wandb 0.18.0rc1__py3-none-macosx_10_13_x86_64.whl

wandb/sdk/wandb_setup.py

@@ -0,0 +1,400 @@
+#
+"""Setup wandb session.
+
+This module configures a wandb session which can extend to multiple wandb runs.
+
+Functions:
+    setup(): Configure wandb session.
+
+Early logging keeps track of logger output until the call to wandb.init() when the
+run_id can be resolved.
+
+"""
+
+import logging
+import os
+import sys
+import threading
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
+
+import wandb
+from wandb.sdk.lib import import_hooks
+
+from . import wandb_manager, wandb_settings
+from .lib import config_util, server, tracelog
+
+Settings = Union["wandb.sdk.wandb_settings.Settings", Dict[str, Any]]
+
+Logger = Union[logging.Logger, "_EarlyLogger"]
+
+if TYPE_CHECKING:
+    from . import wandb_run
+
+# logger will be configured to be either a standard logger instance or _EarlyLogger
+logger: Optional[Logger] = None
+
+
+def _set_logger(log_object: Logger) -> None:
+    """Configure module logger."""
+    global logger
+    logger = log_object
+
+
+class _EarlyLogger:
+    """Early logger which captures logs in memory until logging can be configured."""
+
+    def __init__(self) -> None:
+        self._log: List[tuple] = []
+        self._exception: List[tuple] = []
+        # support old warn() as alias of warning()
+        self.warn = self.warning
+
+    def debug(self, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._log.append((logging.DEBUG, msg, args, kwargs))
+
+    def info(self, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._log.append((logging.INFO, msg, args, kwargs))
+
+    def warning(self, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._log.append((logging.WARNING, msg, args, kwargs))
+
+    def error(self, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._log.append((logging.ERROR, msg, args, kwargs))
+
+    def critical(self, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._log.append((logging.CRITICAL, msg, args, kwargs))
+
+    def exception(self, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._exception.append((msg, args, kwargs))
+
+    def log(self, level: str, msg: str, *args: Any, **kwargs: Any) -> None:
+        self._log.append((level, msg, args, kwargs))
+
+    def _flush(self) -> None:
+        assert self is not logger
+        assert logger is not None
+        for level, msg, args, kwargs in self._log:
+            logger.log(level, msg, *args, **kwargs)
+        for msg, args, kwargs in self._exception:
+            logger.exception(msg, *args, **kwargs)
+
+
+class _WandbSetup__WandbSetup:  # noqa: N801
+    """Inner class of _WandbSetup."""
+
+    _manager: Optional[wandb_manager._Manager]
+    _pid: int
+
+    def __init__(
+        self,
+        pid: int,
+        settings: Optional[Settings] = None,
+        environ: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        self._environ = environ or dict(os.environ)
+        self._sweep_config: Optional[Dict[str, Any]] = None
+        self._config: Optional[Dict[str, Any]] = None
+        self._server: Optional[server.Server] = None
+        self._manager: Optional[wandb_manager._Manager] = None
+        self._pid = pid
+
+        # keep track of multiple runs, so we can unwind with join()s
+        self._global_run_stack: List[wandb_run.Run] = []
+
+        # TODO(jhr): defer strict checks until settings are fully initialized
+        #            and logging is ready
+        self._early_logger = _EarlyLogger()
+        _set_logger(self._early_logger)
+
+        self._settings = self._settings_setup(settings, self._early_logger)
+        # self._settings.freeze()
+
+        wandb.termsetup(self._settings, logger)
+
+        self._check()
+        self._setup()
+
+        tracelog_mode = self._settings._tracelog
+        if tracelog_mode:
+            tracelog.enable(tracelog_mode)
+
+    def _settings_setup(
+        self,
+        settings: Optional[Settings] = None,
+        early_logger: Optional[_EarlyLogger] = None,
+    ) -> "wandb_settings.Settings":
+        s = wandb_settings.Settings()
+        s._apply_base(pid=self._pid, _logger=early_logger)
+        s._apply_config_files(_logger=early_logger)
+        s._apply_env_vars(self._environ, _logger=early_logger)
+
+        if isinstance(settings, wandb_settings.Settings):
+            s._apply_settings(settings, _logger=early_logger)
+        elif isinstance(settings, dict):
+            # if passed settings arg is a mapping, update the settings with it
+            s._apply_setup(settings, _logger=early_logger)
+
+        s._infer_settings_from_environment()
+        if not s._cli_only_mode:
+            s._infer_run_settings_from_environment(_logger=early_logger)
+
+        return s
+
+    def _update(
+        self,
+        settings: Optional[Settings] = None,
+    ) -> None:
+        if settings is None:
+            return
+        # self._settings.unfreeze()
+        if isinstance(settings, wandb_settings.Settings):
+            # todo: check the logic here. this _only_ comes up in tests?
+            self._settings._apply_settings(settings)
+        elif isinstance(settings, dict):
+            # if it is a mapping, update the settings with it
+            self._settings.update(settings, source=wandb_settings.Source.SETUP)
+        # self._settings.freeze()
+
+    def _update_user_settings(self, settings: Optional[Settings] = None) -> None:
+        settings = settings or self._settings
+        # Get rid of cached results to force a refresh.
+        self._server = None
+        user_settings = self._load_user_settings(settings=settings)
+        if user_settings is not None:
+            # self._settings.unfreeze()
+            self._settings._apply_user(user_settings)
+            # self._settings.freeze()
+
+    def _early_logger_flush(self, new_logger: Logger) -> None:
+        if not self._early_logger:
+            return
+        _set_logger(new_logger)
+        # self._settings._clear_early_logger()
+        self._early_logger._flush()
+
+    def _get_logger(self) -> Optional[Logger]:
+        return logger
+
+    @property
+    def settings(self) -> "wandb_settings.Settings":
+        return self._settings
+
+    def _get_entity(self) -> Optional[str]:
+        if self._settings and self._settings._offline:
+            return None
+        if self._server is None:
+            self._load_viewer()
+        assert self._server is not None
+        entity = self._server._viewer.get("entity")
+        return entity
+
+    def _get_username(self) -> Optional[str]:
+        if self._settings and self._settings._offline:
+            return None
+        if self._server is None:
+            self._load_viewer()
+        assert self._server is not None
+        username = self._server._viewer.get("username")
+        return username
+
+    def _get_teams(self) -> List[str]:
+        if self._settings and self._settings._offline:
+            return []
+        if self._server is None:
+            self._load_viewer()
+        assert self._server is not None
+        teams = self._server._viewer.get("teams")
+        if teams:
+            teams = [team["node"]["name"] for team in teams["edges"]]
+        return teams or []
+
+    def _load_viewer(self, settings: Optional[Settings] = None) -> None:
+        if self._settings and self._settings._offline:
+            return
+        if isinstance(settings, dict):
+            settings = wandb_settings.Settings(**settings)
+        s = server.Server(settings=settings)
+        s.query_with_timeout()
+        self._server = s
+
+    def _load_user_settings(
+        self, settings: Optional[Settings] = None
+    ) -> Optional[Dict[str, Any]]:
+        if self._server is None:
+            self._load_viewer(settings=settings)
+
+        # offline?
+        if self._server is None:
+            return None
+
+        flags = self._server._flags
+        user_settings = dict()
+        if "code_saving_enabled" in flags:
+            user_settings["save_code"] = flags["code_saving_enabled"]
+
+        email = self._server._viewer.get("email", None)
+        if email:
+            user_settings["email"] = email
+
+        return user_settings
+
+    def _check(self) -> None:
+        if hasattr(threading, "main_thread"):
+            if threading.current_thread() is not threading.main_thread():
+                pass
+        elif threading.current_thread().name != "MainThread":
+            print("bad thread2", threading.current_thread().name)
+        if getattr(sys, "frozen", False):
+            print("frozen, could be trouble")
+
+    def _setup(self) -> None:
+        self._setup_manager()
+
+        sweep_path = self._settings.sweep_param_path
+        if sweep_path:
+            self._sweep_config = config_util.dict_from_config_file(
+                sweep_path, must_exist=True
+            )
+
+        # if config_paths was set, read in config dict
+        if self._settings.config_paths:
+            # TODO(jhr): handle load errors, handle list of files
+            for config_path in self._settings.config_paths:
+                config_dict = config_util.dict_from_config_file(config_path)
+                if config_dict is None:
+                    continue
+                if self._config is not None:
+                    self._config.update(config_dict)
+                else:
+                    self._config = config_dict
+
+    def _teardown(self, exit_code: Optional[int] = None) -> None:
+        import_hooks.unregister_all_post_import_hooks()
+
+        if not self._manager:
+            return
+
+        internal_exit_code = self._manager._teardown(exit_code or 0)
+        self._manager = None
+
+        if internal_exit_code != 0:
+            sys.exit(internal_exit_code)
+
+    def _setup_manager(self) -> None:
+        if self._settings._disable_service:
+            return
+        self._manager = wandb_manager._Manager(settings=self._settings)
+
+    def _get_manager(self) -> Optional[wandb_manager._Manager]:
+        return self._manager
+
+
+class _WandbSetup:
+    """Wandb singleton class.
+
+    Note: This is a process local singleton.
+    (Forked processes will get a new copy of the object)
+    """
+
+    _instance: Optional["_WandbSetup__WandbSetup"] = None
+
+    def __init__(self, settings: Optional[Settings] = None) -> None:
+        pid = os.getpid()
+        if _WandbSetup._instance and _WandbSetup._instance._pid == pid:
+            _WandbSetup._instance._update(settings=settings)
+            return
+        _WandbSetup._instance = _WandbSetup__WandbSetup(settings=settings, pid=pid)
+
+    def __getattr__(self, name: str) -> Any:
+        return getattr(self._instance, name)
+
+
+def _setup(
+    settings: Optional[Settings] = None,
+    _reset: bool = False,
+) -> Optional["_WandbSetup"]:
+    """Set up library context."""
+    if _reset:
+        teardown()
+        return None
+
+    wl = _WandbSetup(settings=settings)
+    return wl
+
+
+def setup(settings: Optional[Settings] = None) -> Optional["_WandbSetup"]:
+    """Prepares W&B for use in the current process and its children.
+
+    You can usually ignore this as it is implicitly called by `wandb.init()`.
+
+    When using wandb in multiple processes, calling `wandb.setup()`
+    in the parent process before starting child processes may improve
+    performance and resource utilization.
+
+    Note that `wandb.setup()` modifies `os.environ`, and it is important
+    that child processes inherit the modified environment variables.
+
+    See also `wandb.teardown()`.
+
+    Args:
+        settings (Optional[Union[Dict[str, Any], wandb.Settings]]): Configuration settings
+            to apply globally. These can be overridden by subsequent `wandb.init()` calls.
+
+    Example:
+        ```python
+        import multiprocessing
+
+        import wandb
+
+
+        def run_experiment(params):
+            with wandb.init(config=params):
+                # Run experiment
+                pass
+
+
+        if __name__ == "__main__":
+            # Start backend and set global config
+            wandb.setup(settings={"project": "my_project"})
+
+            # Define experiment parameters
+            experiment_params = [
+                {"learning_rate": 0.01, "epochs": 10},
+                {"learning_rate": 0.001, "epochs": 20},
+            ]
+
+            # Start multiple processes, each running a separate experiment
+            processes = []
+            for params in experiment_params:
+                p = multiprocessing.Process(target=run_experiment, args=(params,))
+                p.start()
+                processes.append(p)
+
+            # Wait for all processes to complete
+            for p in processes:
+                p.join()
+
+            # Optional: Explicitly shut down the backend
+            wandb.teardown()
+        ```
+    """
+    ret = _setup(settings=settings)
+    return ret
+
+
+def teardown(exit_code: Optional[int] = None) -> None:
+    """Waits for wandb to finish and frees resources.
+
+    Completes any runs that were not explicitly finished
+    using `run.finish()` and waits for all data to be uploaded.
+
+    It is recommended to call this at the end of a session
+    that used `wandb.setup()`. It is invoked automatically
+    in an `atexit` hook, but this is not reliable in certain setups
+    such as when using Python's `multiprocessing` module.
+    """
+    setup_instance = _WandbSetup._instance
+    _WandbSetup._instance = None
+
+    if setup_instance:
+        setup_instance._teardown(exit_code=exit_code)

wandb/sdk/wandb_summary.py

@@ -0,0 +1,150 @@
+import abc
+import typing as t
+
+from .interface.summary_record import SummaryItem, SummaryRecord
+
+
+def _get_dict(d):
+    if isinstance(d, dict):
+        return d
+    # assume argparse Namespace
+    return vars(d)
+
+
+class SummaryDict(metaclass=abc.ABCMeta):
+    """dict-like wrapper for the nested dictionaries in a SummarySubDict.
+
+    Triggers self._root._callback on property changes.
+    """
+
+    @abc.abstractmethod
+    def _as_dict(self):
+        raise NotImplementedError
+
+    @abc.abstractmethod
+    def _update(self, record: SummaryRecord):
+        raise NotImplementedError
+
+    def keys(self):
+        return [k for k in self._as_dict().keys() if k != "_wandb"]
+
+    def get(self, key, default=None):
+        return self._as_dict().get(key, default)
+
+    def __getitem__(self, key):
+        item = self._as_dict()[key]
+
+        if isinstance(item, dict):
+            # this nested dict needs to be wrapped:
+            wrapped_item = SummarySubDict()
+            object.__setattr__(wrapped_item, "_items", item)
+            object.__setattr__(wrapped_item, "_parent", self)
+            object.__setattr__(wrapped_item, "_parent_key", key)
+
+            return wrapped_item
+
+        # this item isn't a nested dict
+        return item
+
+    __getattr__ = __getitem__
+
+    def __setitem__(self, key, val):
+        self.update({key: val})
+
+    __setattr__ = __setitem__
+
+    def __delattr__(self, key):
+        record = SummaryRecord()
+        item = SummaryItem()
+        item.key = (key,)
+        record.remove = (item,)
+        self._update(record)
+
+    __delitem__ = __delattr__
+
+    def update(self, d: t.Dict):
+        # import ipdb; ipdb.set_trace()
+        record = SummaryRecord()
+        for key, value in d.items():
+            item = SummaryItem()
+            item.key = (key,)
+            item.value = value
+            record.update.append(item)
+
+        self._update(record)
+
+
+class Summary(SummaryDict):
+    """Track single values for each metric for each run.
+
+    By default, a metric's summary is the last value of its History.
+
+    For example, `wandb.log({'accuracy': 0.9})` will add a new step to History and
+    update Summary to the latest value. In some cases, it's more useful to have
+    the maximum or minimum of a metric instead of the final value. You can set
+    history manually `(wandb.summary['accuracy'] = best_acc)`.
+
+    In the UI, summary metrics appear in the table to compare across runs.
+    Summary metrics are also used in visualizations like the scatter plot and
+    parallel coordinates chart.
+
+    After training has completed, you may want to save evaluation metrics to a
+    run. Summary can handle numpy arrays and PyTorch/TensorFlow tensors. When
+    you save one of these types to Summary, we persist the entire tensor in a
+    binary file and store high level metrics in the summary object, such as min,
+    mean, variance, and 95th percentile.
+
+    Examples:
+        ```python
+        wandb.init(config=args)
+
+        best_accuracy = 0
+        for epoch in range(1, args.epochs + 1):
+            test_loss, test_accuracy = test()
+            if test_accuracy > best_accuracy:
+                wandb.run.summary["best_accuracy"] = test_accuracy
+                best_accuracy = test_accuracy
+        ```
+    """
+
+    _update_callback: t.Callable
+    _get_current_summary_callback: t.Callable
+
+    def __init__(self, get_current_summary_callback: t.Callable):
+        super().__init__()
+        object.__setattr__(self, "_update_callback", None)
+        object.__setattr__(
+            self, "_get_current_summary_callback", get_current_summary_callback
+        )
+
+    def _set_update_callback(self, update_callback: t.Callable):
+        object.__setattr__(self, "_update_callback", update_callback)
+
+    def _as_dict(self):
+        return self._get_current_summary_callback()
+
+    def _update(self, record: SummaryRecord):
+        if self._update_callback:  # type: ignore
+            self._update_callback(record)
+
+
+class SummarySubDict(SummaryDict):
+    """Non-root node of the summary data structure.
+
+    Contains a path to itself from the root.
+    """
+
+    _items: t.Dict
+    _parent: SummaryDict
+    _parent_key: str
+
+    def __init__(self):
+        object.__setattr__(self, "_items", dict())
+        object.__setattr__(self, "_parent", None)
+        object.__setattr__(self, "_parent_key", None)
+
+    def _as_dict(self):
+        return self._items
+
+    def _update(self, record: SummaryRecord):
+        return self._parent._update(record._add_next_parent(self._parent_key))

wandb/sdk/wandb_sweep.py

@@ -0,0 +1,119 @@
+import urllib.parse
+from typing import TYPE_CHECKING, Callable, Dict, List, Optional, Union
+
+import wandb
+from wandb import env
+from wandb.apis import InternalApi
+from wandb.sdk.launch.sweeps.utils import handle_sweep_config_violations
+
+from . import wandb_login
+
+if TYPE_CHECKING:
+    from wandb.wandb_controller import _WandbController
+
+
+def _get_sweep_url(api, sweep_id):
+    """Return sweep url if we can figure it out."""
+    if api.api_key:
+        if api.settings("entity") is None:
+            viewer = api.viewer()
+            if viewer.get("entity"):
+                api.set_setting("entity", viewer["entity"])
+        project = api.settings("project")
+        if not project:
+            return
+        if api.settings("entity"):
+            return "{base}/{entity}/{project}/sweeps/{sweepid}".format(
+                base=api.app_url,
+                entity=urllib.parse.quote(api.settings("entity")),
+                project=urllib.parse.quote(project),
+                sweepid=urllib.parse.quote(sweep_id),
+            )
+
+
+def sweep(
+    sweep: Union[dict, Callable],
+    entity: Optional[str] = None,
+    project: Optional[str] = None,
+    prior_runs: Optional[List[str]] = None,
+) -> str:
+    """Initialize a hyperparameter sweep.
+
+    Search for hyperparameters that optimizes a cost function
+    of a machine learning model by testing various combinations.
+
+    Make note the unique identifier, `sweep_id`, that is returned.
+    At a later step provide the `sweep_id` to a sweep agent.
+
+    Args:
+      sweep: The configuration of a hyperparameter search.
+        (or configuration generator). See
+        [Sweep configuration structure](https://docs.wandb.ai/guides/sweeps/define-sweep-configuration)
+        for information on how to define your sweep.
+        If you provide a callable, ensure that the callable does
+        not take arguments and that it returns a dictionary that
+        conforms to the W&B sweep config spec.
+      entity: The username or team name where you want to send W&B
+        runs created by the sweep to. Ensure that the entity you
+        specify already exists. If you don't specify an entity,
+        the run will be sent to your default entity,
+        which is usually your username.
+      project: The name of the project where W&B runs created from
+        the sweep are sent to. If the project is not specified, the
+        run is sent to a project labeled 'Uncategorized'.
+      prior_runs: The run IDs of existing runs to add to this sweep.
+
+    Returns:
+      sweep_id: str. A unique identifier for the sweep.
+    """
+    if callable(sweep):
+        sweep = sweep()
+    """Sweep create for controller api and jupyter (eventually for cli)."""
+
+    # Project may be only found in the sweep config.
+    if project is None and isinstance(sweep, dict):
+        project = sweep.get("project", None)
+
+    if entity:
+        env.set_entity(entity)
+    if project:
+        env.set_project(project)
+
+    # Make sure we are logged in
+    if wandb.run is None:
+        wandb_login._login(_silent=True)
+    api = InternalApi()
+    sweep_id, warnings = api.upsert_sweep(sweep, prior_runs=prior_runs)
+    handle_sweep_config_violations(warnings)
+    print("Create sweep with ID:", sweep_id)
+    sweep_url = _get_sweep_url(api, sweep_id)
+    if sweep_url:
+        print("Sweep URL:", sweep_url)
+    return sweep_id
+
+
+def controller(
+    sweep_id_or_config: Optional[Union[str, Dict]] = None,
+    entity: Optional[str] = None,
+    project: Optional[str] = None,
+) -> "_WandbController":
+    """Public sweep controller constructor.
+
+    Usage:
+        ```python
+        import wandb
+
+        tuner = wandb.controller(...)
+        print(tuner.sweep_config)
+        print(tuner.sweep_id)
+        tuner.configure_search(...)
+        tuner.configure_stopping(...)
+        ```
+
+    """
+    from ..wandb_controller import _WandbController
+
+    c = _WandbController(
+        sweep_id_or_config=sweep_id_or_config, entity=entity, project=project
+    )
+    return c