venv/lib/python3.11/site-packages/wandb/util.py

import colorsys
import contextlib
import dataclasses
import enum
import functools
import gzip
import importlib
import importlib.util
import itertools
import json
import logging
import math
import numbers
import os
import pathlib
import platform
import queue
import random
import re
import secrets
import shlex
import socket
import string
import sys
import tarfile
import tempfile
import threading
import time
import types
import urllib
from dataclasses import asdict, is_dataclass
from datetime import date, datetime, timedelta
from importlib import import_module
from sys import getsizeof
from types import ModuleType
from typing import (
    IO,
    TYPE_CHECKING,
    Callable,
    Dict,
    Iterable,
    List,
    Mapping,
    Optional,
    Sequence,
    TextIO,
    Tuple,
    Union,
)

import requests
import yaml
from typing_extensions import Any, Generator, TypeGuard, TypeVar

import wandb
import wandb.env
from wandb.errors import (
    AuthenticationError,
    CommError,
    UsageError,
    WandbCoreNotAvailableError,
    term,
)
from wandb.sdk.internal.thread_local_settings import _thread_local_api_settings
from wandb.sdk.lib import filesystem, runid
from wandb.sdk.lib.json_util import dump, dumps
from wandb.sdk.lib.paths import FilePathStr, StrPath

if TYPE_CHECKING:
    import wandb.sdk.internal.settings_static
    import wandb.sdk.wandb_settings
    from wandb.sdk.artifacts.artifact import Artifact

CheckRetryFnType = Callable[[Exception], Union[bool, timedelta]]
T = TypeVar("T")


logger = logging.getLogger(__name__)
_not_importable = set()

LAUNCH_JOB_ARTIFACT_SLOT_NAME = "_wandb_job"

MAX_LINE_BYTES = (10 << 20) - (100 << 10)  # imposed by back end
IS_GIT = os.path.exists(os.path.join(os.path.dirname(__file__), "..", ".git"))

# From https://docs.docker.com/engine/reference/commandline/tag/
# "Name components may contain lowercase letters, digits and separators.
# A separator is defined as a period, one or two underscores, or one or more dashes.
# A name component may not start or end with a separator."
DOCKER_IMAGE_NAME_SEPARATOR = "(?:__|[._]|[-]+)"
RE_DOCKER_IMAGE_NAME_SEPARATOR_START = re.compile("^" + DOCKER_IMAGE_NAME_SEPARATOR)
RE_DOCKER_IMAGE_NAME_SEPARATOR_END = re.compile(DOCKER_IMAGE_NAME_SEPARATOR + "$")
RE_DOCKER_IMAGE_NAME_SEPARATOR_REPEAT = re.compile(DOCKER_IMAGE_NAME_SEPARATOR + "{2,}")
RE_DOCKER_IMAGE_NAME_CHARS = re.compile(r"[^a-z0-9._\-]")

# these match the environments for gorilla
if IS_GIT:
    SENTRY_ENV = "development"
else:
    SENTRY_ENV = "production"


POW_10_BYTES = [
    ("B", 10**0),
    ("KB", 10**3),
    ("MB", 10**6),
    ("GB", 10**9),
    ("TB", 10**12),
    ("PB", 10**15),
    ("EB", 10**18),
]

POW_2_BYTES = [
    ("B", 2**0),
    ("KiB", 2**10),
    ("MiB", 2**20),
    ("GiB", 2**30),
    ("TiB", 2**40),
    ("PiB", 2**50),
    ("EiB", 2**60),
]


def vendor_setup() -> Callable:
    """Create a function that restores user paths after vendor imports.

    This enables us to use the vendor directory for packages we don't depend on. Call
    the returned function after imports are complete. If you don't you may modify the
    user's path which is never good.

    Usage:

    ```python
    reset_path = vendor_setup()
    # do any vendor imports...
    reset_path()
    ```
    """
    original_path = [directory for directory in sys.path]

    def reset_import_path() -> None:
        sys.path = original_path

    parent_dir = os.path.abspath(os.path.dirname(__file__))
    vendor_dir = os.path.join(parent_dir, "vendor")
    vendor_packages = (
        "gql-0.2.0",
        "graphql-core-1.1",
        "watchdog_0_9_0",
        "promise-2.3.0",
    )
    package_dirs = [os.path.join(vendor_dir, p) for p in vendor_packages]
    for p in [vendor_dir] + package_dirs:
        if p not in sys.path:
            sys.path.insert(1, p)

    return reset_import_path


def vendor_import(name: str) -> Any:
    reset_path = vendor_setup()
    module = import_module(name)
    reset_path()
    return module


class LazyModuleState:
    def __init__(self, module: types.ModuleType) -> None:
        self.module = module
        self.load_started = False
        self.lock = threading.RLock()

    def load(self) -> None:
        with self.lock:
            if self.load_started:
                return
            self.load_started = True
            assert self.module.__spec__ is not None
            assert self.module.__spec__.loader is not None
            self.module.__spec__.loader.exec_module(self.module)
            self.module.__class__ = types.ModuleType

            # Set the submodule as an attribute on the parent module
            # This enables access to the submodule via normal attribute access.
            parent, _, child = self.module.__name__.rpartition(".")
            if parent:
                parent_module = sys.modules[parent]
                setattr(parent_module, child, self.module)


class LazyModule(types.ModuleType):
    def __getattribute__(self, name: str) -> Any:
        state = object.__getattribute__(self, "__lazy_module_state__")
        state.load()
        return object.__getattribute__(self, name)

    def __setattr__(self, name: str, value: Any) -> None:
        state = object.__getattribute__(self, "__lazy_module_state__")
        state.load()
        object.__setattr__(self, name, value)

    def __delattr__(self, name: str) -> None:
        state = object.__getattribute__(self, "__lazy_module_state__")
        state.load()
        object.__delattr__(self, name)


def import_module_lazy(name: str) -> types.ModuleType:
    """Import a module lazily, only when it is used.

    Inspired by importlib.util.LazyLoader, but improved so that the module loading is
    thread-safe. Circular dependency between modules can lead to a deadlock if the two
    modules are loaded from different threads.

    :param (str) name: Dot-separated module path. E.g., 'scipy.stats'.
    """
    try:
        return sys.modules[name]
    except KeyError:
        spec = importlib.util.find_spec(name)
        if spec is None:
            raise ModuleNotFoundError
        module = importlib.util.module_from_spec(spec)
        module.__lazy_module_state__ = LazyModuleState(module)  # type: ignore
        module.__class__ = LazyModule
        sys.modules[name] = module
        return module


def get_module(
    name: str,
    required: Optional[Union[str, bool]] = None,
    lazy: bool = True,
) -> Any:
    """Return module or None. Absolute import is required.

    :param (str) name: Dot-separated module path. E.g., 'scipy.stats'.
    :param (str) required: A string to raise a ValueError if missing
    :param (bool) lazy: If True, return a lazy loader for the module.
    :return: (module|None) If import succeeds, the module will be returned.
    """
    if name not in _not_importable:
        try:
            if not lazy:
                return import_module(name)
            else:
                return import_module_lazy(name)
        except Exception:
            _not_importable.add(name)
            msg = f"Error importing optional module {name}"
            if required:
                logger.exception(msg)
    if required and name in _not_importable:
        raise wandb.Error(required)


def get_optional_module(name) -> Optional["importlib.ModuleInterface"]:  # type: ignore
    return get_module(name)


np = get_module("numpy")

pd_available = False
pandas_spec = importlib.util.find_spec("pandas")
if pandas_spec is not None:
    pd_available = True

# TODO: Revisit these limits
VALUE_BYTES_LIMIT = 100000


def app_url(api_url: str) -> str:
    """Return the frontend app url without a trailing slash."""
    # TODO: move me to settings
    app_url = wandb.env.get_app_url()
    if app_url is not None:
        return str(app_url.strip("/"))
    if "://api.wandb.test" in api_url:
        # dev mode
        return api_url.replace("://api.", "://app.").strip("/")
    elif "://api.wandb." in api_url:
        # cloud
        return api_url.replace("://api.", "://").strip("/")
    elif "://api." in api_url:
        # onprem cloud
        return api_url.replace("://api.", "://app.").strip("/")
    # wandb/local
    return api_url


def get_full_typename(o: Any) -> Any:
    """Determine types based on type names.

    Avoids needing to to import (and therefore depend on) PyTorch, TensorFlow, etc.
    """
    instance_name = o.__class__.__module__ + "." + o.__class__.__name__
    if instance_name in ["builtins.module", "__builtin__.module"]:
        return o.__name__
    else:
        return instance_name


def get_h5_typename(o: Any) -> Any:
    typename = get_full_typename(o)
    if is_tf_tensor_typename(typename):
        return "tensorflow.Tensor"
    elif is_pytorch_tensor_typename(typename):
        return "torch.Tensor"
    else:
        return o.__class__.__module__.split(".")[0] + "." + o.__class__.__name__


def is_uri(string: str) -> bool:
    parsed_uri = urllib.parse.urlparse(string)
    return len(parsed_uri.scheme) > 0


def local_file_uri_to_path(uri: str) -> str:
    """Convert URI to local filesystem path.

    No-op if the uri does not have the expected scheme.
    """
    path = urllib.parse.urlparse(uri).path if uri.startswith("file:") else uri
    return urllib.request.url2pathname(path)


def get_local_path_or_none(path_or_uri: str) -> Optional[str]:
    """Return path if local, None otherwise.

    Return None if the argument is a local path (not a scheme or file:///). Otherwise
    return `path_or_uri`.
    """
    parsed_uri = urllib.parse.urlparse(path_or_uri)
    if (
        len(parsed_uri.scheme) == 0
        or parsed_uri.scheme == "file"
        and len(parsed_uri.netloc) == 0
    ):
        return local_file_uri_to_path(path_or_uri)
    else:
        return None


def make_tarfile(
    output_filename: str,
    source_dir: str,
    archive_name: str,
    custom_filter: Optional[Callable] = None,
) -> None:
    # Helper for filtering out modification timestamps
    def _filter_timestamps(tar_info: "tarfile.TarInfo") -> Optional["tarfile.TarInfo"]:
        tar_info.mtime = 0
        return tar_info if custom_filter is None else custom_filter(tar_info)

    descriptor, unzipped_filename = tempfile.mkstemp()
    try:
        with tarfile.open(unzipped_filename, "w") as tar:
            tar.add(source_dir, arcname=archive_name, filter=_filter_timestamps)
        # When gzipping the tar, don't include the tar's filename or modification time in the
        # zipped archive (see https://docs.python.org/3/library/gzip.html#gzip.GzipFile)
        with gzip.GzipFile(
            filename="", fileobj=open(output_filename, "wb"), mode="wb", mtime=0
        ) as gzipped_tar, open(unzipped_filename, "rb") as tar_file:
            gzipped_tar.write(tar_file.read())
    finally:
        os.close(descriptor)
        os.remove(unzipped_filename)


def is_tf_tensor(obj: Any) -> bool:
    import tensorflow  # type: ignore

    return isinstance(obj, tensorflow.Tensor)


def is_tf_tensor_typename(typename: str) -> bool:
    return typename.startswith("tensorflow.") and (
        "Tensor" in typename or "Variable" in typename
    )


def is_tf_eager_tensor_typename(typename: str) -> bool:
    return typename.startswith("tensorflow.") and ("EagerTensor" in typename)


def is_pytorch_tensor(obj: Any) -> bool:
    import torch  # type: ignore

    return isinstance(obj, torch.Tensor)


def is_pytorch_tensor_typename(typename: str) -> bool:
    return typename.startswith("torch.") and (
        "Tensor" in typename or "Variable" in typename
    )


def is_jax_tensor_typename(typename: str) -> bool:
    return typename.startswith("jaxlib.") and "Array" in typename


def get_jax_tensor(obj: Any) -> Optional[Any]:
    import jax  # type: ignore

    return jax.device_get(obj)


def is_fastai_tensor_typename(typename: str) -> bool:
    return typename.startswith("fastai.") and ("Tensor" in typename)


def is_pandas_data_frame_typename(typename: str) -> bool:
    return typename.startswith("pandas.") and "DataFrame" in typename


def is_matplotlib_typename(typename: str) -> bool:
    return typename.startswith("matplotlib.")


def is_plotly_typename(typename: str) -> bool:
    return typename.startswith("plotly.")


def is_plotly_figure_typename(typename: str) -> bool:
    return typename.startswith("plotly.") and typename.endswith(".Figure")


def is_numpy_array(obj: Any) -> bool:
    return np and isinstance(obj, np.ndarray)


def is_pandas_data_frame(obj: Any) -> bool:
    if pd_available:
        import pandas as pd

        return isinstance(obj, pd.DataFrame)
    else:
        return is_pandas_data_frame_typename(get_full_typename(obj))


def ensure_matplotlib_figure(obj: Any) -> Any:
    """Extract the current figure from a matplotlib object.

    Return the object itself if it's a figure.
    Raises ValueError if the object can't be converted.
    """
    import matplotlib  # type: ignore
    from matplotlib.figure import Figure  # type: ignore

    # there are combinations of plotly and matplotlib versions that don't work well together,
    # this patches matplotlib to add a removed method that plotly assumes exists
    from matplotlib.spines import Spine  # type: ignore

    def is_frame_like(self: Any) -> bool:
        """Return True if directly on axes frame.

        This is useful for determining if a spine is the edge of an
        old style MPL plot. If so, this function will return True.
        """
        position = self._position or ("outward", 0.0)
        if isinstance(position, str):
            if position == "center":
                position = ("axes", 0.5)
            elif position == "zero":
                position = ("data", 0)
        if len(position) != 2:
            raise ValueError("position should be 2-tuple")
        position_type, amount = position  # type: ignore
        if position_type == "outward" and amount == 0:
            return True
        else:
            return False

    Spine.is_frame_like = is_frame_like

    if obj == matplotlib.pyplot:
        obj = obj.gcf()
    elif not isinstance(obj, Figure):
        if hasattr(obj, "figure"):
            obj = obj.figure
            # Some matplotlib objects have a figure function
            if not isinstance(obj, Figure):
                raise ValueError(
                    "Only matplotlib.pyplot or matplotlib.pyplot.Figure objects are accepted."
                )
    return obj


def matplotlib_to_plotly(obj: Any) -> Any:
    obj = ensure_matplotlib_figure(obj)
    tools = get_module(
        "plotly.tools",
        required=(
            "plotly is required to log interactive plots, install with: "
            "`pip install plotly` or convert the plot to an image with `wandb.Image(plt)`"
        ),
    )
    return tools.mpl_to_plotly(obj)


def matplotlib_contains_images(obj: Any) -> bool:
    obj = ensure_matplotlib_figure(obj)
    return any(len(ax.images) > 0 for ax in obj.axes)


def _numpy_generic_convert(obj: Any) -> Any:
    obj = obj.item()
    if isinstance(obj, float) and math.isnan(obj):
        obj = None
    elif isinstance(obj, np.generic) and (
        obj.dtype.kind == "f" or obj.dtype == "bfloat16"
    ):
        # obj is a numpy float with precision greater than that of native python float
        # (i.e., float96 or float128) or it is of custom type such as bfloat16.
        # in these cases, obj.item() does not return a native
        # python float (in the first case - to avoid loss of precision,
        # so we need to explicitly cast this down to a 64bit float)
        obj = float(obj)
    return obj


def _sanitize_numpy_keys(
    d: Dict,
    visited: Optional[Dict[int, Dict]] = None,
) -> Tuple[Dict, bool]:
    """Returns a dictionary where all NumPy keys are converted.

    Args:
        d: The dictionary to sanitize.

    Returns:
        A sanitized dictionary, and a boolean indicating whether anything was
        changed.
    """
    out: Dict[Any, Any] = dict()
    converted = False

    # Work with recursive dictionaries: if a dictionary has already been
    # converted, reuse its converted value to retain the recursive structure
    # of the input.
    if visited is None:
        visited = {id(d): out}
    elif id(d) in visited:
        return visited[id(d)], False
    visited[id(d)] = out

    for key, value in d.items():
        if isinstance(value, dict):
            value, converted_value = _sanitize_numpy_keys(value, visited)
            converted |= converted_value
        if isinstance(key, np.generic):
            key = _numpy_generic_convert(key)
            converted = True
        out[key] = value

    return out, converted


def json_friendly(  # noqa: C901
    obj: Any,
) -> Union[Tuple[Any, bool], Tuple[Union[None, str, float], bool]]:
    """Convert an object into something that's more becoming of JSON."""
    converted = True
    typename = get_full_typename(obj)

    if is_tf_eager_tensor_typename(typename):
        obj = obj.numpy()
    elif is_tf_tensor_typename(typename):
        try:
            obj = obj.eval()
        except RuntimeError:
            obj = obj.numpy()
    elif is_pytorch_tensor_typename(typename) or is_fastai_tensor_typename(typename):
        try:
            if obj.requires_grad:
                obj = obj.detach()
        except AttributeError:
            pass  # before 0.4 is only present on variables

        try:
            obj = obj.data
        except RuntimeError:
            pass  # happens for Tensors before 0.4

        if obj.size():
            obj = obj.cpu().detach().numpy()
        else:
            return obj.item(), True
    elif is_jax_tensor_typename(typename):
        obj = get_jax_tensor(obj)

    if is_numpy_array(obj):
        if obj.size == 1:
            obj = obj.flatten()[0]
        elif obj.size <= 32:
            obj = obj.tolist()
    elif np and isinstance(obj, np.generic):
        obj = _numpy_generic_convert(obj)
    elif isinstance(obj, bytes):
        obj = obj.decode("utf-8")
    elif isinstance(obj, (datetime, date)):
        obj = obj.isoformat()
    elif callable(obj):
        obj = (
            f"{obj.__module__}.{obj.__qualname__}"
            if hasattr(obj, "__qualname__") and hasattr(obj, "__module__")
            else str(obj)
        )
    elif isinstance(obj, float) and math.isnan(obj):
        obj = None
    elif isinstance(obj, dict) and np:
        obj, converted = _sanitize_numpy_keys(obj)
    elif isinstance(obj, set):
        # set is not json serializable, so we convert it to tuple
        obj = tuple(obj)
    elif isinstance(obj, enum.Enum):
        obj = obj.name
    else:
        converted = False
    if getsizeof(obj) > VALUE_BYTES_LIMIT:
        wandb.termwarn(
            f"Serializing object of type {type(obj).__name__} that is {getsizeof(obj)} bytes"
        )
    return obj, converted


def json_friendly_val(val: Any) -> Any:
    """Make any value (including dict, slice, sequence, dataclass) JSON friendly."""
    converted: Union[dict, list]
    if isinstance(val, dict):
        converted = {}
        for key, value in val.items():
            converted[key] = json_friendly_val(value)
        return converted
    if isinstance(val, slice):
        converted = dict(
            slice_start=val.start, slice_step=val.step, slice_stop=val.stop
        )
        return converted
    val, _ = json_friendly(val)
    if isinstance(val, Sequence) and not isinstance(val, str):
        converted = []
        for value in val:
            converted.append(json_friendly_val(value))
        return converted
    if is_dataclass(val) and not isinstance(val, type):
        converted = asdict(val)
        return converted
    else:
        if val.__class__.__module__ not in ("builtins", "__builtin__"):
            val = str(val)
        return val


def alias_is_version_index(alias: str) -> bool:
    return len(alias) >= 2 and alias[0] == "v" and alias[1:].isnumeric()


def convert_plots(obj: Any) -> Any:
    if is_matplotlib_typename(get_full_typename(obj)):
        tools = get_module(
            "plotly.tools",
            required=(
                "plotly is required to log interactive plots, install with: "
                "`pip install plotly` or convert the plot to an image with `wandb.Image(plt)`"
            ),
        )
        obj = tools.mpl_to_plotly(obj)

    if is_plotly_typename(get_full_typename(obj)):
        return {"_type": "plotly", "plot": obj.to_plotly_json()}
    else:
        return obj


def maybe_compress_history(obj: Any) -> Tuple[Any, bool]:
    if np and isinstance(obj, np.ndarray) and obj.size > 32:
        return wandb.Histogram(obj, num_bins=32).to_json(), True
    else:
        return obj, False


def maybe_compress_summary(obj: Any, h5_typename: str) -> Tuple[Any, bool]:
    if np and isinstance(obj, np.ndarray) and obj.size > 32:
        return (
            {
                "_type": h5_typename,  # may not be ndarray
                "var": np.var(obj).item(),
                "mean": np.mean(obj).item(),
                "min": np.amin(obj).item(),
                "max": np.amax(obj).item(),
                "10%": np.percentile(obj, 10),
                "25%": np.percentile(obj, 25),
                "75%": np.percentile(obj, 75),
                "90%": np.percentile(obj, 90),
                "size": obj.size,
            },
            True,
        )
    else:
        return obj, False


def launch_browser(attempt_launch_browser: bool = True) -> bool:
    """Decide if we should launch a browser."""
    _display_variables = ["DISPLAY", "WAYLAND_DISPLAY", "MIR_SOCKET"]
    _webbrowser_names_blocklist = ["www-browser", "lynx", "links", "elinks", "w3m"]

    import webbrowser

    launch_browser = attempt_launch_browser
    if launch_browser:
        if "linux" in sys.platform and not any(
            os.getenv(var) for var in _display_variables
        ):
            launch_browser = False
        try:
            browser = webbrowser.get()
            if hasattr(browser, "name") and browser.name in _webbrowser_names_blocklist:
                launch_browser = False
        except webbrowser.Error:
            launch_browser = False

    return launch_browser


def generate_id(length: int = 8) -> str:
    # Do not use this; use wandb.sdk.lib.runid.generate_id instead.
    # This is kept only for legacy code.
    return runid.generate_id(length)


def parse_tfjob_config() -> Any:
    """Attempt to parse TFJob config, returning False if it can't find it."""
    if os.getenv("TF_CONFIG"):
        try:
            return json.loads(os.environ["TF_CONFIG"])
        except ValueError:
            return False
    else:
        return False


class WandBJSONEncoder(json.JSONEncoder):
    """A JSON Encoder that handles some extra types."""

    def default(self, obj: Any) -> Any:
        if hasattr(obj, "json_encode"):
            return obj.json_encode()
        # if hasattr(obj, 'to_json'):
        #     return obj.to_json()
        tmp_obj, converted = json_friendly(obj)
        if converted:
            return tmp_obj
        return json.JSONEncoder.default(self, obj)


class WandBJSONEncoderOld(json.JSONEncoder):
    """A JSON Encoder that handles some extra types."""

    def default(self, obj: Any) -> Any:
        tmp_obj, converted = json_friendly(obj)
        tmp_obj, compressed = maybe_compress_summary(tmp_obj, get_h5_typename(obj))
        if converted:
            return tmp_obj
        return json.JSONEncoder.default(self, tmp_obj)


class WandBHistoryJSONEncoder(json.JSONEncoder):
    """A JSON Encoder that handles some extra types.

    This encoder turns numpy like objects with a size > 32 into histograms.
    """

    def default(self, obj: Any) -> Any:
        obj, converted = json_friendly(obj)
        obj, compressed = maybe_compress_history(obj)
        if converted:
            return obj
        return json.JSONEncoder.default(self, obj)


class JSONEncoderUncompressed(json.JSONEncoder):
    """A JSON Encoder that handles some extra types.

    This encoder turns numpy like objects with a size > 32 into histograms.
    """

    def default(self, obj: Any) -> Any:
        if is_numpy_array(obj):
            return obj.tolist()
        elif np and isinstance(obj, np.number):
            return obj.item()
        elif np and isinstance(obj, np.generic):
            obj = obj.item()
        return json.JSONEncoder.default(self, obj)


def json_dump_safer(obj: Any, fp: IO[str], **kwargs: Any) -> None:
    """Convert obj to json, with some extra encodable types."""
    return dump(obj, fp, cls=WandBJSONEncoder, **kwargs)


def json_dumps_safer(obj: Any, **kwargs: Any) -> str:
    """Convert obj to json, with some extra encodable types."""
    return dumps(obj, cls=WandBJSONEncoder, **kwargs)


# This is used for dumping raw json into files
def json_dump_uncompressed(obj: Any, fp: IO[str], **kwargs: Any) -> None:
    """Convert obj to json, with some extra encodable types."""
    return dump(obj, fp, cls=JSONEncoderUncompressed, **kwargs)


def json_dumps_safer_history(obj: Any, **kwargs: Any) -> str:
    """Convert obj to json, with some extra encodable types, including histograms."""
    return dumps(obj, cls=WandBHistoryJSONEncoder, **kwargs)


def make_json_if_not_number(
    v: Union[int, float, str, Mapping, Sequence],
) -> Union[int, float, str]:
    """If v is not a basic type convert it to json."""
    if isinstance(v, (float, int)):
        return v
    return json_dumps_safer(v)


def make_safe_for_json(obj: Any) -> Any:
    """Replace invalid json floats with strings. Also converts to lists and dicts."""
    if isinstance(obj, Mapping):
        return {k: make_safe_for_json(v) for k, v in obj.items()}
    elif isinstance(obj, str):
        # str's are Sequence, so we need to short-circuit
        return obj
    elif isinstance(obj, Sequence):
        return [make_safe_for_json(v) for v in obj]
    elif isinstance(obj, float):
        # W&B backend and UI handle these strings
        if obj != obj:  # standard way to check for NaN
            return "NaN"
        elif obj == float("+inf"):
            return "Infinity"
        elif obj == float("-inf"):
            return "-Infinity"
    return obj


def no_retry_4xx(e: Exception) -> bool:
    if not isinstance(e, requests.HTTPError):
        return True
    assert e.response is not None
    if not (400 <= e.response.status_code < 500) or e.response.status_code == 429:
        return True
    body = json.loads(e.response.content)
    raise UsageError(body["errors"][0]["message"])


def parse_backend_error_messages(response: requests.Response) -> List[str]:
    """Returns error messages stored in a backend response.

    If the response is not in an expected format, an empty list is returned.

    Args:
        response: A response to an HTTP request to the W&B server.
    """
    try:
        data = response.json()
    except requests.JSONDecodeError:
        return []

    if not isinstance(data, dict):
        return []

    # Backend error values are returned in one of two ways:
    # - A string containing the error message
    # - A JSON object with a "message" field that is a string
    def get_message(error: Any) -> Optional[str]:
        if isinstance(error, str):
            return error
        elif (
            isinstance(error, dict)
            and (message := error.get("message"))
            and isinstance(message, str)
        ):
            return message
        else:
            return None

    # The response can contain an "error" field with a single error
    # or an "errors" field with a list of errors.
    if error := data.get("error"):
        message = get_message(error)
        return [message] if message else []

    elif (errors := data.get("errors")) and isinstance(errors, list):
        messages: List[str] = []
        for error in errors:
            message = get_message(error)
            if message:
                messages.append(message)
        return messages

    else:
        return []


def no_retry_auth(e: Any) -> bool:
    if hasattr(e, "exception"):
        e = e.exception
    if not isinstance(e, requests.HTTPError):
        return True
    if e.response is None:
        return True
    # Don't retry bad request errors; raise immediately
    if e.response.status_code in (400, 409):
        return False
    # Retry all non-forbidden/unauthorized/not-found errors.
    if e.response.status_code not in (401, 403, 404):
        return True

    # Crash with more informational message on forbidden/unauthorized errors.
    # UnauthorizedError
    if e.response.status_code == 401:
        raise AuthenticationError(
            "The API key you provided is either invalid or missing.  "
            f"If the `{wandb.env.API_KEY}` environment variable is set, make sure it is correct. "
            "Otherwise, to resolve this issue, you may try running the 'wandb login --relogin' command. "
            "If you are using a local server, make sure that you're using the correct hostname. "
            "If you're not sure, you can try logging in again using the 'wandb login --relogin --host [hostname]' command."
            f"(Error {e.response.status_code}: {e.response.reason})"
        )
    # ForbiddenError
    if e.response.status_code == 403:
        if wandb.run:
            raise CommError(f"Permission denied to access {wandb.run.path}")
        else:
            raise CommError(
                "It appears that you do not have permission to access the requested resource. "
                "Please reach out to the project owner to grant you access. "
                "If you have the correct permissions, verify that there are no issues with your networking setup."
                f"(Error {e.response.status_code}: {e.response.reason})"
            )

    # NotFoundError
    if e.response.status_code == 404:
        # If error message is empty, raise a more generic NotFoundError message.
        if parse_backend_error_messages(e.response):
            return False
        else:
            raise LookupError(
                f"Failed to find resource. Please make sure you have the correct resource path. "
                f"(Error {e.response.status_code}: {e.response.reason})"
            )
    return False


def check_retry_conflict(e: Any) -> Optional[bool]:
    """Check if the exception is a conflict type so it can be retried.

    Returns:
        True - Should retry this operation
        False - Should not retry this operation
        None - No decision, let someone else decide
    """
    if hasattr(e, "exception"):
        e = e.exception
    if isinstance(e, requests.HTTPError) and e.response is not None:
        if e.response.status_code == 409:
            return True
    return None


def check_retry_conflict_or_gone(e: Any) -> Optional[bool]:
    """Check if the exception is a conflict or gone type, so it can be retried or not.

    Returns:
        True - Should retry this operation
        False - Should not retry this operation
        None - No decision, let someone else decide
    """
    if hasattr(e, "exception"):
        e = e.exception
    if isinstance(e, requests.HTTPError) and e.response is not None:
        if e.response.status_code == 409:
            return True
        if e.response.status_code == 410:
            return False
    return None


def make_check_retry_fn(
    fallback_retry_fn: CheckRetryFnType,
    check_fn: Callable[[Exception], Optional[bool]],
    check_timedelta: Optional[timedelta] = None,
) -> CheckRetryFnType:
    """Return a check_retry_fn which can be used by lib.Retry().

    Args:
        fallback_fn: Use this function if check_fn didn't decide if a retry should happen.
        check_fn: Function which returns bool if retry should happen or None if unsure.
        check_timedelta: Optional retry timeout if we check_fn matches the exception
    """

    def check_retry_fn(e: Exception) -> Union[bool, timedelta]:
        check = check_fn(e)
        if check is None:
            return fallback_retry_fn(e)
        if check is False:
            return False
        if check_timedelta:
            return check_timedelta
        return True

    return check_retry_fn


def find_runner(program: str) -> Union[None, list, List[str]]:
    """Return a command that will run program.

    Args:
        program: The string name of the program to try to run.

    Returns:
        commandline list of strings to run the program (eg. with subprocess.call()) or None
    """
    if os.path.isfile(program) and not os.access(program, os.X_OK):
        # program is a path to a non-executable file
        try:
            opened = open(program)
        except OSError:  # PermissionError doesn't exist in 2.7
            return None
        first_line = opened.readline().strip()
        if first_line.startswith("#!"):
            return shlex.split(first_line[2:])
        if program.endswith(".py"):
            return [sys.executable]
    return None


def downsample(values: Sequence, target_length: int) -> list:
    """Downsample 1d values to target_length, including start and end.

    Algorithm just rounds index down.

    Values can be any sequence, including a generator.
    """
    if not target_length > 1:
        raise UsageError("target_length must be > 1")
    values = list(values)
    if len(values) < target_length:
        return values
    ratio = float(len(values) - 1) / (target_length - 1)
    result = []
    for i in range(target_length):
        result.append(values[int(i * ratio)])
    return result


def has_num(dictionary: Mapping, key: Any) -> bool:
    return key in dictionary and isinstance(dictionary[key], numbers.Number)


def docker_image_regex(image: str) -> Any:
    """Regex match for valid docker image names."""
    if image:
        return re.match(
            r"^(?:(?=[^:\/]{1,253})(?!-)[a-zA-Z0-9-]{1,63}(?<!-)(?:\.(?!-)[a-zA-Z0-9-]{1,63}(?<!-))*(?::[0-9]{1,5})?/)?((?![._-])(?:[a-z0-9._-]*)(?<![._-])(?:/(?![._-])[a-z0-9._-]*(?<![._-]))*)(?::(?![.-])[a-zA-Z0-9_.-]{1,128})?$",
            image,
        )
    return None


def image_from_docker_args(args: List[str]) -> Optional[str]:
    """Scan docker run args and attempt to find the most likely docker image argument.

    It excludes any arguments that start with a dash, and the argument after it if it
    isn't a boolean switch. This can be improved, we currently fallback gracefully when
    this fails.
    """
    bool_args = [
        "-t",
        "--tty",
        "--rm",
        "--privileged",
        "--oom-kill-disable",
        "--no-healthcheck",
        "-i",
        "--interactive",
        "--init",
        "--help",
        "--detach",
        "-d",
        "--sig-proxy",
        "-it",
        "-itd",
    ]
    last_flag = -2
    last_arg = ""
    possible_images = []
    if len(args) > 0 and args[0] == "run":
        args.pop(0)
    for i, arg in enumerate(args):
        if arg.startswith("-"):
            last_flag = i
            last_arg = arg
        elif "@sha256:" in arg:
            # Because our regex doesn't match digests
            possible_images.append(arg)
        elif docker_image_regex(arg):
            if last_flag == i - 2:
                possible_images.append(arg)
            elif "=" in last_arg:
                possible_images.append(arg)
            elif last_arg in bool_args and last_flag == i - 1:
                possible_images.append(arg)
    most_likely = None
    for img in possible_images:
        if ":" in img or "@" in img or "/" in img:
            most_likely = img
            break
    if most_likely is None and len(possible_images) > 0:
        most_likely = possible_images[0]
    return most_likely


def load_yaml(file: Any) -> Any:
    return yaml.safe_load(file)


def image_id_from_k8s() -> Optional[str]:
    """Ping the k8s metadata service for the image id.

    Specify the KUBERNETES_NAMESPACE environment variable if your pods are not in the
    default namespace:

    - name: KUBERNETES_NAMESPACE valueFrom:
        fieldRef:
          fieldPath: metadata.namespace
    """
    token_path = "/var/run/secrets/kubernetes.io/serviceaccount/token"

    if not os.path.exists(token_path):
        return None

    try:
        with open(token_path) as token_file:
            token = token_file.read()
    except FileNotFoundError:
        logger.warning(f"Token file not found at {token_path}.")
        return None
    except PermissionError as e:
        current_uid = os.getuid()
        warning = (
            f"Unable to read the token file at {token_path} due to permission error ({e})."
            f"The current user id is {current_uid}. "
            "Consider changing the securityContext to run the container as the current user."
        )
        logger.warning(warning)
        wandb.termwarn(warning)
        return None

    if not token:
        return None

    k8s_server = "https://{}:{}/api/v1/namespaces/{}/pods/{}".format(
        os.getenv("KUBERNETES_SERVICE_HOST"),
        os.getenv("KUBERNETES_PORT_443_TCP_PORT"),
        os.getenv("KUBERNETES_NAMESPACE", "default"),
        os.getenv("HOSTNAME"),
    )
    try:
        res = requests.get(
            k8s_server,
            verify="/var/run/secrets/kubernetes.io/serviceaccount/ca.crt",
            timeout=3,
            headers={"Authorization": f"Bearer {token}"},
        )
        res.raise_for_status()
    except requests.RequestException:
        return None
    try:
        return str(  # noqa: B005
            res.json()["status"]["containerStatuses"][0]["imageID"]
        ).strip("docker-pullable://")
    except (ValueError, KeyError, IndexError):
        logger.exception("Error checking kubernetes for image id")
        return None


def async_call(
    target: Callable, timeout: Optional[Union[int, float]] = None
) -> Callable:
    """Wrap a method to run in the background with an optional timeout.

    Returns a new method that will call the original with any args, waiting for upto
    timeout seconds. This new method blocks on the original and returns the result or
    None if timeout was reached, along with the thread. You can check thread.is_alive()
    to determine if a timeout was reached. If an exception is thrown in the thread, we
    reraise it.
    """
    q: queue.Queue = queue.Queue()

    def wrapped_target(q: "queue.Queue", *args: Any, **kwargs: Any) -> Any:
        try:
            q.put(target(*args, **kwargs))
        except Exception as e:
            q.put(e)

    def wrapper(
        *args: Any, **kwargs: Any
    ) -> Union[Tuple[Exception, "threading.Thread"], Tuple[None, "threading.Thread"]]:
        thread = threading.Thread(
            target=wrapped_target, args=(q,) + args, kwargs=kwargs
        )
        thread.daemon = True
        thread.start()

        try:
            result = q.get(True, timeout)
        except queue.Empty:
            return None, thread

        if isinstance(result, Exception):
            raise result.with_traceback(sys.exc_info()[2])
        return result, thread

    return wrapper


def read_many_from_queue(
    q: "queue.Queue", max_items: int, queue_timeout: Union[int, float]
) -> list:
    try:
        item = q.get(True, queue_timeout)
    except queue.Empty:
        return []
    items = [item]
    for _ in range(max_items):
        try:
            item = q.get_nowait()
        except queue.Empty:
            return items
        items.append(item)
    return items


def stopwatch_now() -> float:
    """Get a time value for interval comparisons.

    When possible it is a monotonic clock to prevent backwards time issues.
    """
    return time.monotonic()


def class_colors(class_count: int) -> List[List[int]]:
    # make class 0 black, and the rest equally spaced fully saturated hues
    return [[0, 0, 0]] + [
        colorsys.hsv_to_rgb(i / (class_count - 1.0), 1.0, 1.0)  # type: ignore
        for i in range(class_count - 1)
    ]


def _prompt_choice(
    input_timeout: Union[int, float, None] = None,
    jupyter: bool = False,
) -> str:
    input_fn: Callable = input
    prompt = term.LOG_STRING
    if input_timeout is not None:
        # delayed import to mitigate risk of timed_input complexity
        from wandb.sdk.lib import timed_input

        input_fn = functools.partial(timed_input.timed_input, timeout=input_timeout)
        # timed_input doesn't handle enhanced prompts
        if platform.system() == "Windows":
            prompt = "wandb"

    text = f"{prompt}: Enter your choice: "
    if input_fn == input:
        choice = input_fn(text)
    else:
        choice = input_fn(text, jupyter=jupyter)
    return choice  # type: ignore


def prompt_choices(
    choices: Sequence[str],
    input_timeout: Union[int, float, None] = None,
    jupyter: bool = False,
) -> str:
    """Allow a user to choose from a list of options."""
    for i, choice in enumerate(choices):
        wandb.termlog(f"({i + 1}) {choice}")

    idx = -1
    while idx < 0 or idx > len(choices) - 1:
        choice = _prompt_choice(input_timeout=input_timeout, jupyter=jupyter)
        if not choice:
            continue
        idx = -1
        try:
            idx = int(choice) - 1
        except ValueError:
            pass
        if idx < 0 or idx > len(choices) - 1:
            wandb.termwarn("Invalid choice")
    result = choices[idx]
    wandb.termlog(f"You chose {result!r}")
    return result


def guess_data_type(shape: Sequence[int], risky: bool = False) -> Optional[str]:
    """Infer the type of data based on the shape of the tensors.

    Args:
        shape (Sequence[int]): The shape of the data
        risky(bool): some guesses are more likely to be wrong.
    """
    # (samples,) or (samples,logits)
    if len(shape) in (1, 2):
        return "label"
    # Assume image mask like fashion mnist: (no color channel)
    # This is risky because RNNs often have 3 dim tensors: batch, time, channels
    if risky and len(shape) == 3:
        return "image"
    if len(shape) == 4:
        if shape[-1] in (1, 3, 4):
            # (samples, height, width, Y \ RGB \ RGBA)
            return "image"
        else:
            # (samples, height, width, logits)
            return "segmentation_mask"
    return None


def download_file_from_url(
    dest_path: str, source_url: str, api_key: Optional[str] = None
) -> None:
    auth = None
    if not _thread_local_api_settings.cookies:
        auth = ("api", api_key or "")
    response = requests.get(
        source_url,
        auth=auth,
        headers=_thread_local_api_settings.headers,
        cookies=_thread_local_api_settings.cookies,
        stream=True,
        timeout=5,
    )
    response.raise_for_status()

    if os.sep in dest_path:
        filesystem.mkdir_exists_ok(os.path.dirname(dest_path))
    with fsync_open(dest_path, "wb") as file:
        for data in response.iter_content(chunk_size=1024):
            file.write(data)


def download_file_into_memory(source_url: str, api_key: Optional[str] = None) -> bytes:
    auth = None
    if not _thread_local_api_settings.cookies:
        auth = ("api", api_key or "")
    response = requests.get(
        source_url,
        auth=auth,
        headers=_thread_local_api_settings.headers,
        cookies=_thread_local_api_settings.cookies,
        stream=True,
        timeout=5,
    )
    response.raise_for_status()
    return response.content


def isatty(ob: IO) -> bool:
    return hasattr(ob, "isatty") and ob.isatty()


def to_human_size(size: int, units: Optional[List[Tuple[str, Any]]] = None) -> str:
    units = units or POW_10_BYTES
    unit, value = units[0]
    factor = round(float(size) / value, 1)
    return (
        f"{factor}{unit}"
        if factor < 1024 or len(units) == 1
        else to_human_size(size, units[1:])
    )


def from_human_size(size: str, units: Optional[List[Tuple[str, Any]]] = None) -> int:
    units = units or POW_10_BYTES
    units_dict = {unit.upper(): value for (unit, value) in units}
    regex = re.compile(
        r"(\d+\.?\d*)\s*({})?".format("|".join(units_dict.keys())), re.IGNORECASE
    )
    match = re.match(regex, size)
    if not match:
        raise ValueError("size must be of the form `10`, `10B` or `10 B`.")
    factor, unit = (
        float(match.group(1)),
        units_dict[match.group(2).upper()] if match.group(2) else 1,
    )
    return int(factor * unit)


def auto_project_name(program: Optional[str]) -> str:
    # if we're in git, set project name to git repo name + relative path within repo
    from wandb.sdk.lib.gitlib import GitRepo

    root_dir = GitRepo().root_dir
    if root_dir is None:
        return "uncategorized"
    # On windows, GitRepo returns paths in unix style, but os.path is windows
    # style. Coerce here.
    root_dir = to_native_slash_path(root_dir)
    repo_name = os.path.basename(root_dir)
    if program is None:
        return str(repo_name)
    if not os.path.isabs(program):
        program = os.path.join(os.curdir, program)
    prog_dir = os.path.dirname(os.path.abspath(program))
    if not prog_dir.startswith(root_dir):
        return str(repo_name)
    project = repo_name
    sub_path = os.path.relpath(prog_dir, root_dir)
    if sub_path != ".":
        project += "-" + sub_path
    return str(project.replace(os.sep, "_"))


def are_paths_on_same_drive(path1: str, path2: str) -> bool:
    """Check if two paths are on the same drive.

    This check is only relevant on Windows,
    since the concept of drives only exists on Windows.
    """
    if platform.system() != "Windows":
        return True

    try:
        path1_drive = pathlib.Path(path1).resolve().drive
        path2_drive = pathlib.Path(path2).resolve().drive
    except OSError:
        # If either path is not a valid Windows path, an OSError is raised.
        return False

    return path1_drive == path2_drive


# TODO(hugh): Deprecate version here and use wandb/sdk/lib/paths.py
def to_forward_slash_path(path: str) -> str:
    if platform.system() == "Windows":
        path = path.replace("\\", "/")
    return path


# TODO(hugh): Deprecate version here and use wandb/sdk/lib/paths.py
def to_native_slash_path(path: str) -> FilePathStr:
    return FilePathStr(path.replace("/", os.sep))


def check_and_warn_old(files: List[str]) -> bool:
    if "wandb-metadata.json" in files:
        wandb.termwarn("These runs were logged with a previous version of wandb.")
        wandb.termwarn(
            "Run pip install wandb<0.10.0 to get the old library and sync your runs."
        )
        return True
    return False


class ImportMetaHook:
    def __init__(self) -> None:
        self.modules: Dict[str, ModuleType] = dict()
        self.on_import: Dict[str, list] = dict()

    def add(self, fullname: str, on_import: Callable) -> None:
        self.on_import.setdefault(fullname, []).append(on_import)

    def install(self) -> None:
        sys.meta_path.insert(0, self)  # type: ignore

    def uninstall(self) -> None:
        sys.meta_path.remove(self)  # type: ignore

    def find_module(
        self, fullname: str, path: Optional[str] = None
    ) -> Optional["ImportMetaHook"]:
        if fullname in self.on_import:
            return self
        return None

    def load_module(self, fullname: str) -> ModuleType:
        self.uninstall()
        mod = importlib.import_module(fullname)
        self.install()
        self.modules[fullname] = mod
        on_imports = self.on_import.get(fullname)
        if on_imports:
            for f in on_imports:
                f()
        return mod

    def get_modules(self) -> Tuple[str, ...]:
        return tuple(self.modules)

    def get_module(self, module: str) -> ModuleType:
        return self.modules[module]


_import_hook: Optional[ImportMetaHook] = None


def add_import_hook(fullname: str, on_import: Callable) -> None:
    global _import_hook
    if _import_hook is None:
        _import_hook = ImportMetaHook()
        _import_hook.install()
    _import_hook.add(fullname, on_import)


def host_from_path(path: Optional[str]) -> str:
    """Return the host of the path."""
    url = urllib.parse.urlparse(path)
    return str(url.netloc)


def uri_from_path(path: Optional[str]) -> str:
    """Return the URI of the path."""
    url = urllib.parse.urlparse(path)
    uri = url.path if url.path[0] != "/" else url.path[1:]
    return str(uri)


def is_unicode_safe(stream: TextIO) -> bool:
    """Return True if the stream supports UTF-8."""
    encoding = getattr(stream, "encoding", None)
    return encoding.lower() in {"utf-8", "utf_8"} if encoding else False


def _has_internet() -> bool:
    """Returns whether we have internet access.

    Checks for internet access by attempting to open a DNS connection to
    Google's root servers.
    """
    try:
        s = socket.create_connection(("8.8.8.8", 53), 0.5)
        s.close()
    except OSError:
        return False

    return True


def rand_alphanumeric(
    length: int = 8, rand: Optional[Union[ModuleType, random.Random]] = None
) -> str:
    wandb.termerror("rand_alphanumeric is deprecated, use 'secrets.token_hex'")
    rand = rand or random
    return "".join(rand.choice("0123456789ABCDEF") for _ in range(length))


@contextlib.contextmanager
def fsync_open(
    path: StrPath, mode: str = "w", encoding: Optional[str] = None
) -> Generator[IO[Any], None, None]:
    """Open a path for I/O and guarantee that the file is flushed and synced."""
    with open(path, mode, encoding=encoding) as f:
        yield f

        f.flush()
        os.fsync(f.fileno())


def _is_kaggle() -> bool:
    return (
        os.getenv("KAGGLE_KERNEL_RUN_TYPE") is not None
        or "kaggle_environments" in sys.modules
    )


def _is_likely_kaggle() -> bool:
    # Telemetry to mark first runs from Kagglers.
    return (
        _is_kaggle()
        or os.path.exists(
            os.path.expanduser(os.path.join("~", ".kaggle", "kaggle.json"))
        )
        or "kaggle" in sys.modules
    )


def _is_databricks() -> bool:
    # check if we are running inside a databricks notebook by
    # inspecting sys.modules, searching for dbutils and verifying that
    # it has the appropriate structure

    if "dbutils" in sys.modules:
        dbutils = sys.modules["dbutils"]
        if hasattr(dbutils, "shell"):
            shell = dbutils.shell
            if hasattr(shell, "sc"):
                sc = shell.sc
                if hasattr(sc, "appName"):
                    return bool(sc.appName == "Databricks Shell")
    return False


def _is_py_requirements_or_dockerfile(path: str) -> bool:
    file = os.path.basename(path)
    return (
        file.endswith(".py")
        or file.startswith("Dockerfile")
        or file == "requirements.txt"
    )


def artifact_to_json(artifact: "Artifact") -> Dict[str, Any]:
    return {
        "_type": "artifactVersion",
        "_version": "v0",
        "id": artifact.id,
        "version": artifact.source_version,
        "sequenceName": artifact.source_name.split(":")[0],
        "usedAs": artifact.use_as,
    }


def check_dict_contains_nested_artifact(d: dict, nested: bool = False) -> bool:
    for item in d.values():
        if isinstance(item, dict):
            contains_artifacts = check_dict_contains_nested_artifact(item, True)
            if contains_artifacts:
                return True
        elif (isinstance(item, wandb.Artifact) or _is_artifact_string(item)) and nested:
            return True
    return False


def load_json_yaml_dict(config: str) -> Any:
    ext = os.path.splitext(config)[-1]
    if ext == ".json":
        with open(config) as f:
            return json.load(f)
    elif ext == ".yaml":
        with open(config) as f:
            return yaml.safe_load(f)
    else:
        try:
            return json.loads(config)
        except ValueError:
            return None


def _parse_entity_project_item(path: str) -> tuple:
    """Parse paths with the following formats: {item}, {project}/{item}, & {entity}/{project}/{item}.

    Args:
        path: `str`, input path; must be between 0 and 3 in length.

    Returns:
        tuple of length 3 - (item, project, entity)

    Example:
        alias, project, entity = _parse_entity_project_item("myproj/mymodel:best")

        assert entity   == ""
        assert project  == "myproj"
        assert alias    == "mymodel:best"

    """
    words = path.split("/")
    if len(words) > 3:
        raise ValueError(
            "Invalid path: must be str the form {item}, {project}/{item}, or {entity}/{project}/{item}"
        )
    padded_words = [""] * (3 - len(words)) + words
    return tuple(reversed(padded_words))


def _resolve_aliases(aliases: Optional[Union[str, Iterable[str]]]) -> List[str]:
    """Add the 'latest' alias and ensure that all aliases are unique.

    Takes in `aliases` which can be None, str, or List[str] and returns List[str].
    Ensures that "latest" is always present in the returned list.

    Args:
        aliases: `Optional[Union[str, List[str]]]`

    Returns:
        List[str], with "latest" always present.

    Usage:

    ```python
    aliases = _resolve_aliases(["best", "dev"])
    assert aliases == ["best", "dev", "latest"]

    aliases = _resolve_aliases("boom")
    assert aliases == ["boom", "latest"]
    ```
    """
    aliases = aliases or ["latest"]

    if isinstance(aliases, str):
        aliases = [aliases]

    try:
        return list(set(aliases) | {"latest"})
    except TypeError as exc:
        raise ValueError("`aliases` must be Iterable or None") from exc


def _is_artifact_object(v: Any) -> "TypeGuard[wandb.Artifact]":
    return isinstance(v, wandb.Artifact)


def _is_artifact_string(v: Any) -> "TypeGuard[str]":
    return isinstance(v, str) and v.startswith("wandb-artifact://")


def _is_artifact_version_weave_dict(v: Any) -> "TypeGuard[dict]":
    return isinstance(v, dict) and v.get("_type") == "artifactVersion"


def _is_artifact_representation(v: Any) -> bool:
    return (
        _is_artifact_object(v)
        or _is_artifact_string(v)
        or _is_artifact_version_weave_dict(v)
    )


def parse_artifact_string(v: str) -> Tuple[str, Optional[str], bool]:
    if not v.startswith("wandb-artifact://"):
        raise ValueError(f"Invalid artifact string: {v}")
    parsed_v = v[len("wandb-artifact://") :]
    base_uri = None
    url_info = urllib.parse.urlparse(parsed_v)
    if url_info.scheme != "":
        base_uri = f"{url_info.scheme}://{url_info.netloc}"
        parts = url_info.path.split("/")[1:]
    else:
        parts = parsed_v.split("/")
    if parts[0] == "_id":
        # for now can't fetch paths but this will be supported in the future
        # when we allow passing typed media objects, this can be extended
        # to include paths
        return parts[1], base_uri, True

    if len(parts) < 3:
        raise ValueError(f"Invalid artifact string: {v}")

    # for now can't fetch paths but this will be supported in the future
    # when we allow passing typed media objects, this can be extended
    # to include paths
    entity, project, name_and_alias_or_version = parts[:3]
    return f"{entity}/{project}/{name_and_alias_or_version}", base_uri, False


def _get_max_cli_version() -> Union[str, None]:
    max_cli_version = wandb.api.max_cli_version()
    return str(max_cli_version) if max_cli_version is not None else None


def ensure_text(
    string: Union[str, bytes], encoding: str = "utf-8", errors: str = "strict"
) -> str:
    """Coerce s to str."""
    if isinstance(string, bytes):
        return string.decode(encoding, errors)
    elif isinstance(string, str):
        return string
    else:
        raise TypeError(f"not expecting type {type(string)!r}")


def make_artifact_name_safe(name: str) -> str:
    """Make an artifact name safe for use in artifacts."""
    # artifact names may only contain alphanumeric characters, dashes, underscores, and dots.
    cleaned = re.sub(r"[^a-zA-Z0-9_\-.]", "_", name)
    if len(cleaned) <= 128:
        return cleaned
    # truncate with dots in the middle using regex
    return re.sub(r"(^.{63}).*(.{63}$)", r"\g<1>..\g<2>", cleaned)


def make_docker_image_name_safe(name: str) -> str:
    """Make a docker image name safe for use in artifacts."""
    safe_chars = RE_DOCKER_IMAGE_NAME_CHARS.sub("__", name.lower())
    deduped = RE_DOCKER_IMAGE_NAME_SEPARATOR_REPEAT.sub("__", safe_chars)
    trimmed_start = RE_DOCKER_IMAGE_NAME_SEPARATOR_START.sub("", deduped)
    trimmed = RE_DOCKER_IMAGE_NAME_SEPARATOR_END.sub("", trimmed_start)
    return trimmed if trimmed else "image"


def merge_dicts(
    source: Dict[str, Any],
    destination: Dict[str, Any],
) -> Dict[str, Any]:
    """Recursively merge two dictionaries.

    This mutates the destination and its nested dictionaries and lists.

    Instances of `dict` are recursively merged and instances of `list`
    are appended to the destination. If the destination type is not
    `dict` or `list`, respectively, the key is overwritten with the
    source value.

    For all other types, the source value overwrites the destination value.
    """
    for key, value in source.items():
        if isinstance(value, dict):
            node = destination.get(key)
            if isinstance(node, dict):
                merge_dicts(value, node)
            else:
                destination[key] = value

        elif isinstance(value, list):
            dest_value = destination.get(key)
            if isinstance(dest_value, list):
                dest_value.extend(value)
            else:
                destination[key] = value

        else:
            destination[key] = value

    return destination


def coalesce(*arg: Any) -> Any:
    """Return the first non-none value in the list of arguments.

    Similar to ?? in C#.
    """
    return next((a for a in arg if a is not None), None)


def recursive_cast_dictlike_to_dict(d: Dict[str, Any]) -> Dict[str, Any]:
    for k, v in d.items():
        if isinstance(v, dict):
            recursive_cast_dictlike_to_dict(v)
        elif hasattr(v, "keys"):
            d[k] = dict(v)
            recursive_cast_dictlike_to_dict(d[k])
    return d


def remove_keys_with_none_values(
    d: Union[Dict[str, Any], Any],
) -> Union[Dict[str, Any], Any]:
    # otherwise iterrows will create a bunch of ugly charts
    if not isinstance(d, dict):
        return d

    if isinstance(d, dict):
        new_dict = {}
        for k, v in d.items():
            new_v = remove_keys_with_none_values(v)
            if new_v is not None and not (isinstance(new_v, dict) and len(new_v) == 0):
                new_dict[k] = new_v
        return new_dict if new_dict else None


def batched(n: int, iterable: Iterable[T]) -> Generator[List[T], None, None]:
    i = iter(iterable)
    batch = list(itertools.islice(i, n))
    while batch:
        yield batch
        batch = list(itertools.islice(i, n))


def random_string(length: int = 12) -> str:
    """Generate a random string of a given length.

    :param length: Length of the string to generate.
    :return: Random string.
    """
    return "".join(
        secrets.choice(string.ascii_lowercase + string.digits) for _ in range(length)
    )


def sample_with_exponential_decay_weights(
    xs: Union[Iterable, Iterable[Iterable]],
    ys: Iterable[Iterable],
    keys: Optional[Iterable] = None,
    sample_size: int = 1500,
) -> Tuple[List, List, Optional[List]]:
    """Sample from a list of lists with weights that decay exponentially.

    May be used with the wandb.plot.line_series function.
    """
    xs_array = np.array(xs)
    ys_array = np.array(ys)
    keys_array = np.array(keys) if keys else None
    weights = np.exp(-np.arange(len(xs_array)) / len(xs_array))
    weights /= np.sum(weights)
    sampled_indices = np.random.choice(len(xs_array), size=sample_size, p=weights)
    sampled_xs = xs_array[sampled_indices].tolist()
    sampled_ys = ys_array[sampled_indices].tolist()
    sampled_keys = keys_array[sampled_indices].tolist() if keys_array else None

    return sampled_xs, sampled_ys, sampled_keys


@dataclasses.dataclass(frozen=True)
class InstalledDistribution:
    """An installed distribution.

    Attributes:
        key: The distribution name as it would be imported.
        version: The distribution's version string.
    """

    key: str
    version: str


def working_set() -> Iterable[InstalledDistribution]:
    """Return the working set of installed distributions."""
    from importlib.metadata import distributions

    for d in distributions():
        try:
            # In some distributions, the "Name" attribute may not be present,
            # which can raise a KeyError. To handle this, we catch the exception
            # and skip those distributions.
            # For additional context, see: https://github.com/python/importlib_metadata/issues/371.

            # From Sentry events we observed that UnicodeDecodeError can occur when
            # trying to decode the metadata of a distribution. To handle this, we catch
            # the exception and skip those distributions.
            yield InstalledDistribution(key=d.metadata["Name"], version=d.version)
        except (KeyError, UnicodeDecodeError):
            pass


def get_core_path() -> str:
    """Returns the path to the wandb-core binary.

    The path can be set explicitly via the _WANDB_CORE_PATH environment
    variable. Otherwise, the path to the binary in the current package
    is returned.

    Returns:
        str: The path to the wandb-core package.

    Raises:
        WandbCoreNotAvailableError: If wandb-core was not built for the current system.
    """
    # NOTE: Environment variable _WANDB_CORE_PATH is a temporary development feature
    #       to assist in running the core service from a live development directory.
    path_from_env: str = os.environ.get("_WANDB_CORE_PATH", "")
    if path_from_env:
        wandb.termwarn(
            f"Using wandb-core from path `_WANDB_CORE_PATH={path_from_env}`. "
            "This is a development feature and may not work as expected."
        )
        return path_from_env

    bin_path = pathlib.Path(__file__).parent / "bin" / "wandb-core"
    if not bin_path.exists():
        raise WandbCoreNotAvailableError(
            f"File not found: {bin_path}."
            " Please contact support at support@wandb.com."
            f" Your platform is: {platform.platform()}."
        )

    return str(bin_path)


class NonOctalStringDumper(yaml.Dumper):
    """Prevents strings containing non-octal values like "008" and "009" from being converted to numbers in in the yaml string saved as the sweep config."""

    def represent_scalar(self, tag, value, style=None):
        if tag == "tag:yaml.org,2002:str" and value.startswith("0") and len(value) > 1:
            return super().represent_scalar(tag, value, style="'")
        return super().represent_scalar(tag, value, style)