wandb 0.19.1rc1__py3-none-musllinux_1_2_aarch64.whl

wandb/sdk/data_types/_private.py

@@ -0,0 +1,10 @@
+import atexit
+import tempfile
+
+# Staging directory, so we can encode raw data into files, then hash them before
+# we put them into the Run directory to be uploaded.
+MEDIA_TMP = tempfile.TemporaryDirectory("wandb-media")
+
+
+def _cleanup_media_tmp_dir() -> None:
+    atexit.register(MEDIA_TMP.cleanup)

wandb/sdk/data_types/audio.py

@@ -0,0 +1,165 @@
+import hashlib
+import os
+from typing import Optional
+
+from wandb import util
+from wandb.sdk.lib import filesystem, runid
+
+from . import _dtypes
+from ._private import MEDIA_TMP
+from .base_types.media import BatchableMedia
+
+
+class Audio(BatchableMedia):
+    """Wandb class for audio clips.
+
+    Args:
+        data_or_path: (string or numpy array) A path to an audio file
+            or a numpy array of audio data.
+        sample_rate: (int) Sample rate, required when passing in raw
+            numpy array of audio data.
+        caption: (string) Caption to display with audio.
+    """
+
+    _log_type = "audio-file"
+
+    def __init__(self, data_or_path, sample_rate=None, caption=None):
+        """Accept a path to an audio file or a numpy array of audio data."""
+        super().__init__()
+        self._duration = None
+        self._sample_rate = sample_rate
+        self._caption = caption
+
+        if isinstance(data_or_path, str):
+            if self.path_is_reference(data_or_path):
+                self._path = data_or_path
+                self._sha256 = hashlib.sha256(data_or_path.encode("utf-8")).hexdigest()
+                self._is_tmp = False
+            else:
+                self._set_file(data_or_path, is_tmp=False)
+        else:
+            if sample_rate is None:
+                raise ValueError(
+                    'Argument "sample_rate" is required when instantiating wandb.Audio with raw data.'
+                )
+
+            soundfile = util.get_module(
+                "soundfile",
+                required='Raw audio requires the soundfile package. To get it, run "pip install soundfile"',
+            )
+
+            tmp_path = os.path.join(MEDIA_TMP.name, runid.generate_id() + ".wav")
+            soundfile.write(tmp_path, data_or_path, sample_rate)
+            self._duration = len(data_or_path) / float(sample_rate)
+
+            self._set_file(tmp_path, is_tmp=True)
+
+    @classmethod
+    def get_media_subdir(cls):
+        return os.path.join("media", "audio")
+
+    @classmethod
+    def from_json(cls, json_obj, source_artifact):
+        return cls(
+            source_artifact.get_entry(json_obj["path"]).download(),
+            caption=json_obj["caption"],
+        )
+
+    def bind_to_run(
+        self, run, key, step, id_=None, ignore_copy_err: Optional[bool] = None
+    ):
+        if self.path_is_reference(self._path):
+            raise ValueError(
+                "Audio media created by a reference to external storage cannot currently be added to a run"
+            )
+
+        return super().bind_to_run(run, key, step, id_, ignore_copy_err)
+
+    def to_json(self, run):
+        json_dict = super().to_json(run)
+        json_dict.update(
+            {
+                "_type": self._log_type,
+                "caption": self._caption,
+            }
+        )
+        return json_dict
+
+    @classmethod
+    def seq_to_json(cls, seq, run, key, step):
+        audio_list = list(seq)
+
+        util.get_module(
+            "soundfile",
+            required="wandb.Audio requires the soundfile package. To get it, run: pip install soundfile",
+        )
+        base_path = os.path.join(run.dir, "media", "audio")
+        filesystem.mkdir_exists_ok(base_path)
+        meta = {
+            "_type": "audio",
+            "count": len(audio_list),
+            "audio": [a.to_json(run) for a in audio_list],
+        }
+        sample_rates = cls.sample_rates(audio_list)
+        if sample_rates:
+            meta["sampleRates"] = sample_rates
+        durations = cls.durations(audio_list)
+        if durations:
+            meta["durations"] = durations
+        captions = cls.captions(audio_list)
+        if captions:
+            meta["captions"] = captions
+
+        return meta
+
+    @classmethod
+    def durations(cls, audio_list):
+        return [a._duration for a in audio_list]
+
+    @classmethod
+    def sample_rates(cls, audio_list):
+        return [a._sample_rate for a in audio_list]
+
+    @classmethod
+    def captions(cls, audio_list):
+        captions = [a._caption for a in audio_list]
+        if all(c is None for c in captions):
+            return False
+        else:
+            return ["" if c is None else c for c in captions]
+
+    def resolve_ref(self):
+        if self.path_is_reference(self._path):
+            # this object was already created using a ref:
+            return self._path
+        source_artifact = self._artifact_source.artifact
+
+        resolved_name = source_artifact._local_path_to_name(self._path)
+        if resolved_name is not None:
+            target_entry = source_artifact.manifest.get_entry_by_path(resolved_name)
+            if target_entry is not None:
+                return target_entry.ref
+
+        return None
+
+    def __eq__(self, other):
+        if self.path_is_reference(self._path) or self.path_is_reference(other._path):
+            # one or more of these objects is an unresolved reference -- we'll compare
+            # their reference paths instead of their SHAs:
+            return (
+                self.resolve_ref() == other.resolve_ref()
+                and self._caption == other._caption
+            )
+
+        return super().__eq__(other) and self._caption == other._caption
+
+    def __ne__(self, other):
+        return not self.__eq__(other)
+
+
+class _AudioFileType(_dtypes.Type):
+    name = "audio-file"
+    types = [Audio]
+
+
+_dtypes.TypeRegistry.add(_AudioFileType)

wandb/sdk/data_types/base_types/json_metadata.py

@@ -0,0 +1,55 @@
+import codecs
+import os
+from typing import TYPE_CHECKING, Type, Union
+
+from wandb import util
+from wandb.sdk.lib import runid
+
+from .._private import MEDIA_TMP
+from .media import Media
+
+if TYPE_CHECKING:  # pragma: no cover
+    from wandb.sdk.artifacts.artifact import Artifact
+
+    from ...wandb_run import Run as LocalRun
+
+
+# Allows encoding of arbitrary JSON structures
+# as a file
+#
+# This class should be used as an abstract class
+# extended to have validation methods
+
+
+class JSONMetadata(Media):
+    """JSONMetadata is a type for encoding arbitrary metadata as files."""
+
+    def __init__(self, val: dict) -> None:
+        super().__init__()
+
+        self.validate(val)
+        self._val = val
+
+        ext = "." + self.type_name() + ".json"
+        tmp_path = os.path.join(MEDIA_TMP.name, runid.generate_id() + ext)
+        with codecs.open(tmp_path, "w", encoding="utf-8") as fp:
+            util.json_dump_uncompressed(self._val, fp)
+        self._set_file(tmp_path, is_tmp=True, extension=ext)
+
+    @classmethod
+    def get_media_subdir(cls: Type["JSONMetadata"]) -> str:
+        return os.path.join("media", "metadata", cls.type_name())
+
+    def to_json(self, run_or_artifact: Union["LocalRun", "Artifact"]) -> dict:
+        json_dict = super().to_json(run_or_artifact)
+        json_dict["_type"] = self.type_name()
+
+        return json_dict
+
+    # These methods should be overridden in the child class
+    @classmethod
+    def type_name(cls) -> str:
+        return "metadata"
+
+    def validate(self, val: dict) -> bool:
+        return True

wandb/sdk/data_types/base_types/media.py

@@ -0,0 +1,376 @@
+import hashlib
+import os
+import platform
+import re
+import shutil
+from typing import TYPE_CHECKING, Optional, Sequence, Type, Union, cast
+
+import wandb
+from wandb import util
+from wandb._globals import _datatypes_callback
+from wandb.sdk.lib import filesystem
+from wandb.sdk.lib.paths import LogicalPath
+
+from .wb_value import WBValue
+
+if TYPE_CHECKING:  # pragma: no cover
+    import numpy as np
+
+    from wandb.sdk.artifacts.artifact import Artifact
+
+    from ...wandb_run import Run as LocalRun
+
+
+SYS_PLATFORM = platform.system()
+
+
+def check_windows_valid_filename(path: Union[int, str]) -> bool:
+    r"""Verify that the given path does not contain any invalid characters for a Windows filename.
+
+    Windows filenames cannot contain the following characters:
+    < > : " \ / | ? *
+
+    For more details, refer to the official documentation:
+    https://learn.microsoft.com/en-us/windows/win32/fileio/naming-a-file#naming-conventions
+
+    Args:
+        path: The file path to check, which can be either an integer or a string.
+
+    Returns:
+        bool: True if the path does not contain any invalid characters, False otherwise.
+    """
+    return not bool(re.search(r'[<>:"\\?*]', path))  # type: ignore
+
+
+def _wb_filename(
+    key: Union[str, int], step: Union[str, int], id: Union[str, int], extension: str
+) -> str:
+    r"""Generates a safe filename/path for storing media files, using the provided key, step, and id.
+
+    The filename is made safe by:
+    1. Removing any leading slashes to prevent writing to absolute paths
+    2. Replacing '.' and '..' with underscores to prevent directory traversal attacks
+
+    If the key contains slashes (e.g. 'images/cats/fluffy.jpg'), subdirectories will be created:
+        media/
+          images/
+            cats/
+              fluffy.jpg_step_id.ext
+
+    Args:
+        key: Name/path for the media file
+        step: Training step number
+        id: Unique identifier
+        extension: File extension (e.g. '.jpg', '.mp3')
+
+    Returns:
+        A sanitized filename string in the format: key_step_id.extension
+
+    Raises:
+        ValueError: If running on Windows and the key contains invalid filename characters
+                   (\\, :, *, ?, ", <, >, |)
+    """
+    if SYS_PLATFORM == "Windows" and not check_windows_valid_filename(key):
+        raise ValueError(
+            f'Media {key} is invalid. Please remove invalid filename characters (\\, :, *, ?, ", <, >, |)'
+        )
+
+    # On Windows, convert forward slashes to backslashes.
+    # This ensures that the key is a valid filename on Windows.
+    if SYS_PLATFORM == "Windows":
+        key = str(key).replace("/", os.sep)
+
+    # Avoid writing to absolute paths by striping any leading slashes.
+    # The key has already been validated for windows operating systems in util.check_windows_valid_filename
+    # This ensures the key does not contain invalid characters for windows, such as '\' or ':'.
+    # So we can check only for '/' in the key.
+    key = str(key).lstrip(os.sep)
+
+    # Avoid directory traversal by replacing dots with underscores.
+    keys = key.split(os.sep)
+    keys = [k.replace(".", "_") if k in (os.curdir, os.pardir) else k for k in keys]
+
+    # Recombine the key into a relative path.
+    key = os.sep.join(keys)
+
+    return f"{str(key)}_{str(step)}_{str(id)}{extension}"
+
+
+class Media(WBValue):
+    """A WBValue stored as a file outside JSON that can be rendered in a media panel.
+
+    If necessary, we move or copy the file into the Run's media directory so that it
+    gets uploaded.
+    """
+
+    _path: Optional[str]
+    _run: Optional["LocalRun"]
+    _caption: Optional[str]
+    _is_tmp: Optional[bool]
+    _extension: Optional[str]
+    _sha256: Optional[str]
+    _size: Optional[int]
+
+    def __init__(self, caption: Optional[str] = None) -> None:
+        super().__init__()
+        self._path = None
+        # The run under which this object is bound, if any.
+        self._run = None
+        self._caption = caption
+
+    def _set_file(
+        self, path: str, is_tmp: bool = False, extension: Optional[str] = None
+    ) -> None:
+        self._path = path
+        self._is_tmp = is_tmp
+        self._extension = extension
+        assert extension is None or path.endswith(
+            extension
+        ), f'Media file extension "{extension}" must occur at the end of path "{path}".'
+
+        with open(self._path, "rb") as f:
+            self._sha256 = hashlib.sha256(f.read()).hexdigest()
+        self._size = os.path.getsize(self._path)
+
+    @classmethod
+    def get_media_subdir(cls: Type["Media"]) -> str:
+        raise NotImplementedError
+
+    @staticmethod
+    def captions(
+        media_items: Sequence["Media"],
+    ) -> Union[bool, Sequence[Optional[str]]]:
+        if media_items[0]._caption is not None:
+            return [m._caption for m in media_items]
+        else:
+            return False
+
+    def is_bound(self) -> bool:
+        return self._run is not None
+
+    def file_is_set(self) -> bool:
+        return self._path is not None and self._sha256 is not None
+
+    def bind_to_run(
+        self,
+        run: "LocalRun",
+        key: Union[int, str],
+        step: Union[int, str],
+        id_: Optional[Union[int, str]] = None,
+        ignore_copy_err: Optional[bool] = None,
+    ) -> None:
+        """Bind this object to a particular Run.
+
+        Calling this function is necessary so that we have somewhere specific to put the
+        file associated with this object, from which other Runs can refer to it.
+        """
+        assert self.file_is_set(), "bind_to_run called before _set_file"
+
+        # The following two assertions are guaranteed to pass
+        # by definition file_is_set, but are needed for
+        # mypy to understand that these are strings below.
+        assert isinstance(self._path, str)
+        assert isinstance(self._sha256, str)
+
+        assert run is not None, 'Argument "run" must not be None.'
+        self._run = run
+
+        if self._extension is None:
+            _, extension = os.path.splitext(os.path.basename(self._path))
+        else:
+            extension = self._extension
+
+        if id_ is None:
+            id_ = self._sha256[:20]
+
+        file_path = _wb_filename(key, step, id_, extension)
+        media_path = os.path.join(self.get_media_subdir(), file_path)
+        new_path = os.path.join(self._run.dir, media_path)
+        filesystem.mkdir_exists_ok(os.path.dirname(new_path))
+
+        if self._is_tmp:
+            shutil.move(self._path, new_path)
+            self._path = new_path
+            self._is_tmp = False
+            _datatypes_callback(media_path)
+        else:
+            try:
+                shutil.copy(self._path, new_path)
+            except shutil.SameFileError as e:
+                if not ignore_copy_err:
+                    raise e
+            self._path = new_path
+            _datatypes_callback(media_path)
+
+    def to_json(self, run: Union["LocalRun", "Artifact"]) -> dict:
+        """Serialize the object into a JSON blob.
+
+        Uses run or artifact to store additional data. If `run_or_artifact` is a
+        wandb.Run then `self.bind_to_run()` must have been previously been called.
+
+        Args:
+            run_or_artifact (wandb.Run | wandb.Artifact): the Run or Artifact for which
+                this object should be generating JSON for - this is useful to store
+                additional data if needed.
+
+        Returns:
+            dict: JSON representation
+        """
+        # NOTE: uses of Audio in this class are a temporary hack -- when Ref support moves up
+        # into Media itself we should get rid of them
+        from wandb import Image
+        from wandb.data_types import Audio
+        from wandb.sdk.wandb_run import Run
+
+        json_obj = {}
+
+        if isinstance(run, Run):
+            json_obj.update(
+                {
+                    "_type": "file",  # TODO(adrian): This isn't (yet) a real media type we support on the frontend.
+                    "sha256": self._sha256,
+                    "size": self._size,
+                }
+            )
+            artifact_entry_url = self._get_artifact_entry_ref_url()
+            if artifact_entry_url is not None:
+                json_obj["artifact_path"] = artifact_entry_url
+            artifact_entry_latest_url = self._get_artifact_entry_latest_ref_url()
+            if artifact_entry_latest_url is not None:
+                json_obj["_latest_artifact_path"] = artifact_entry_latest_url
+
+            if artifact_entry_url is None or self.is_bound():
+                assert self.is_bound(), "Value of type {} must be bound to a run with bind_to_run() before being serialized to JSON.".format(
+                    type(self).__name__
+                )
+
+                assert (
+                    self._run is run
+                ), "We don't support referring to media files across runs."
+
+                # The following two assertions are guaranteed to pass
+                # by definition is_bound, but are needed for
+                # mypy to understand that these are strings below.
+                assert isinstance(self._path, str)
+                json_obj["path"] = LogicalPath(
+                    os.path.relpath(self._path, self._run.dir)
+                )
+
+        elif isinstance(run, wandb.Artifact):
+            if self.file_is_set():
+                # The following two assertions are guaranteed to pass
+                # by definition of the call above, but are needed for
+                # mypy to understand that these are strings below.
+                assert isinstance(self._path, str)
+                assert isinstance(self._sha256, str)
+                artifact = run  # Checks if the concrete image has already been added to this artifact
+                name = artifact.get_added_local_path_name(self._path)
+                if name is None:
+                    if self._is_tmp:
+                        name = os.path.join(
+                            self.get_media_subdir(), os.path.basename(self._path)
+                        )
+                    else:
+                        # If the files is not temporary, include the first 8 characters of the file's SHA256 to
+                        # avoid name collisions. This way, if there are two images `dir1/img.png` and `dir2/img.png`
+                        # we end up with a unique path for each.
+                        name = os.path.join(
+                            self.get_media_subdir(),
+                            self._sha256[:20],
+                            os.path.basename(self._path),
+                        )
+
+                    # if not, check to see if there is a source artifact for this object
+                    if (
+                        self._artifact_source is not None
+                        # and self._artifact_source.artifact != artifact
+                    ):
+                        default_root = self._artifact_source.artifact._default_root()
+                        # if there is, get the name of the entry (this might make sense to move to a helper off artifact)
+                        if self._path.startswith(default_root):
+                            name = self._path[len(default_root) :]
+                            name = name.lstrip(os.sep)
+
+                        # Add this image as a reference
+                        path = self._artifact_source.artifact.get_entry(name)
+                        artifact.add_reference(path.ref_url(), name=name)
+                    elif (
+                        isinstance(self, Audio) or isinstance(self, Image)
+                    ) and self.path_is_reference(self._path):
+                        artifact.add_reference(self._path, name=name)
+                    else:
+                        entry = artifact.add_file(
+                            self._path, name=name, is_tmp=self._is_tmp
+                        )
+                        name = entry.path
+
+                json_obj["path"] = name
+                json_obj["sha256"] = self._sha256
+            json_obj["_type"] = self._log_type
+        return json_obj
+
+    @classmethod
+    def from_json(
+        cls: Type["Media"], json_obj: dict, source_artifact: "Artifact"
+    ) -> "Media":
+        """Likely will need to override for any more complicated media objects."""
+        return cls(source_artifact.get_entry(json_obj["path"]).download())
+
+    def __eq__(self, other: object) -> bool:
+        """Likely will need to override for any more complicated media objects."""
+        return (
+            isinstance(other, self.__class__)
+            and hasattr(self, "_sha256")
+            and hasattr(other, "_sha256")
+            and self._sha256 == other._sha256
+        )
+
+    @staticmethod
+    def path_is_reference(path: Optional[str]) -> bool:
+        return bool(path and re.match(r"^(gs|s3|https?)://", path))
+
+
+class BatchableMedia(Media):
+    """Media that is treated in batches.
+
+    E.g. images and thumbnails. Apart from images, we just use these batches to help
+    organize files by name in the media directory.
+    """
+
+    def __init__(self) -> None:
+        super().__init__()
+
+    @classmethod
+    def seq_to_json(
+        cls: Type["BatchableMedia"],
+        seq: Sequence["BatchableMedia"],
+        run: "LocalRun",
+        key: str,
+        step: Union[int, str],
+    ) -> dict:
+        raise NotImplementedError
+
+
+def _numpy_arrays_to_lists(
+    payload: Union[dict, Sequence, "np.ndarray"],
+) -> Union[Sequence, dict, str, int, float, bool]:
+    # Casts all numpy arrays to lists so we don't convert them to histograms, primarily for Plotly
+
+    if isinstance(payload, dict):
+        res = {}
+        for key, val in payload.items():
+            res[key] = _numpy_arrays_to_lists(val)
+        return res
+    elif isinstance(payload, Sequence) and not isinstance(payload, str):
+        return [_numpy_arrays_to_lists(v) for v in payload]
+    elif util.is_numpy_array(payload):
+        if TYPE_CHECKING:
+            payload = cast("np.ndarray", payload)
+        return [
+            _numpy_arrays_to_lists(v)
+            for v in (payload.tolist() if payload.ndim > 0 else [payload.tolist()])
+        ]
+    # Protects against logging non serializable objects
+    elif isinstance(payload, Media):
+        return str(payload.__class__.__name__)
+    return payload  # type: ignore