Skip to content

Reference for ultralytics/hub/session.py

Note

This file is available at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/hub/session.py. If you spot a problem please help fix it by contributing a Pull Request 🛠️. Thank you 🙏!


ultralytics.hub.session.HUBTrainingSession

HUBTrainingSession(identifier)

HUB training session for Ultralytics HUB YOLO models. Handles model initialization, heartbeats, and checkpointing.

Attributes:

Name Type Description
model_id str

Identifier for the YOLO model being trained.

model_url str

URL for the model in Ultralytics HUB.

rate_limits dict

Rate limits for different API calls (in seconds).

timers dict

Timers for rate limiting.

metrics_queue dict

Queue for the model's metrics.

model dict

Model data fetched from Ultralytics HUB.

Parameters:

Name Type Description Default
identifier str

Model identifier used to initialize the HUB training session. It can be a URL string or a model key with specific format.

required

Raises:

Type Description
ValueError

If the provided model identifier is invalid.

ConnectionError

If connecting with global API key is not supported.

ModuleNotFoundError

If hub-sdk package is not installed.

Source code in ultralytics/hub/session.py
def __init__(self, identifier):
    """
    Initialize the HUBTrainingSession with the provided model identifier.

    Args:
        identifier (str): Model identifier used to initialize the HUB training session.
            It can be a URL string or a model key with specific format.

    Raises:
        ValueError: If the provided model identifier is invalid.
        ConnectionError: If connecting with global API key is not supported.
        ModuleNotFoundError: If hub-sdk package is not installed.
    """
    from hub_sdk import HUBClient

    self.rate_limits = {"metrics": 3, "ckpt": 900, "heartbeat": 300}  # rate limits (seconds)
    self.metrics_queue = {}  # holds metrics for each epoch until upload
    self.metrics_upload_failed_queue = {}  # holds metrics for each epoch if upload failed
    self.timers = {}  # holds timers in ultralytics/utils/callbacks/hub.py
    self.model = None
    self.model_url = None

    # Parse input
    api_key, model_id, self.filename = self._parse_identifier(identifier)

    # Get credentials
    active_key = api_key or SETTINGS.get("api_key")
    credentials = {"api_key": active_key} if active_key else None  # set credentials

    # Initialize client
    self.client = HUBClient(credentials)

    # Load models if authenticated
    if self.client.authenticated:
        if model_id:
            self.load_model(model_id)  # load existing model
        else:
            self.model = self.client.model()  # load empty model

create_model

create_model(model_args)

Initializes a HUB training session with the specified model identifier.

Source code in ultralytics/hub/session.py
def create_model(self, model_args):
    """Initializes a HUB training session with the specified model identifier."""
    payload = {
        "config": {
            "batchSize": model_args.get("batch", -1),
            "epochs": model_args.get("epochs", 300),
            "imageSize": model_args.get("imgsz", 640),
            "patience": model_args.get("patience", 100),
            "device": str(model_args.get("device", "")),  # convert None to string
            "cache": str(model_args.get("cache", "ram")),  # convert True, False, None to string
        },
        "dataset": {"name": model_args.get("data")},
        "lineage": {
            "architecture": {"name": self.filename.replace(".pt", "").replace(".yaml", "")},
            "parent": {},
        },
        "meta": {"name": self.filename},
    }

    if self.filename.endswith(".pt"):
        payload["lineage"]["parent"]["name"] = self.filename

    self.model.create_model(payload)

    # Model could not be created
    # TODO: improve error handling
    if not self.model.id:
        return None

    self.model_url = f"{HUB_WEB_ROOT}/models/{self.model.id}"

    # Start heartbeats for HUB to monitor agent
    self.model.start_heartbeat(self.rate_limits["heartbeat"])

    LOGGER.info(f"{PREFIX}View model at {self.model_url} 🚀")

create_session classmethod

create_session(identifier, args=None)

Class method to create an authenticated HUBTrainingSession or return None.

Source code in ultralytics/hub/session.py
@classmethod
def create_session(cls, identifier, args=None):
    """Class method to create an authenticated HUBTrainingSession or return None."""
    try:
        session = cls(identifier)
        if not session.client.authenticated:
            if identifier.startswith(f"{HUB_WEB_ROOT}/models/"):
                LOGGER.warning(f"{PREFIX}WARNING ⚠️ Login to Ultralytics HUB with 'yolo hub login API_KEY'.")
                exit()
            return None
        if args and not identifier.startswith(f"{HUB_WEB_ROOT}/models/"):  # not a HUB model URL
            session.create_model(args)
            assert session.model.id, "HUB model not loaded correctly"
        return session
    # PermissionError and ModuleNotFoundError indicate hub-sdk not installed
    except (PermissionError, ModuleNotFoundError, AssertionError):
        return None

load_model

load_model(model_id)

Loads an existing model from Ultralytics HUB using the provided model identifier.

Source code in ultralytics/hub/session.py
def load_model(self, model_id):
    """Loads an existing model from Ultralytics HUB using the provided model identifier."""
    self.model = self.client.model(model_id)
    if not self.model.data:  # then model does not exist
        raise ValueError(emojis("❌ The specified HUB model does not exist"))  # TODO: improve error handling

    self.model_url = f"{HUB_WEB_ROOT}/models/{self.model.id}"

    self._set_train_args()

    # Start heartbeats for HUB to monitor agent
    self.model.start_heartbeat(self.rate_limits["heartbeat"])
    LOGGER.info(f"{PREFIX}View model at {self.model_url} 🚀")

request_queue

request_queue(request_func, retry=3, timeout=30, thread=True, verbose=True, progress_total=None, stream_response=None, *args, **kwargs)

Attempts to execute request_func with retries, timeout handling, optional threading, and progress.

Source code in ultralytics/hub/session.py
def request_queue(
    self,
    request_func,
    retry=3,
    timeout=30,
    thread=True,
    verbose=True,
    progress_total=None,
    stream_response=None,
    *args,
    **kwargs,
):
    """Attempts to execute `request_func` with retries, timeout handling, optional threading, and progress."""

    def retry_request():
        """Attempts to call `request_func` with retries, timeout, and optional threading."""
        t0 = time.time()  # Record the start time for the timeout
        response = None
        for i in range(retry + 1):
            if (time.time() - t0) > timeout:
                LOGGER.warning(f"{PREFIX}Timeout for request reached. {HELP_MSG}")
                break  # Timeout reached, exit loop

            response = request_func(*args, **kwargs)
            if response is None:
                LOGGER.warning(f"{PREFIX}Received no response from the request. {HELP_MSG}")
                time.sleep(2**i)  # Exponential backoff before retrying
                continue  # Skip further processing and retry

            if progress_total:
                self._show_upload_progress(progress_total, response)
            elif stream_response:
                self._iterate_content(response)

            if HTTPStatus.OK <= response.status_code < HTTPStatus.MULTIPLE_CHOICES:
                # if request related to metrics upload
                if kwargs.get("metrics"):
                    self.metrics_upload_failed_queue = {}
                return response  # Success, no need to retry

            if i == 0:
                # Initial attempt, check status code and provide messages
                message = self._get_failure_message(response, retry, timeout)

                if verbose:
                    LOGGER.warning(f"{PREFIX}{message} {HELP_MSG} ({response.status_code})")

            if not self._should_retry(response.status_code):
                LOGGER.warning(f"{PREFIX}Request failed. {HELP_MSG} ({response.status_code}")
                break  # Not an error that should be retried, exit loop

            time.sleep(2**i)  # Exponential backoff for retries

        # if request related to metrics upload and exceed retries
        if response is None and kwargs.get("metrics"):
            self.metrics_upload_failed_queue.update(kwargs.get("metrics", None))

        return response

    if thread:
        # Start a new thread to run the retry_request function
        threading.Thread(target=retry_request, daemon=True).start()
    else:
        # If running in the main thread, call retry_request directly
        return retry_request()

upload_metrics

upload_metrics()

Upload model metrics to Ultralytics HUB.

Source code in ultralytics/hub/session.py
def upload_metrics(self):
    """Upload model metrics to Ultralytics HUB."""
    return self.request_queue(self.model.upload_metrics, metrics=self.metrics_queue.copy(), thread=True)

upload_model

upload_model(epoch: int, weights: str, is_best: bool = False, map: float = 0.0, final: bool = False) -> None

Upload a model checkpoint to Ultralytics HUB.

Parameters:

Name Type Description Default
epoch int

The current training epoch.

required
weights str

Path to the model weights file.

required
is_best bool

Indicates if the current model is the best one so far.

False
map float

Mean average precision of the model.

0.0
final bool

Indicates if the model is the final model after training.

False
Source code in ultralytics/hub/session.py
def upload_model(
    self,
    epoch: int,
    weights: str,
    is_best: bool = False,
    map: float = 0.0,
    final: bool = False,
) -> None:
    """
    Upload a model checkpoint to Ultralytics HUB.

    Args:
        epoch (int): The current training epoch.
        weights (str): Path to the model weights file.
        is_best (bool): Indicates if the current model is the best one so far.
        map (float): Mean average precision of the model.
        final (bool): Indicates if the model is the final model after training.
    """
    if Path(weights).is_file():
        progress_total = Path(weights).stat().st_size if final else None  # Only show progress if final
        self.request_queue(
            self.model.upload_model,
            epoch=epoch,
            weights=weights,
            is_best=is_best,
            map=map,
            final=final,
            retry=10,
            timeout=3600,
            thread=not final,
            progress_total=progress_total,
            stream_response=True,
        )
    else:
        LOGGER.warning(f"{PREFIX}WARNING ⚠️ Model upload issue. Missing model {weights}.")





Created 2023-11-12, Updated 2024-07-21
Authors: glenn-jocher (6), Burhan-Q (1)