,
+        private: bool | None = None,
+        token: str | None = None,
+        allow_patterns: list[str] | str | None = None,
+        ignore_patterns: list[str] | str | None = None,
+        squash_history: bool = False,
+        hf_api: HfApi | None = None,
+        on_before_commit: Callable[[], None] | None = None,
+    ) -> None:
+        self.api = hf_api or HfApi(token=token)
+        self.on_before_commit = on_before_commit
+
+        # Folder
+        self.folder_path = Path(folder_path).expanduser().resolve()
+        self.path_in_repo = path_in_repo or ""
+        self.allow_patterns = allow_patterns
+
+        if ignore_patterns is None:
+            ignore_patterns = []
+        elif isinstance(ignore_patterns, str):
+            ignore_patterns = [ignore_patterns]
+        self.ignore_patterns = ignore_patterns + DEFAULT_IGNORE_PATTERNS
+
+        if self.folder_path.is_file():
+            raise ValueError(
+                f"'folder_path' must be a directory, not a file: '{self.folder_path}'."
+            )
+        self.folder_path.mkdir(parents=True, exist_ok=True)
+
+        # Repository
+        repo_url = self.api.create_repo(
+            repo_id=repo_id, private=private, repo_type=repo_type, exist_ok=True
+        )
+        self.repo_id = repo_url.repo_id
+        self.repo_type = repo_type
+        self.revision = revision
+        self.token = token
+
+        self.last_uploaded: Dict[Path, float] = {}
+        self.last_push_time: float | None = None
+
+        if not every > 0:
+            raise ValueError(f"'every' must be a positive integer, not '{every}'.")
+        self.lock = Lock()
+        self.every = every
+        self.squash_history = squash_history
+
+        logger.info(
+            f"Scheduled job to push '{self.folder_path}' to '{self.repo_id}' every {self.every} minutes."
+        )
+        self._scheduler_thread = Thread(target=self._run_scheduler, daemon=True)
+        self._scheduler_thread.start()
+        atexit.register(self._push_to_hub)
+
+        self.__stopped = False
+
+    def stop(self) -> None:
+        """Stop the scheduler.
+
+        A stopped scheduler cannot be restarted. Mostly for tests purposes.
+        """
+        self.__stopped = True
+
+    def __enter__(self) -> "CommitScheduler":
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback) -> None:
+        # Upload last changes before exiting
+        self.trigger().result()
+        self.stop()
+        return
+
+    def _run_scheduler(self) -> None:
+        """Dumb thread waiting between each scheduled push to Hub."""
+        while True:
+            self.last_future = self.trigger()
+            time.sleep(self.every * 60)
+            if self.__stopped:
+                break
+
+    def trigger(self) -> Future:
+        """Trigger a `push_to_hub` and return a future.
+
+        This method is automatically called every `every` minutes. You can also call it manually to trigger a commit
+        immediately, without waiting for the next scheduled commit.
+        """
+        return self.api.run_as_future(self._push_to_hub)
+
+    def _push_to_hub(self) -> CommitInfo | None:
+        if self.__stopped:  # If stopped, already scheduled commits are ignored
+            return None
+
+        logger.info("(Background) scheduled commit triggered.")
+        try:
+            value = self.push_to_hub()
+            if self.squash_history:
+                logger.info("(Background) squashing repo history.")
+                self.api.super_squash_history(
+                    repo_id=self.repo_id, repo_type=self.repo_type, branch=self.revision
+                )
+            return value
+        except Exception as e:
+            logger.error(
+                f"Error while pushing to Hub: {e}"
+            )  # Depending on the setup, error might be silenced
+            raise
+
+    def push_to_hub(self) -> CommitInfo | None:
+        """
+        Push folder to the Hub and return the commit info.
+
+        <Tip warning={true}>
+
+        This method is not meant to be called directly. It is run in the background by the scheduler, respecting a
+        queue mechanism to avoid concurrent commits. Making a direct call to the method might lead to concurrency
+        issues.
+
+        </Tip>
+
+        The default behavior of `push_to_hub` is to assume an append-only folder. It lists all files in the folder and
+        uploads only changed files. If no changes are found, the method returns without committing anything. If you want
+        to change this behavior, you can inherit from [`CommitScheduler`] and override this method. This can be useful
+        for example to compress data together in a single file before committing. For more details and examples, check
+        out our [integration guide](https://huggingface.co/docs/huggingface_hub/main/en/guides/upload#scheduled-uploads).
+        """
+        # Check files to upload (with lock)
+        with self.lock:
+            if self.on_before_commit is not None:
+                self.on_before_commit()
+
+            logger.debug("Listing files to upload for scheduled commit.")
+
+            # List files from folder (taken from `_prepare_upload_folder_additions`)
+            relpath_to_abspath = {
+                path.relative_to(self.folder_path).as_posix(): path
+                for path in sorted(
+                    self.folder_path.glob("**/*")
+                )  # sorted to be deterministic
+                if path.is_file()
+            }
+            prefix = f"{self.path_in_repo.strip('/')}/" if self.path_in_repo else ""
+
+            # Filter with pattern + filter out unchanged files + retrieve current file size
+            files_to_upload: List[_FileToUpload] = []
+            for relpath in filter_repo_objects(
+                relpath_to_abspath.keys(),
+                allow_patterns=self.allow_patterns,
+                ignore_patterns=self.ignore_patterns,
+            ):
+                local_path = relpath_to_abspath[relpath]
+                stat = local_path.stat()
+                if (
+                    self.last_uploaded.get(local_path) is None
+                    or self.last_uploaded[local_path] != stat.st_mtime
+                ):
+                    files_to_upload.append(
+                        _FileToUpload(
+                            local_path=local_path,
+                            path_in_repo=prefix + relpath,
+                            size_limit=stat.st_size,
+                            last_modified=stat.st_mtime,
+                        )
+                    )
+
+        # Return if nothing to upload
+        if len(files_to_upload) == 0:
+            logger.debug("Dropping schedule commit: no changed file to upload.")
+            return None
+
+        # Convert `_FileToUpload` as `CommitOperationAdd` (=> compute file shas + limit to file size)
+        logger.debug("Removing unchanged files since previous scheduled commit.")
+        add_operations = [
+            CommitOperationAdd(
+                # TODO: Cap the file to its current size, even if the user append data to it while a scheduled commit is happening
+                # (requires an upstream fix for XET-535: `hf_xet` should support `BinaryIO` for upload)
+                path_or_fileobj=file_to_upload.local_path,
+                path_in_repo=file_to_upload.path_in_repo,
+            )
+            for file_to_upload in files_to_upload
+        ]
+
+        # Upload files (append mode expected - no need for lock)
+        logger.debug("Uploading files for scheduled commit.")
+        commit_info = self.api.create_commit(
+            repo_id=self.repo_id,
+            repo_type=self.repo_type,
+            operations=add_operations,
+            commit_message="Scheduled Commit",
+            revision=self.revision,
+        )
+
+        for file in files_to_upload:
+            self.last_uploaded[file.local_path] = file.last_modified
+
+        self.last_push_time = time.time()
+
+        return commit_info
+
+
+class PartialFileIO(BytesIO):
+    """A file-like object that reads only the first part of a file.
+
+    Useful to upload a file to the Hub when the user might still be appending data to it. Only the first part of the
+    file is uploaded (i.e. the part that was available when the filesystem was first scanned).
+
+    In practice, only used internally by the CommitScheduler to regularly push a folder to the Hub with minimal
+    disturbance for the user. The object is passed to `CommitOperationAdd`.
+
+    Only supports `read`, `tell` and `seek` methods.
+
+    Args:
+        file_path (`str` or `Path`):
+            Path to the file to read.
+        size_limit (`int`):
+            The maximum number of bytes to read from the file. If the file is larger than this, only the first part
+            will be read (and uploaded).
+    """
+
+    def __init__(self, file_path: Union[str, Path], size_limit: int) -> None:
+        self._file_path = Path(file_path)
+        self._file = self._file_path.open("rb")
+        self._size_limit = min(size_limit, os.fstat(self._file.fileno()).st_size)
+
+    def __del__(self) -> None:
+        self._file.close()
+        return super().__del__()
+
+    def __repr__(self) -> str:
+        return (
+            f"<PartialFileIO file_path={self._file_path} size_limit={self._size_limit}>"
+        )
+
+    def __len__(self) -> int:
+        return self._size_limit
+
+    def __getattribute__(self, name: str):
+        if name.startswith("_") or name in (
+            "read",
+            "tell",
+            "seek",
+        ):  # only 3 public methods supported
+            return super().__getattribute__(name)
+        raise NotImplementedError(f"PartialFileIO does not support '{name}'.")
+
+    def tell(self) -> int:
+        """Return the current file position."""
+        return self._file.tell()
+
+    def seek(self, __offset: int, __whence: int = SEEK_SET) -> int:
+        """Change the stream position to the given offset.
+
+        Behavior is the same as a regular file, except that the position is capped to the size limit.
+        """
+        if __whence == SEEK_END:
+            # SEEK_END => set from the truncated end
+            __offset = len(self) + __offset
+            __whence = SEEK_SET
+
+        pos = self._file.seek(__offset, __whence)
+        if pos > self._size_limit:
+            return self._file.seek(self._size_limit)
+        return pos
+
+    def read(self, __size: int | None = -1) -> bytes:
+        """Read at most `__size` bytes from the file.
+
+        Behavior is the same as a regular file, except that it is capped to the size limit.
+        """
+        current = self._file.tell()
+        if __size is None or __size < 0:
+            # Read until file limit
+            truncated_size = self._size_limit - current
+        else:
+            # Read until file limit or __size
+            truncated_size = min(__size, self._size_limit - current)
+        return self._file.read(truncated_size)

context_vars.py

@@ -0,0 +1,18 @@
+import contextvars
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from trackio.run import Run
+
+current_run: contextvars.ContextVar["Run | None"] = contextvars.ContextVar(
+    "current_run", default=None
+)
+current_project: contextvars.ContextVar[str | None] = contextvars.ContextVar(
+    "current_project", default=None
+)
+current_server: contextvars.ContextVar[str | None] = contextvars.ContextVar(
+    "current_server", default=None
+)
+current_share_server: contextvars.ContextVar[str | None] = contextvars.ContextVar(
+    "current_share_server", default=None
+)

deploy.py

@@ -0,0 +1,225 @@
+import importlib.metadata
+import io
+import os
+import time
+from importlib.resources import files
+from pathlib import Path
+
+import gradio
+import huggingface_hub
+from gradio_client import Client, handle_file
+from httpx import ReadTimeout
+from huggingface_hub.errors import RepositoryNotFoundError
+from requests import HTTPError
+
+import trackio
+from trackio.sqlite_storage import SQLiteStorage
+
+SPACE_HOST_URL = "https://{user_name}-{space_name}.hf.space/"
+SPACE_URL = "https://huggingface.co/spaces/{space_id}"
+
+
+def _is_trackio_installed_from_source() -> bool:
+    """Check if trackio is installed from source/editable install vs PyPI."""
+    try:
+        trackio_file = trackio.__file__
+        if "site-packages" not in trackio_file:
+            return True
+
+        dist = importlib.metadata.distribution("trackio")
+        if dist.files:
+            files = list(dist.files)
+            has_pth = any(".pth" in str(f) for f in files)
+            if has_pth:
+                return True
+
+        return False
+    except (
+        AttributeError,
+        importlib.metadata.PackageNotFoundError,
+        importlib.metadata.MetadataError,
+        ValueError,
+        TypeError,
+    ):
+        return True
+
+
+def deploy_as_space(
+    space_id: str,
+    space_storage: huggingface_hub.SpaceStorage | None = None,
+    dataset_id: str | None = None,
+    private: bool | None = None,
+):
+    if (
+        os.getenv("SYSTEM") == "spaces"
+    ):  # in case a repo with this function is uploaded to spaces
+        return
+
+    trackio_path = files("trackio")
+
+    hf_api = huggingface_hub.HfApi()
+
+    try:
+        huggingface_hub.create_repo(
+            space_id,
+            private=private,
+            space_sdk="gradio",
+            space_storage=space_storage,
+            repo_type="space",
+            exist_ok=True,
+        )
+    except HTTPError as e:
+        if e.response.status_code in [401, 403]:  # unauthorized or forbidden
+            print("Need 'write' access token to create a Spaces repo.")
+            huggingface_hub.login(add_to_git_credential=False)
+            huggingface_hub.create_repo(
+                space_id,
+                private=private,
+                space_sdk="gradio",
+                space_storage=space_storage,
+                repo_type="space",
+                exist_ok=True,
+            )
+        else:
+            raise ValueError(f"Failed to create Space: {e}")
+
+    with open(Path(trackio_path, "README.md"), "r") as f:
+        readme_content = f.read()
+        readme_content = readme_content.replace("{GRADIO_VERSION}", gradio.__version__)
+        readme_buffer = io.BytesIO(readme_content.encode("utf-8"))
+        hf_api.upload_file(
+            path_or_fileobj=readme_buffer,
+            path_in_repo="README.md",
+            repo_id=space_id,
+            repo_type="space",
+        )
+
+    # We can assume pandas, gradio, and huggingface-hub are already installed in a Gradio Space.
+    # Make sure necessary dependencies are installed by creating a requirements.txt.
+    is_source_install = _is_trackio_installed_from_source()
+
+    if is_source_install:
+        requirements_content = """pyarrow>=21.0"""
+    else:
+        requirements_content = f"""pyarrow>=21.0
+trackio=={trackio.__version__}"""
+
+    requirements_buffer = io.BytesIO(requirements_content.encode("utf-8"))
+    hf_api.upload_file(
+        path_or_fileobj=requirements_buffer,
+        path_in_repo="requirements.txt",
+        repo_id=space_id,
+        repo_type="space",
+    )
+
+    huggingface_hub.utils.disable_progress_bars()
+
+    if is_source_install:
+        hf_api.upload_folder(
+            repo_id=space_id,
+            repo_type="space",
+            folder_path=trackio_path,
+            ignore_patterns=["README.md"],
+        )
+    else:
+        app_file_content = """import trackio
+trackio.show()"""
+        app_file_buffer = io.BytesIO(app_file_content.encode("utf-8"))
+        hf_api.upload_file(
+            path_or_fileobj=app_file_buffer,
+            path_in_repo="ui/main.py",
+            repo_id=space_id,
+            repo_type="space",
+        )
+
+    if hf_token := huggingface_hub.utils.get_token():
+        huggingface_hub.add_space_secret(space_id, "HF_TOKEN", hf_token)
+    if dataset_id is not None:
+        huggingface_hub.add_space_variable(space_id, "TRACKIO_DATASET_ID", dataset_id)
+
+
+def create_space_if_not_exists(
+    space_id: str,
+    space_storage: huggingface_hub.SpaceStorage | None = None,
+    dataset_id: str | None = None,
+    private: bool | None = None,
+) -> None:
+    """
+    Creates a new Hugging Face Space if it does not exist. If a dataset_id is provided, it will be added as a space variable.
+
+    Args:
+        space_id: The ID of the Space to create.
+        dataset_id: The ID of the Dataset to add to the Space.
+        private: Whether to make the Space private. If None (default), the repo will be
+          public unless the organization's default is private. This value is ignored if
+          the repo already exists.
+    """
+    if "/" not in space_id:
+        raise ValueError(
+            f"Invalid space ID: {space_id}. Must be in the format: username/reponame or orgname/reponame."
+        )
+    if dataset_id is not None and "/" not in dataset_id:
+        raise ValueError(
+            f"Invalid dataset ID: {dataset_id}. Must be in the format: username/datasetname or orgname/datasetname."
+        )
+    try:
+        huggingface_hub.repo_info(space_id, repo_type="space")
+        print(f"* Found existing space: {SPACE_URL.format(space_id=space_id)}")
+        if dataset_id is not None:
+            huggingface_hub.add_space_variable(
+                space_id, "TRACKIO_DATASET_ID", dataset_id
+            )
+        return
+    except RepositoryNotFoundError:
+        pass
+    except HTTPError as e:
+        if e.response.status_code in [401, 403]:  # unauthorized or forbidden
+            print("Need 'write' access token to create a Spaces repo.")
+            huggingface_hub.login(add_to_git_credential=False)
+            huggingface_hub.add_space_variable(
+                space_id, "TRACKIO_DATASET_ID", dataset_id
+            )
+        else:
+            raise ValueError(f"Failed to create Space: {e}")
+
+    print(f"* Creating new space: {SPACE_URL.format(space_id=space_id)}")
+    deploy_as_space(space_id, space_storage, dataset_id, private)
+
+
+def wait_until_space_exists(
+    space_id: str,
+) -> None:
+    """
+    Blocks the current thread until the space exists.
+    May raise a TimeoutError if this takes quite a while.
+
+    Args:
+        space_id: The ID of the Space to wait for.
+    """
+    delay = 1
+    for _ in range(10):
+        try:
+            Client(space_id, verbose=False)
+            return
+        except (ReadTimeout, ValueError):
+            time.sleep(delay)
+            delay = min(delay * 2, 30)
+    raise TimeoutError("Waiting for space to exist took longer than expected")
+
+
+def upload_db_to_space(project: str, space_id: str) -> None:
+    """
+    Uploads the database of a local Trackio project to a Hugging Face Space.
+
+    Args:
+        project: The name of the project to upload.
+        space_id: The ID of the Space to upload to.
+    """
+    db_path = SQLiteStorage.get_project_db_path(project)
+    client = Client(space_id, verbose=False)
+    client.predict(
+        api_name="/upload_db_to_space",
+        project=project,
+        uploaded_db=handle_file(db_path),
+        hf_token=huggingface_hub.utils.get_token(),
+    )

dummy_commit_scheduler.py

@@ -0,0 +1,12 @@
+# A dummy object to fit the interface of huggingface_hub's CommitScheduler
+class DummyCommitSchedulerLock:
+    def __enter__(self):
+        return None
+
+    def __exit__(self, exception_type, exception_value, exception_traceback):
+        pass
+
+
+class DummyCommitScheduler:
+    def __init__(self):
+        self.lock = DummyCommitSchedulerLock()

file_storage.py

@@ -0,0 +1,37 @@
+from pathlib import Path
+
+try:  # absolute imports when installed
+    from trackio.utils import MEDIA_DIR
+except ImportError:  # relative imports for local execution on Spaces
+    from utils import MEDIA_DIR
+
+
+class FileStorage:
+    @staticmethod
+    def get_project_media_path(
+        project: str,
+        run: str | None = None,
+        step: int | None = None,
+        filename: str | None = None,
+    ) -> Path:
+        if filename is not None and step is None:
+            raise ValueError("filename requires step")
+        if step is not None and run is None:
+            raise ValueError("step requires run")
+
+        path = MEDIA_DIR / project
+        if run:
+            path /= run
+        if step is not None:
+            path /= str(step)
+        if filename:
+            path /= filename
+        return path
+
+    @staticmethod
+    def init_project_media_path(
+        project: str, run: str | None = None, step: int | None = None
+    ) -> Path:
+        path = FileStorage.get_project_media_path(project, run, step)
+        path.mkdir(parents=True, exist_ok=True)
+        return path

imports.py

@@ -0,0 +1,302 @@
+import os
+from pathlib import Path
+
+import pandas as pd
+
+from trackio import deploy, utils
+from trackio.sqlite_storage import SQLiteStorage
+
+
+def import_csv(
+    csv_path: str | Path,
+    project: str,
+    name: str | None = None,
+    space_id: str | None = None,
+    dataset_id: str | None = None,
+    private: bool | None = None,
+) -> None:
+    """
+    Imports a CSV file into a Trackio project. The CSV file must contain a `"step"`
+    column, may optionally contain a `"timestamp"` column, and any other columns will be
+    treated as metrics. It should also include a header row with the column names.
+
+    TODO: call init() and return a Run object so that the user can continue to log metrics to it.
+
+    Args:
+        csv_path (`str` or `Path`):
+            The str or Path to the CSV file to import.
+        project (`str`):
+            The name of the project to import the CSV file into. Must not be an existing
+            project.
+        name (`str`, *optional*):
+            The name of the Run to import the CSV file into. If not provided, a default
+            name will be generated.
+        name (`str`, *optional*):
+            The name of the run (if not provided, a default name will be generated).
+        space_id (`str`, *optional*):
+            If provided, the project will be logged to a Hugging Face Space instead of a
+            local directory. Should be a complete Space name like `"username/reponame"`
+            or `"orgname/reponame"`, or just `"reponame"` in which case the Space will
+            be created in the currently-logged-in Hugging Face user's namespace. If the
+            Space does not exist, it will be created. If the Space already exists, the
+            project will be logged to it.
+        dataset_id (`str`, *optional*):
+            If provided, a persistent Hugging Face Dataset will be created and the
+            metrics will be synced to it every 5 minutes. Should be a complete Dataset
+            name like `"username/datasetname"` or `"orgname/datasetname"`, or just
+            `"datasetname"` in which case the Dataset will be created in the
+            currently-logged-in Hugging Face user's namespace. If the Dataset does not
+            exist, it will be created. If the Dataset already exists, the project will
+            be appended to it. If not provided, the metrics will be logged to a local
+            SQLite database, unless a `space_id` is provided, in which case a Dataset
+            will be automatically created with the same name as the Space but with the
+            `"_dataset"` suffix.
+        private (`bool`, *optional*):
+            Whether to make the Space private. If None (default), the repo will be
+            public unless the organization's default is private. This value is ignored
+            if the repo already exists.
+    """
+    if SQLiteStorage.get_runs(project):
+        raise ValueError(
+            f"Project '{project}' already exists. Cannot import CSV into existing project."
+        )
+
+    csv_path = Path(csv_path)
+    if not csv_path.exists():
+        raise FileNotFoundError(f"CSV file not found: {csv_path}")
+
+    df = pd.read_csv(csv_path)
+    if df.empty:
+        raise ValueError("CSV file is empty")
+
+    column_mapping = utils.simplify_column_names(df.columns.tolist())
+    df = df.rename(columns=column_mapping)
+
+    step_column = None
+    for col in df.columns:
+        if col.lower() == "step":
+            step_column = col
+            break
+
+    if step_column is None:
+        raise ValueError("CSV file must contain a 'step' or 'Step' column")
+
+    if name is None:
+        name = csv_path.stem
+
+    metrics_list = []
+    steps = []
+    timestamps = []
+
+    numeric_columns = []
+    for column in df.columns:
+        if column == step_column:
+            continue
+        if column == "timestamp":
+            continue
+
+        try:
+            pd.to_numeric(df[column], errors="raise")
+            numeric_columns.append(column)
+        except (ValueError, TypeError):
+            continue
+
+    for _, row in df.iterrows():
+        metrics = {}
+        for column in numeric_columns:
+            value = row[column]
+            if bool(pd.notna(value)):
+                metrics[column] = float(value)
+
+        if metrics:
+            metrics_list.append(metrics)
+            steps.append(int(row[step_column]))
+
+            if "timestamp" in df.columns and bool(pd.notna(row["timestamp"])):
+                timestamps.append(str(row["timestamp"]))
+            else:
+                timestamps.append("")
+
+    if metrics_list:
+        SQLiteStorage.bulk_log(
+            project=project,
+            run=name,
+            metrics_list=metrics_list,
+            steps=steps,
+            timestamps=timestamps,
+        )
+
+    print(
+        f"* Imported {len(metrics_list)} rows from {csv_path} into project '{project}' as run '{name}'"
+    )
+    print(f"* Metrics found: {', '.join(metrics_list[0].keys())}")
+
+    space_id, dataset_id = utils.preprocess_space_and_dataset_ids(space_id, dataset_id)
+    if dataset_id is not None:
+        os.environ["TRACKIO_DATASET_ID"] = dataset_id
+        print(f"* Trackio metrics will be synced to Hugging Face Dataset: {dataset_id}")
+
+    if space_id is None:
+        utils.print_dashboard_instructions(project)
+    else:
+        deploy.create_space_if_not_exists(
+            space_id=space_id, dataset_id=dataset_id, private=private
+        )
+        deploy.wait_until_space_exists(space_id=space_id)
+        deploy.upload_db_to_space(project=project, space_id=space_id)
+        print(
+            f"* View dashboard by going to: {deploy.SPACE_URL.format(space_id=space_id)}"
+        )
+
+
+def import_tf_events(
+    log_dir: str | Path,
+    project: str,
+    name: str | None = None,
+    space_id: str | None = None,
+    dataset_id: str | None = None,
+    private: bool | None = None,
+) -> None:
+    """
+    Imports TensorFlow Events files from a directory into a Trackio project. Each
+    subdirectory in the log directory will be imported as a separate run.
+
+    Args:
+        log_dir (`str` or `Path`):
+            The str or Path to the directory containing TensorFlow Events files.
+        project (`str`):
+            The name of the project to import the TensorFlow Events files into. Must not
+            be an existing project.
+        name (`str`, *optional*):
+            The name prefix for runs (if not provided, will use directory names). Each
+            subdirectory will create a separate run.
+        space_id (`str`, *optional*):
+            If provided, the project will be logged to a Hugging Face Space instead of a
+            local directory. Should be a complete Space name like `"username/reponame"`
+            or `"orgname/reponame"`, or just `"reponame"` in which case the Space will
+            be created in the currently-logged-in Hugging Face user's namespace. If the
+            Space does not exist, it will be created. If the Space already exists, the
+            project will be logged to it.
+        dataset_id (`str`, *optional*):
+            If provided, a persistent Hugging Face Dataset will be created and the
+            metrics will be synced to it every 5 minutes. Should be a complete Dataset
+            name like `"username/datasetname"` or `"orgname/datasetname"`, or just
+            `"datasetname"` in which case the Dataset will be created in the
+            currently-logged-in Hugging Face user's namespace. If the Dataset does not
+            exist, it will be created. If the Dataset already exists, the project will
+            be appended to it. If not provided, the metrics will be logged to a local
+            SQLite database, unless a `space_id` is provided, in which case a Dataset
+            will be automatically created with the same name as the Space but with the
+            `"_dataset"` suffix.
+        private (`bool`, *optional*):
+            Whether to make the Space private. If None (default), the repo will be
+            public unless the organization's default is private. This value is ignored
+            if the repo already exists.
+    """
+    try:
+        from tbparse import SummaryReader
+    except ImportError:
+        raise ImportError(
+            "The `tbparse` package is not installed but is required for `import_tf_events`. Please install trackio with the `tensorboard` extra: `pip install trackio[tensorboard]`."
+        )
+
+    if SQLiteStorage.get_runs(project):
+        raise ValueError(
+            f"Project '{project}' already exists. Cannot import TF events into existing project."
+        )
+
+    path = Path(log_dir)
+    if not path.exists():
+        raise FileNotFoundError(f"TF events directory not found: {path}")
+
+    # Use tbparse to read all tfevents files in the directory structure
+    reader = SummaryReader(str(path), extra_columns={"dir_name"})
+    df = reader.scalars
+
+    if df.empty:
+        raise ValueError(f"No TensorFlow events data found in {path}")
+
+    total_imported = 0
+    imported_runs = []
+
+    # Group by dir_name to create separate runs
+    for dir_name, group_df in df.groupby("dir_name"):
+        try:
+            # Determine run name based on directory name
+            if dir_name == "":
+                run_name = "main"  # For files in the root directory
+            else:
+                run_name = dir_name  # Use directory name
+
+            if name:
+                run_name = f"{name}_{run_name}"
+
+            if group_df.empty:
+                print(f"* Skipping directory {dir_name}: no scalar data found")
+                continue
+
+            metrics_list = []
+            steps = []
+            timestamps = []
+
+            for _, row in group_df.iterrows():
+                # Convert row values to appropriate types
+                tag = str(row["tag"])
+                value = float(row["value"])
+                step = int(row["step"])
+
+                metrics = {tag: value}
+                metrics_list.append(metrics)
+                steps.append(step)
+
+                # Use wall_time if present, else fallback
+                if "wall_time" in group_df.columns and not bool(
+                    pd.isna(row["wall_time"])
+                ):
+                    timestamps.append(str(row["wall_time"]))
+                else:
+                    timestamps.append("")
+
+            if metrics_list:
+                SQLiteStorage.bulk_log(
+                    project=project,
+                    run=str(run_name),
+                    metrics_list=metrics_list,
+                    steps=steps,
+                    timestamps=timestamps,
+                )
+
+                total_imported += len(metrics_list)
+                imported_runs.append(run_name)
+
+                print(
+                    f"* Imported {len(metrics_list)} scalar events from directory '{dir_name}' as run '{run_name}'"
+                )
+                print(f"* Metrics in this run: {', '.join(set(group_df['tag']))}")
+
+        except Exception as e:
+            print(f"* Error processing directory {dir_name}: {e}")
+            continue
+
+    if not imported_runs:
+        raise ValueError("No valid TensorFlow events data could be imported")
+
+    print(f"* Total imported events: {total_imported}")
+    print(f"* Created runs: {', '.join(imported_runs)}")
+
+    space_id, dataset_id = utils.preprocess_space_and_dataset_ids(space_id, dataset_id)
+    if dataset_id is not None:
+        os.environ["TRACKIO_DATASET_ID"] = dataset_id
+        print(f"* Trackio metrics will be synced to Hugging Face Dataset: {dataset_id}")
+
+    if space_id is None:
+        utils.print_dashboard_instructions(project)
+    else:
+        deploy.create_space_if_not_exists(
+            space_id, dataset_id=dataset_id, private=private
+        )
+        deploy.wait_until_space_exists(space_id)
+        deploy.upload_db_to_space(project, space_id)
+        print(
+            f"* View dashboard by going to: {deploy.SPACE_URL.format(space_id=space_id)}"
+        )

media.py

@@ -0,0 +1,286 @@
+import os
+import shutil
+import uuid
+from abc import ABC, abstractmethod
+from pathlib import Path
+from typing import Literal
+
+import numpy as np
+from PIL import Image as PILImage
+
+try:  # absolute imports when installed
+    from trackio.file_storage import FileStorage
+    from trackio.utils import MEDIA_DIR
+    from trackio.video_writer import write_video
+except ImportError:  # relative imports for local execution on Spaces
+    from file_storage import FileStorage
+    from utils import MEDIA_DIR
+    from video_writer import write_video
+
+
+class TrackioMedia(ABC):
+    """
+    Abstract base class for Trackio media objects
+    Provides shared functionality for file handling and serialization.
+    """
+
+    TYPE: str
+
+    def __init_subclass__(cls, **kwargs):
+        """Ensure subclasses define the TYPE attribute."""
+        super().__init_subclass__(**kwargs)
+        if not hasattr(cls, "TYPE") or cls.TYPE is None:
+            raise TypeError(f"Class {cls.__name__} must define TYPE attribute")
+
+    def __init__(self, value, caption: str | None = None):
+        self.caption = caption
+        self._value = value
+        self._file_path: Path | None = None
+
+        # Validate file existence for string/Path inputs
+        if isinstance(self._value, str | Path):
+            if not os.path.isfile(self._value):
+                raise ValueError(f"File not found: {self._value}")
+
+    def _file_extension(self) -> str:
+        if self._file_path:
+            return self._file_path.suffix[1:].lower()
+        if isinstance(self._value, str | Path):
+            path = Path(self._value)
+            return path.suffix[1:].lower()
+        if hasattr(self, "_format") and self._format:
+            return self._format
+        return "unknown"
+
+    def _get_relative_file_path(self) -> Path | None:
+        return self._file_path
+
+    def _get_absolute_file_path(self) -> Path | None:
+        if self._file_path:
+            return MEDIA_DIR / self._file_path
+        return None
+
+    def _save(self, project: str, run: str, step: int = 0):
+        if self._file_path:
+            return
+
+        media_dir = FileStorage.init_project_media_path(project, run, step)
+        filename = f"{uuid.uuid4()}.{self._file_extension()}"
+        file_path = media_dir / filename
+
+        # Delegate to subclass-specific save logic
+        self._save_media(file_path)
+
+        self._file_path = file_path.relative_to(MEDIA_DIR)
+
+    @abstractmethod
+    def _save_media(self, file_path: Path):
+        """
+        Performs the actual media saving logic.
+        """
+        pass
+
+    def _to_dict(self) -> dict:
+        if not self._file_path:
+            raise ValueError("Media must be saved to file before serialization")
+        return {
+            "_type": self.TYPE,
+            "file_path": str(self._get_relative_file_path()),
+            "caption": self.caption,
+        }
+
+
+TrackioImageSourceType = str | Path | np.ndarray | PILImage.Image
+
+
+class TrackioImage(TrackioMedia):
+    """
+    Initializes an Image object.
+
+    Example:
+        ```python
+        import trackio
+        import numpy as np
+        from PIL import Image
+
+        # Create an image from numpy array
+        image_data = np.random.randint(0, 255, (64, 64, 3), dtype=np.uint8)
+        image = trackio.Image(image_data, caption="Random image")
+        trackio.log({"my_image": image})
+
+        # Create an image from PIL Image
+        pil_image = Image.new('RGB', (100, 100), color='red')
+        image = trackio.Image(pil_image, caption="Red square")
+        trackio.log({"red_image": image})
+
+        # Create an image from file path
+        image = trackio.Image("path/to/image.jpg", caption="Photo from file")
+        trackio.log({"file_image": image})
+        ```
+
+    Args:
+        value (`str`, `Path`, `numpy.ndarray`, or `PIL.Image`, *optional*):
+            A path to an image, a PIL Image, or a numpy array of shape (height, width, channels).
+        caption (`str`, *optional*):
+            A string caption for the image.
+    """
+
+    TYPE = "trackio.image"
+
+    def __init__(self, value: TrackioImageSourceType, caption: str | None = None):
+        super().__init__(value, caption)
+        self._format: str | None = None
+
+        if (
+            isinstance(self._value, np.ndarray | PILImage.Image)
+            and self._format is None
+        ):
+            self._format = "png"
+
+    def _as_pil(self) -> PILImage.Image | None:
+        try:
+            if isinstance(self._value, np.ndarray):
+                arr = np.asarray(self._value).astype("uint8")
+                return PILImage.fromarray(arr).convert("RGBA")
+            if isinstance(self._value, PILImage.Image):
+                return self._value.convert("RGBA")
+        except Exception as e:
+            raise ValueError(f"Failed to process image data: {self._value}") from e
+        return None
+
+    def _save_media(self, file_path: Path):
+        if pil := self._as_pil():
+            pil.save(file_path, format=self._format)
+        elif isinstance(self._value, str | Path):
+            if os.path.isfile(self._value):
+                shutil.copy(self._value, file_path)
+            else:
+                raise ValueError(f"File not found: {self._value}")
+
+
+TrackioVideoSourceType = str | Path | np.ndarray
+TrackioVideoFormatType = Literal["gif", "mp4", "webm"]
+
+
+class TrackioVideo(TrackioMedia):
+    """
+    Initializes a Video object.
+
+    Example:
+        ```python
+        import trackio
+        import numpy as np
+
+        # Create a simple video from numpy array
+        frames = np.random.randint(0, 255, (10, 3, 64, 64), dtype=np.uint8)
+        video = trackio.Video(frames, caption="Random video", fps=30)
+
+        # Create a batch of videos
+        batch_frames = np.random.randint(0, 255, (3, 10, 3, 64, 64), dtype=np.uint8)
+        batch_video = trackio.Video(batch_frames, caption="Batch of videos", fps=15)
+
+        # Create video from file path
+        video = trackio.Video("path/to/video.mp4", caption="Video from file")
+        ```
+
+    Args:
+        value (`str`, `Path`, or `numpy.ndarray`, *optional*):
+            A path to a video file, or a numpy array.
+            The array should be of type `np.uint8` with RGB values in the range `[0, 255]`.
+            It is expected to have shape of either (frames, channels, height, width) or (batch, frames, channels, height, width).
+            For the latter, the videos will be tiled into a grid.
+        caption (`str`, *optional*):
+            A string caption for the video.
+        fps (`int`, *optional*):
+            Frames per second for the video. Only used when value is an ndarray. Default is `24`.
+        format (`Literal["gif", "mp4", "webm"]`, *optional*):
+            Video format ("gif", "mp4", or "webm"). Only used when value is an ndarray. Default is "gif".
+    """
+
+    TYPE = "trackio.video"
+
+    def __init__(
+        self,
+        value: TrackioVideoSourceType,
+        caption: str | None = None,
+        fps: int | None = None,
+        format: TrackioVideoFormatType | None = None,
+    ):
+        super().__init__(value, caption)
+        if isinstance(value, np.ndarray):
+            if format is None:
+                format = "gif"
+            if fps is None:
+                fps = 24
+        self._fps = fps
+        self._format = format
+
+    @property
+    def _codec(self) -> str:
+        match self._format:
+            case "gif":
+                return "gif"
+            case "mp4":
+                return "h264"
+            case "webm":
+                return "vp9"
+            case _:
+                raise ValueError(f"Unsupported format: {self._format}")
+
+    def _save_media(self, file_path: Path):
+        if isinstance(self._value, np.ndarray):
+            video = TrackioVideo._process_ndarray(self._value)
+            write_video(file_path, video, fps=self._fps, codec=self._codec)
+        elif isinstance(self._value, str | Path):
+            if os.path.isfile(self._value):
+                shutil.copy(self._value, file_path)
+            else:
+                raise ValueError(f"File not found: {self._value}")
+
+    @staticmethod
+    def _process_ndarray(value: np.ndarray) -> np.ndarray:
+        # Verify value is either 4D (single video) or 5D array (batched videos).
+        # Expected format: (frames, channels, height, width) or (batch, frames, channels, height, width)
+        if value.ndim < 4:
+            raise ValueError(
+                "Video requires at least 4 dimensions (frames, channels, height, width)"
+            )
+        if value.ndim > 5:
+            raise ValueError(
+                "Videos can have at most 5 dimensions (batch, frames, channels, height, width)"
+            )
+        if value.ndim == 4:
+            # Reshape to 5D with single batch: (1, frames, channels, height, width)
+            value = value[np.newaxis, ...]
+
+        value = TrackioVideo._tile_batched_videos(value)
+        return value
+
+    @staticmethod
+    def _tile_batched_videos(video: np.ndarray) -> np.ndarray:
+        """
+        Tiles a batch of videos into a grid of videos.
+
+        Input format: (batch, frames, channels, height, width) - original FCHW format
+        Output format: (frames, total_height, total_width, channels)
+        """
+        batch_size, frames, channels, height, width = video.shape
+
+        next_pow2 = 1 << (batch_size - 1).bit_length()
+        if batch_size != next_pow2:
+            pad_len = next_pow2 - batch_size
+            pad_shape = (pad_len, frames, channels, height, width)
+            padding = np.zeros(pad_shape, dtype=video.dtype)
+            video = np.concatenate((video, padding), axis=0)
+            batch_size = next_pow2
+
+        n_rows = 1 << ((batch_size.bit_length() - 1) // 2)
+        n_cols = batch_size // n_rows
+
+        # Reshape to grid layout: (n_rows, n_cols, frames, channels, height, width)
+        video = video.reshape(n_rows, n_cols, frames, channels, height, width)
+
+        # Rearrange dimensions to (frames, total_height, total_width, channels)
+        video = video.transpose(2, 0, 4, 1, 5, 3)
+        video = video.reshape(frames, n_rows * height, n_cols * width, channels)
+        return video

run.py

@@ -0,0 +1,180 @@
+import threading
+import time
+from datetime import datetime, timezone
+
+import huggingface_hub
+from gradio_client import Client, handle_file
+
+from trackio import utils
+from trackio.media import TrackioMedia
+from trackio.sqlite_storage import SQLiteStorage
+from trackio.table import Table
+from trackio.typehints import LogEntry, UploadEntry
+
+BATCH_SEND_INTERVAL = 0.5
+
+
+class Run:
+    def __init__(
+        self,
+        url: str,
+        project: str,
+        client: Client | None,
+        name: str | None = None,
+        group: str | None = None,
+        config: dict | None = None,
+        space_id: str | None = None,
+    ):
+        self.url = url
+        self.project = project
+        self._client_lock = threading.Lock()
+        self._client_thread = None
+        self._client = client
+        self._space_id = space_id
+        self.name = name or utils.generate_readable_name(
+            SQLiteStorage.get_runs(project), space_id
+        )
+        self.group = group
+        self.config = utils.to_json_safe(config or {})
+
+        if isinstance(self.config, dict):
+            for key in self.config:
+                if key.startswith("_"):
+                    raise ValueError(
+                        f"Config key '{key}' is reserved (keys starting with '_' are reserved for internal use)"
+                    )
+
+        self.config["_Username"] = self._get_username()
+        self.config["_Created"] = datetime.now(timezone.utc).isoformat()
+        self.config["_Group"] = self.group
+
+        self._queued_logs: list[LogEntry] = []
+        self._queued_uploads: list[UploadEntry] = []
+        self._stop_flag = threading.Event()
+        self._config_logged = False
+
+        self._client_thread = threading.Thread(target=self._init_client_background)
+        self._client_thread.daemon = True
+        self._client_thread.start()
+
+    def _get_username(self) -> str | None:
+        """Get the current HuggingFace username if logged in, otherwise None."""
+        try:
+            who = huggingface_hub.whoami()
+            return who["name"] if who else None
+        except Exception:
+            return None
+
+    def _batch_sender(self):
+        """Send batched logs every BATCH_SEND_INTERVAL."""
+        while not self._stop_flag.is_set() or len(self._queued_logs) > 0:
+            # If the stop flag has been set, then just quickly send all
+            # the logs and exit.
+            if not self._stop_flag.is_set():
+                time.sleep(BATCH_SEND_INTERVAL)
+
+            with self._client_lock:
+                if self._client is None:
+                    return
+                if self._queued_logs:
+                    logs_to_send = self._queued_logs.copy()
+                    self._queued_logs.clear()
+                    self._client.predict(
+                        api_name="/bulk_log",
+                        logs=logs_to_send,
+                        hf_token=huggingface_hub.utils.get_token(),
+                    )
+                if self._queued_uploads:
+                    uploads_to_send = self._queued_uploads.copy()
+                    self._queued_uploads.clear()
+                    self._client.predict(
+                        api_name="/bulk_upload_media",
+                        uploads=uploads_to_send,
+                        hf_token=huggingface_hub.utils.get_token(),
+                    )
+
+    def _init_client_background(self):
+        if self._client is None:
+            fib = utils.fibo()
+            for sleep_coefficient in fib:
+                try:
+                    client = Client(self.url, verbose=False)
+
+                    with self._client_lock:
+                        self._client = client
+                    break
+                except Exception:
+                    pass
+                if sleep_coefficient is not None:
+                    time.sleep(0.1 * sleep_coefficient)
+
+        self._batch_sender()
+
+    def _process_media(self, metrics, step: int | None) -> dict:
+        """
+        Serialize media in metrics and upload to space if needed.
+        """
+        serializable_metrics = {}
+        if not step:
+            step = 0
+        for key, value in metrics.items():
+            if isinstance(value, TrackioMedia):
+                value._save(self.project, self.name, step)
+                serializable_metrics[key] = value._to_dict()
+                if self._space_id:
+                    # Upload local media when deploying to space
+                    upload_entry: UploadEntry = {
+                        "project": self.project,
+                        "run": self.name,
+                        "step": step,
+                        "uploaded_file": handle_file(value._get_absolute_file_path()),
+                    }
+                    with self._client_lock:
+                        self._queued_uploads.append(upload_entry)
+            else:
+                serializable_metrics[key] = value
+        return serializable_metrics
+
+    @staticmethod
+    def _replace_tables(metrics):
+        for k, v in metrics.items():
+            if isinstance(v, Table):
+                metrics[k] = v._to_dict()
+
+    def log(self, metrics: dict, step: int | None = None):
+        for k in metrics.keys():
+            if k in utils.RESERVED_KEYS or k.startswith("__"):
+                raise ValueError(
+                    f"Please do not use this reserved key as a metric: {k}"
+                )
+        Run._replace_tables(metrics)
+
+        metrics = self._process_media(metrics, step)
+        metrics = utils.serialize_values(metrics)
+
+        config_to_log = None
+        if not self._config_logged and self.config:
+            config_to_log = utils.to_json_safe(self.config)
+            self._config_logged = True
+
+        log_entry: LogEntry = {
+            "project": self.project,
+            "run": self.name,
+            "metrics": metrics,
+            "step": step,
+            "config": config_to_log,
+        }
+
+        with self._client_lock:
+            self._queued_logs.append(log_entry)
+
+    def finish(self):
+        """Cleanup when run is finished."""
+        self._stop_flag.set()
+
+        # Wait for the batch sender to finish before joining the client thread.
+        time.sleep(2 * BATCH_SEND_INTERVAL)
+
+        if self._client_thread is not None:
+            print("* Run finished. Uploading logs to Trackio (please wait...)")
+            self._client_thread.join()

sqlite_storage.py

@@ -0,0 +1,580 @@
+import os
+import platform
+import sqlite3
+import time
+from datetime import datetime
+from pathlib import Path
+from threading import Lock
+
+try:
+    import fcntl
+except ImportError:  # fcntl is not available on Windows
+    fcntl = None
+
+import huggingface_hub as hf
+import orjson
+import pandas as pd
+
+try:  # absolute imports when installed from PyPI
+    from trackio.commit_scheduler import CommitScheduler
+    from trackio.dummy_commit_scheduler import DummyCommitScheduler
+    from trackio.utils import (
+        TRACKIO_DIR,
+        deserialize_values,
+        serialize_values,
+    )
+except ImportError:  # relative imports when installed from source on Spaces
+    from commit_scheduler import CommitScheduler
+    from dummy_commit_scheduler import DummyCommitScheduler
+    from utils import TRACKIO_DIR, deserialize_values, serialize_values
+
+
+class ProcessLock:
+    """A file-based lock that works across processes. Is a no-op on Windows."""
+
+    def __init__(self, lockfile_path: Path):
+        self.lockfile_path = lockfile_path
+        self.lockfile = None
+        self.is_windows = platform.system() == "Windows"
+
+    def __enter__(self):
+        """Acquire the lock with retry logic."""
+        if self.is_windows:
+            return self
+        self.lockfile_path.parent.mkdir(parents=True, exist_ok=True)
+        self.lockfile = open(self.lockfile_path, "w")
+
+        max_retries = 100
+        for attempt in range(max_retries):
+            try:
+                fcntl.flock(self.lockfile.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+                return self
+            except IOError:
+                if attempt < max_retries - 1:
+                    time.sleep(0.1)
+                else:
+                    raise IOError("Could not acquire database lock after 10 seconds")
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        """Release the lock."""
+        if self.is_windows:
+            return
+
+        if self.lockfile:
+            fcntl.flock(self.lockfile.fileno(), fcntl.LOCK_UN)
+            self.lockfile.close()
+
+
+class SQLiteStorage:
+    _dataset_import_attempted = False
+    _current_scheduler: CommitScheduler | DummyCommitScheduler | None = None
+    _scheduler_lock = Lock()
+
+    @staticmethod
+    def _get_connection(db_path: Path) -> sqlite3.Connection:
+        conn = sqlite3.connect(str(db_path), timeout=30.0)
+        conn.execute("PRAGMA journal_mode = WAL")
+        conn.row_factory = sqlite3.Row
+        return conn
+
+    @staticmethod
+    def _get_process_lock(project: str) -> ProcessLock:
+        lockfile_path = TRACKIO_DIR / f"{project}.lock"
+        return ProcessLock(lockfile_path)
+
+    @staticmethod
+    def get_project_db_filename(project: str) -> Path:
+        """Get the database filename for a specific project."""
+        safe_project_name = "".join(
+            c for c in project if c.isalnum() or c in ("-", "_")
+        ).rstrip()
+        if not safe_project_name:
+            safe_project_name = "default"
+        return f"{safe_project_name}.db"
+
+    @staticmethod
+    def get_project_db_path(project: str) -> Path:
+        """Get the database path for a specific project."""
+        filename = SQLiteStorage.get_project_db_filename(project)
+        return TRACKIO_DIR / filename
+
+    @staticmethod
+    def init_db(project: str) -> Path:
+        """
+        Initialize the SQLite database with required tables.
+        If there is a dataset ID provided, copies from that dataset instead.
+        Returns the database path.
+        """
+        db_path = SQLiteStorage.get_project_db_path(project)
+        db_path.parent.mkdir(parents=True, exist_ok=True)
+        with SQLiteStorage._get_process_lock(project):
+            with sqlite3.connect(db_path, timeout=30.0) as conn:
+                conn.execute("PRAGMA journal_mode = WAL")
+                cursor = conn.cursor()
+                cursor.execute("""
+                    CREATE TABLE IF NOT EXISTS metrics (
+                        id INTEGER PRIMARY KEY AUTOINCREMENT,
+                        timestamp TEXT NOT NULL,
+                        run_name TEXT NOT NULL,
+                        step INTEGER NOT NULL,
+                        metrics TEXT NOT NULL
+                    )
+                """)
+                cursor.execute("""
+                    CREATE TABLE IF NOT EXISTS configs (
+                        id INTEGER PRIMARY KEY AUTOINCREMENT,
+                        run_name TEXT NOT NULL,
+                        config TEXT NOT NULL,
+                        created_at TEXT NOT NULL,
+                        UNIQUE(run_name)
+                    )
+                """)
+                cursor.execute(
+                    """
+                    CREATE INDEX IF NOT EXISTS idx_metrics_run_step
+                    ON metrics(run_name, step)
+                    """
+                )
+                cursor.execute(
+                    """
+                    CREATE INDEX IF NOT EXISTS idx_configs_run_name
+                    ON configs(run_name)
+                    """
+                )
+                cursor.execute(
+                    """
+                    CREATE INDEX IF NOT EXISTS idx_metrics_run_timestamp
+                    ON metrics(run_name, timestamp)
+                    """
+                )
+                conn.commit()
+        return db_path
+
+    @staticmethod
+    def export_to_parquet():
+        """
+        Exports all projects' DB files as Parquet under the same path but with extension ".parquet".
+        """
+        # don't attempt to export (potentially wrong/blank) data before importing for the first time
+        if not SQLiteStorage._dataset_import_attempted:
+            return
+        all_paths = os.listdir(TRACKIO_DIR)
+        db_paths = [f for f in all_paths if f.endswith(".db")]
+        for db_path in db_paths:
+            db_path = TRACKIO_DIR / db_path
+            parquet_path = db_path.with_suffix(".parquet")
+            if (not parquet_path.exists()) or (
+                db_path.stat().st_mtime > parquet_path.stat().st_mtime
+            ):
+                with sqlite3.connect(db_path) as conn:
+                    df = pd.read_sql("SELECT * from metrics", conn)
+                # break out the single JSON metrics column into individual columns
+                metrics = df["metrics"].copy()
+                metrics = pd.DataFrame(
+                    metrics.apply(
+                        lambda x: deserialize_values(orjson.loads(x))
+                    ).values.tolist(),
+                    index=df.index,
+                )
+                del df["metrics"]
+                for col in metrics.columns:
+                    df[col] = metrics[col]
+                df.to_parquet(parquet_path)
+
+    @staticmethod
+    def import_from_parquet():
+        """
+        Imports to all DB files that have matching files under the same path but with extension ".parquet".
+        """
+        all_paths = os.listdir(TRACKIO_DIR)
+        parquet_paths = [f for f in all_paths if f.endswith(".parquet")]
+        for parquet_path in parquet_paths:
+            parquet_path = TRACKIO_DIR / parquet_path
+            db_path = parquet_path.with_suffix(".db")
+            df = pd.read_parquet(parquet_path)
+            with sqlite3.connect(db_path) as conn:
+                # fix up df to have a single JSON metrics column
+                if "metrics" not in df.columns:
+                    # separate other columns from metrics
+                    metrics = df.copy()
+                    other_cols = ["id", "timestamp", "run_name", "step"]
+                    df = df[other_cols]
+                    for col in other_cols:
+                        del metrics[col]
+                    # combine them all into a single metrics col
+                    metrics = orjson.loads(metrics.to_json(orient="records"))
+                    df["metrics"] = [
+                        orjson.dumps(serialize_values(row)) for row in metrics
+                    ]
+                df.to_sql("metrics", conn, if_exists="replace", index=False)
+
+    @staticmethod
+    def get_scheduler():
+        """
+        Get the scheduler for the database based on the environment variables.
+        This applies to both local and Spaces.
+        """
+        with SQLiteStorage._scheduler_lock:
+            if SQLiteStorage._current_scheduler is not None:
+                return SQLiteStorage._current_scheduler
+            hf_token = os.environ.get("HF_TOKEN")
+            dataset_id = os.environ.get("TRACKIO_DATASET_ID")
+            space_repo_name = os.environ.get("SPACE_REPO_NAME")
+            if dataset_id is None or space_repo_name is None:
+                scheduler = DummyCommitScheduler()
+            else:
+                scheduler = CommitScheduler(
+                    repo_id=dataset_id,
+                    repo_type="dataset",
+                    folder_path=TRACKIO_DIR,
+                    private=True,
+                    allow_patterns=["*.parquet", "media/**/*"],
+                    squash_history=True,
+                    token=hf_token,
+                    on_before_commit=SQLiteStorage.export_to_parquet,
+                )
+            SQLiteStorage._current_scheduler = scheduler
+            return scheduler
+
+    @staticmethod
+    def log(project: str, run: str, metrics: dict, step: int | None = None):
+        """
+        Safely log metrics to the database. Before logging, this method will ensure the database exists
+        and is set up with the correct tables. It also uses a cross-process lock to prevent
+        database locking errors when multiple processes access the same database.
+
+        This method is not used in the latest versions of Trackio (replaced by bulk_log) but
+        is kept for backwards compatibility for users who are connecting to a newer version of
+        a Trackio Spaces dashboard with an older version of Trackio installed locally.
+        """
+        db_path = SQLiteStorage.init_db(project)
+
+        with SQLiteStorage._get_process_lock(project):
+            with SQLiteStorage._get_connection(db_path) as conn:
+                cursor = conn.cursor()
+
+                cursor.execute(
+                    """
+                    SELECT MAX(step) 
+                    FROM metrics 
+                    WHERE run_name = ?
+                    """,
+                    (run,),
+                )
+                last_step = cursor.fetchone()[0]
+                if step is None:
+                    current_step = 0 if last_step is None else last_step + 1
+                else:
+                    current_step = step
+
+                current_timestamp = datetime.now().isoformat()
+
+                cursor.execute(
+                    """
+                    INSERT INTO metrics
+                    (timestamp, run_name, step, metrics)
+                    VALUES (?, ?, ?, ?)
+                    """,
+                    (
+                        current_timestamp,
+                        run,
+                        current_step,
+                        orjson.dumps(serialize_values(metrics)),
+                    ),
+                )
+                conn.commit()
+
+    @staticmethod
+    def bulk_log(
+        project: str,
+        run: str,
+        metrics_list: list[dict],
+        steps: list[int] | None = None,
+        timestamps: list[str] | None = None,
+        config: dict | None = None,
+    ):
+        """
+        Safely log bulk metrics to the database. Before logging, this method will ensure the database exists
+        and is set up with the correct tables. It also uses a cross-process lock to prevent
+        database locking errors when multiple processes access the same database.
+        """
+        if not metrics_list:
+            return
+
+        if timestamps is None:
+            timestamps = [datetime.now().isoformat()] * len(metrics_list)
+
+        db_path = SQLiteStorage.init_db(project)
+        with SQLiteStorage._get_process_lock(project):
+            with SQLiteStorage._get_connection(db_path) as conn:
+                cursor = conn.cursor()
+
+                if steps is None:
+                    steps = list(range(len(metrics_list)))
+                elif any(s is None for s in steps):
+                    cursor.execute(
+                        "SELECT MAX(step) FROM metrics WHERE run_name = ?", (run,)
+                    )
+                    last_step = cursor.fetchone()[0]
+                    current_step = 0 if last_step is None else last_step + 1
+
+                    processed_steps = []
+                    for step in steps:
+                        if step is None:
+                            processed_steps.append(current_step)
+                            current_step += 1
+                        else:
+                            processed_steps.append(step)
+                    steps = processed_steps
+
+                if len(metrics_list) != len(steps) or len(metrics_list) != len(
+                    timestamps
+                ):
+                    raise ValueError(
+                        "metrics_list, steps, and timestamps must have the same length"
+                    )
+
+                data = []
+                for i, metrics in enumerate(metrics_list):
+                    data.append(
+                        (
+                            timestamps[i],
+                            run,
+                            steps[i],
+                            orjson.dumps(serialize_values(metrics)),
+                        )
+                    )
+
+                cursor.executemany(
+                    """
+                    INSERT INTO metrics
+                    (timestamp, run_name, step, metrics)
+                    VALUES (?, ?, ?, ?)
+                    """,
+                    data,
+                )
+
+                if config:
+                    current_timestamp = datetime.now().isoformat()
+                    cursor.execute(
+                        """
+                        INSERT OR REPLACE INTO configs
+                        (run_name, config, created_at)
+                        VALUES (?, ?, ?)
+                        """,
+                        (
+                            run,
+                            orjson.dumps(serialize_values(config)),
+                            current_timestamp,
+                        ),
+                    )
+
+                conn.commit()
+
+    @staticmethod
+    def get_logs(project: str, run: str) -> list[dict]:
+        """Retrieve logs for a specific run. Logs include the step count (int) and the timestamp (datetime object)."""
+        db_path = SQLiteStorage.get_project_db_path(project)
+        if not db_path.exists():
+            return []
+
+        with SQLiteStorage._get_connection(db_path) as conn:
+            cursor = conn.cursor()
+            cursor.execute(
+                """
+                SELECT timestamp, step, metrics
+                FROM metrics
+                WHERE run_name = ?
+                ORDER BY timestamp
+                """,
+                (run,),
+            )
+
+            rows = cursor.fetchall()
+            results = []
+            for row in rows:
+                metrics = orjson.loads(row["metrics"])
+                metrics = deserialize_values(metrics)
+                metrics["timestamp"] = row["timestamp"]
+                metrics["step"] = row["step"]
+                results.append(metrics)
+            return results
+
+    @staticmethod
+    def load_from_dataset():
+        dataset_id = os.environ.get("TRACKIO_DATASET_ID")
+        space_repo_name = os.environ.get("SPACE_REPO_NAME")
+        if dataset_id is not None and space_repo_name is not None:
+            hfapi = hf.HfApi()
+            updated = False
+            if not TRACKIO_DIR.exists():
+                TRACKIO_DIR.mkdir(parents=True, exist_ok=True)
+            with SQLiteStorage.get_scheduler().lock:
+                try:
+                    files = hfapi.list_repo_files(dataset_id, repo_type="dataset")
+                    for file in files:
+                        # Download parquet and media assets
+                        if not (file.endswith(".parquet") or file.startswith("media/")):
+                            continue
+                        if (TRACKIO_DIR / file).exists():
+                            continue
+                        hf.hf_hub_download(
+                            dataset_id, file, repo_type="dataset", local_dir=TRACKIO_DIR
+                        )
+                        updated = True
+                except hf.errors.EntryNotFoundError:
+                    pass
+                except hf.errors.RepositoryNotFoundError:
+                    pass
+                if updated:
+                    SQLiteStorage.import_from_parquet()
+        SQLiteStorage._dataset_import_attempted = True
+
+    @staticmethod
+    def get_projects() -> list[str]:
+        """
+        Get list of all projects by scanning the database files in the trackio directory.
+        """
+        if not SQLiteStorage._dataset_import_attempted:
+            SQLiteStorage.load_from_dataset()
+
+        projects: set[str] = set()
+        if not TRACKIO_DIR.exists():
+            return []
+
+        for db_file in TRACKIO_DIR.glob("*.db"):
+            project_name = db_file.stem
+            projects.add(project_name)
+        return sorted(projects)
+
+    @staticmethod
+    def get_runs(project: str) -> list[str]:
+        """Get list of all runs for a project."""
+        db_path = SQLiteStorage.get_project_db_path(project)
+        if not db_path.exists():
+            return []
+
+        with SQLiteStorage._get_connection(db_path) as conn:
+            cursor = conn.cursor()
+            cursor.execute(
+                "SELECT DISTINCT run_name FROM metrics",
+            )
+            return [row[0] for row in cursor.fetchall()]
+
+    @staticmethod
+    def get_max_steps_for_runs(project: str) -> dict[str, int]:
+        """Get the maximum step for each run in a project."""
+        db_path = SQLiteStorage.get_project_db_path(project)
+        if not db_path.exists():
+            return {}
+
+        with SQLiteStorage._get_connection(db_path) as conn:
+            cursor = conn.cursor()
+            cursor.execute(
+                """
+                SELECT run_name, MAX(step) as max_step
+                FROM metrics
+                GROUP BY run_name
+                """
+            )
+
+            results = {}
+            for row in cursor.fetchall():
+                results[row["run_name"]] = row["max_step"]
+
+            return results
+
+    @staticmethod
+    def store_config(project: str, run: str, config: dict) -> None:
+        """Store configuration for a run."""
+        db_path = SQLiteStorage.init_db(project)
+
+        with SQLiteStorage._get_process_lock(project):
+            with SQLiteStorage._get_connection(db_path) as conn:
+                cursor = conn.cursor()
+                current_timestamp = datetime.now().isoformat()
+
+                cursor.execute(
+                    """
+                    INSERT OR REPLACE INTO configs
+                    (run_name, config, created_at)
+                    VALUES (?, ?, ?)
+                    """,
+                    (run, orjson.dumps(serialize_values(config)), current_timestamp),
+                )
+                conn.commit()
+
+    @staticmethod
+    def get_run_config(project: str, run: str) -> dict | None:
+        """Get configuration for a specific run."""
+        db_path = SQLiteStorage.get_project_db_path(project)
+        if not db_path.exists():
+            return None
+
+        with SQLiteStorage._get_connection(db_path) as conn:
+            cursor = conn.cursor()
+            try:
+                cursor.execute(
+                    """
+                    SELECT config FROM configs WHERE run_name = ?
+                    """,
+                    (run,),
+                )
+
+                row = cursor.fetchone()
+                if row:
+                    config = orjson.loads(row["config"])
+                    return deserialize_values(config)
+                return None
+            except sqlite3.OperationalError as e:
+                if "no such table: configs" in str(e):
+                    return None
+                raise
+
+    @staticmethod
+    def delete_run(project: str, run: str) -> bool:
+        """Delete a run from the database (both metrics and config)."""
+        db_path = SQLiteStorage.get_project_db_path(project)
+        if not db_path.exists():
+            return False
+
+        with SQLiteStorage._get_process_lock(project):
+            with SQLiteStorage._get_connection(db_path) as conn:
+                cursor = conn.cursor()
+                try:
+                    cursor.execute("DELETE FROM metrics WHERE run_name = ?", (run,))
+                    cursor.execute("DELETE FROM configs WHERE run_name = ?", (run,))
+                    conn.commit()
+                    return True
+                except sqlite3.Error:
+                    return False
+
+    @staticmethod
+    def get_all_run_configs(project: str) -> dict[str, dict]:
+        """Get configurations for all runs in a project."""
+        db_path = SQLiteStorage.get_project_db_path(project)
+        if not db_path.exists():
+            return {}
+
+        with SQLiteStorage._get_connection(db_path) as conn:
+            cursor = conn.cursor()
+            try:
+                cursor.execute(
+                    """
+                    SELECT run_name, config FROM configs
+                    """
+                )
+
+                results = {}
+                for row in cursor.fetchall():
+                    config = orjson.loads(row["config"])
+                    results[row["run_name"]] = deserialize_values(config)
+                return results
+            except sqlite3.OperationalError as e:
+                if "no such table: configs" in str(e):
+                    return {}
+                raise
+
+    def finish(self):
+        """Cleanup when run is finished."""
+        pass

table.py

@@ -0,0 +1,53 @@
+from typing import Any, Literal
+
+from pandas import DataFrame
+
+
+class Table:
+    """
+    Initializes a Table object.
+
+    Args:
+        columns (`list[str]`, *optional*):
+            Names of the columns in the table. Optional if `data` is provided. Not
+            expected if `dataframe` is provided. Currently ignored.
+        data (`list[list[Any]]`, *optional*):
+            2D row-oriented array of values.
+        dataframe (`pandas.`DataFrame``, *optional*):
+            DataFrame object used to create the table. When set, `data` and `columns`
+            arguments are ignored.
+        rows (`list[list[any]]`, *optional*):
+            Currently ignored.
+        optional (`bool` or `list[bool]`, *optional*, defaults to `True`):
+            Currently ignored.
+        allow_mixed_types (`bool`, *optional*, defaults to `False`):
+            Currently ignored.
+        log_mode: (`Literal["IMMUTABLE", "MUTABLE", "INCREMENTAL"]` or `None`, *optional*, defaults to `"IMMUTABLE"`):
+            Currently ignored.
+    """
+
+    TYPE = "trackio.table"
+
+    def __init__(
+        self,
+        columns: list[str] | None = None,
+        data: list[list[Any]] | None = None,
+        dataframe: DataFrame | None = None,
+        rows: list[list[Any]] | None = None,
+        optional: bool | list[bool] = True,
+        allow_mixed_types: bool = False,
+        log_mode: Literal["IMMUTABLE", "MUTABLE", "INCREMENTAL"] | None = "IMMUTABLE",
+    ):
+        # TODO: implement support for columns, dtype, optional, allow_mixed_types, and log_mode.
+        # for now (like `rows`) they are included for API compat but don't do anything.
+
+        if dataframe is None:
+            self.data = data
+        else:
+            self.data = dataframe.to_dict(orient="records")
+
+    def _to_dict(self):
+        return {
+            "_type": self.TYPE,
+            "_value": self.data,
+        }

typehints.py

@@ -0,0 +1,18 @@
+from typing import Any, TypedDict
+
+from gradio import FileData
+
+
+class LogEntry(TypedDict):
+    project: str
+    run: str
+    metrics: dict[str, Any]
+    step: int | None
+    config: dict[str, Any] | None
+
+
+class UploadEntry(TypedDict):
+    project: str
+    run: str
+    step: int | None
+    uploaded_file: FileData