import os from argparse import Namespace from pathlib import Path from typing import TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional, Union from packaging import version from typing_extensions import override import wandb from wandb import Artifact from wandb.sdk.lib import telemetry from wandb.sdk.wandb_run import Run try: import lightning import torch.nn as nn from lightning.fabric.loggers.logger import Logger, rank_zero_experiment from lightning.fabric.utilities.exceptions import MisconfigurationException from lightning.fabric.utilities.logger import ( _add_prefix, _convert_params, _sanitize_callable_params, ) from lightning.fabric.utilities.rank_zero import rank_zero_only, rank_zero_warn from lightning.fabric.utilities.types import _PATH from torch import Tensor from torch.nn import Module if version.parse(lightning.__version__) > version.parse("2.1.3"): wandb.termwarn( """This integration is tested and supported for lightning Fabric 2.1.3. Please report any issues to https://github.com/wandb/wandb/issues with the tag `lightning-fabric`.""", repeat=False, ) if TYPE_CHECKING: from lightning.pytorch.callbacks.model_checkpoint import ModelCheckpoint except ImportError as e: wandb.Error(e) class WandbLogger(Logger): r"""Log using `Weights and Biases `_. **Installation and set-up** Install with pip: .. code-block:: bash pip install wandb Create a `WandbLogger` instance: .. code-block:: python from lightning.fabric.loggers import WandbLogger wandb_logger = WandbLogger(project="MNIST") Pass the logger instance to the `Trainer`: .. code-block:: python trainer = Trainer(logger=wandb_logger) A new W&B run will be created when training starts if you have not created one manually before with `wandb.init()`. **Log metrics** Log from :class:`~lightning.pytorch.core.LightningModule`: .. code-block:: python class LitModule(LightningModule): def training_step(self, batch, batch_idx): self.log("train/loss", loss) Use directly wandb module: .. code-block:: python wandb.log({"train/loss": loss}) **Log hyper-parameters** Save :class:`~lightning.pytorch.core.LightningModule` parameters: .. code-block:: python class LitModule(LightningModule): def __init__(self, *args, **kwarg): self.save_hyperparameters() Add other config parameters: .. code-block:: python # add one parameter wandb_logger.experiment.config["key"] = value # add multiple parameters wandb_logger.experiment.config.update({key1: val1, key2: val2}) # use directly wandb module wandb.config["key"] = value wandb.config.update() **Log gradients, parameters and model topology** Call the `watch` method for automatically tracking gradients: .. code-block:: python # log gradients and model topology wandb_logger.watch(model) # log gradients, parameter histogram and model topology wandb_logger.watch(model, log="all") # change log frequency of gradients and parameters (100 steps by default) wandb_logger.watch(model, log_freq=500) # do not log graph (in case of errors) wandb_logger.watch(model, log_graph=False) The `watch` method adds hooks to the model which can be removed at the end of training: .. code-block:: python wandb_logger.experiment.unwatch(model) **Log model checkpoints** Log model checkpoints at the end of training: .. code-block:: python wandb_logger = WandbLogger(log_model=True) Log model checkpoints as they get created during training: .. code-block:: python wandb_logger = WandbLogger(log_model="all") Custom checkpointing can be set up through :class:`~lightning.pytorch.callbacks.ModelCheckpoint`: .. code-block:: python # log model only if `val_accuracy` increases wandb_logger = WandbLogger(log_model="all") checkpoint_callback = ModelCheckpoint(monitor="val_accuracy", mode="max") trainer = Trainer(logger=wandb_logger, callbacks=[checkpoint_callback]) `latest` and `best` aliases are automatically set to easily retrieve a model checkpoint: .. code-block:: python # reference can be retrieved in artifacts panel # "VERSION" can be a version (ex: "v2") or an alias ("latest or "best") checkpoint_reference = "USER/PROJECT/MODEL-RUN_ID:VERSION" # download checkpoint locally (if not already cached) run = wandb.init(project="MNIST") artifact = run.use_artifact(checkpoint_reference, type="model") artifact_dir = artifact.download() # load checkpoint model = LitModule.load_from_checkpoint(Path(artifact_dir) / "model.ckpt") **Log media** Log text with: .. code-block:: python # using columns and data columns = ["input", "label", "prediction"] data = [["cheese", "english", "english"], ["fromage", "french", "spanish"]] wandb_logger.log_text(key="samples", columns=columns, data=data) # using a pandas DataFrame wandb_logger.log_text(key="samples", dataframe=my_dataframe) Log images with: .. code-block:: python # using tensors, numpy arrays or PIL images wandb_logger.log_image(key="samples", images=[img1, img2]) # adding captions wandb_logger.log_image( key="samples", images=[img1, img2], caption=["tree", "person"] ) # using file path wandb_logger.log_image(key="samples", images=["img_1.jpg", "img_2.jpg"]) More arguments can be passed for logging segmentation masks and bounding boxes. Refer to `Image Overlays documentation `_. **Log Tables** `W&B Tables `_ can be used to log, query and analyze tabular data. They support any type of media (text, image, video, audio, molecule, html, etc) and are great for storing, understanding and sharing any form of data, from datasets to model predictions. .. code-block:: python columns = ["caption", "image", "sound"] data = [ ["cheese", wandb.Image(img_1), wandb.Audio(snd_1)], ["wine", wandb.Image(img_2), wandb.Audio(snd_2)], ] wandb_logger.log_table(key="samples", columns=columns, data=data) **Downloading and Using Artifacts** To download an artifact without starting a run, call the ``download_artifact`` function on the class: .. code-block:: python artifact_dir = wandb_logger.download_artifact(artifact="path/to/artifact") To download an artifact and link it to an ongoing run call the ``download_artifact`` function on the logger instance: .. code-block:: python class MyModule(LightningModule): def any_lightning_module_function_or_hook(self): self.logger.download_artifact(artifact="path/to/artifact") To link an artifact from a previous run you can use ``use_artifact`` function: .. code-block:: python wandb_logger.use_artifact(artifact="path/to/artifact") See Also: - `Demo in Google Colab `__ with hyperparameter search and model logging - `W&B Documentation `__ Args: name: Display name for the run. save_dir: Path where data is saved. version: Sets the version, mainly used to resume a previous run. offline: Run offline (data can be streamed later to wandb servers). dir: Same as save_dir. id: Same as version. anonymous: Enables or explicitly disables anonymous logging. project: The name of the project to which this run will belong. If not set, the environment variable `WANDB_PROJECT` will be used as a fallback. If both are not set, it defaults to ``'lightning_logs'``. log_model: Log checkpoints created by :class:`~lightning.pytorch.callbacks.ModelCheckpoint` as W&B artifacts. `latest` and `best` aliases are automatically set. * if ``log_model == 'all'``, checkpoints are logged during training. * if ``log_model == True``, checkpoints are logged at the end of training, except when `~lightning.pytorch.callbacks.ModelCheckpoint.save_top_k` ``== -1`` which also logs every checkpoint during training. * if ``log_model == False`` (default), no checkpoint is logged. prefix: A string to put at the beginning of metric keys. experiment: WandB experiment object. Automatically set when creating a run. checkpoint_name: Name of the model checkpoint artifact being logged. log_checkpoint_on: When to log model checkpoints as W&B artifacts. Only used if ``log_model`` is ``True``. Options: ``"success"``, ``"all"``. Default: ``"success"``. \**kwargs: Arguments passed to :func:`wandb.init` like `entity`, `group`, `tags`, etc. Raises: ModuleNotFoundError: If required WandB package is not installed on the device. MisconfigurationException: If both ``log_model`` and ``offline`` is set to ``True``. """ LOGGER_JOIN_CHAR = "-" def __init__( self, name: Optional[str] = None, save_dir: _PATH = ".", version: Optional[str] = None, offline: bool = False, dir: Optional[_PATH] = None, id: Optional[str] = None, anonymous: Optional[bool] = None, project: Optional[str] = None, log_model: Union[Literal["all"], bool] = False, experiment: Optional["Run"] = None, prefix: str = "", checkpoint_name: Optional[str] = None, log_checkpoint_on: Union[Literal["success"], Literal["all"]] = "success", **kwargs: Any, ) -> None: if offline and log_model: raise MisconfigurationException( f"Providing log_model={log_model} and offline={offline} is an invalid configuration" " since model checkpoints cannot be uploaded in offline mode.\n" "Hint: Set `offline=False` to log your model." ) super().__init__() self._offline = offline self._log_model = log_model self._prefix = prefix self._experiment = experiment self._logged_model_time: Dict[str, float] = {} self._checkpoint_callback: Optional[ModelCheckpoint] = None # paths are processed as strings if save_dir is not None: save_dir = os.fspath(save_dir) elif dir is not None: dir = os.fspath(dir) project = project or os.environ.get("WANDB_PROJECT", "lightning_fabric_logs") # set wandb init arguments self._wandb_init: Dict[str, Any] = { "name": name, "project": project, "dir": save_dir or dir, "id": version or id, "resume": "allow", "anonymous": ("allow" if anonymous else None), } self._wandb_init.update(**kwargs) # extract parameters self._project = self._wandb_init.get("project") self._save_dir = self._wandb_init.get("dir") self._name = self._wandb_init.get("name") self._id = self._wandb_init.get("id") self._checkpoint_name = checkpoint_name self._log_checkpoint_on = log_checkpoint_on def __getstate__(self) -> Dict[str, Any]: # Hack: If the 'spawn' launch method is used, the logger will get pickled and this `__getstate__` gets called. # We create an experiment here in the main process, and attach to it in the worker process. # Using wandb-service, we persist the same experiment even if multiple `Trainer.fit/test/validate` calls # are made. _ = self.experiment state = self.__dict__.copy() # args needed to reload correct experiment if self._experiment is not None: state["_id"] = getattr(self._experiment, "id", None) state["_attach_id"] = getattr(self._experiment, "_attach_id", None) state["_name"] = self._experiment.name # cannot be pickled state["_experiment"] = None return state @property @rank_zero_experiment def experiment(self) -> "Run": r"""Actual wandb object. To use wandb features in your :class:`~lightning.pytorch.core.LightningModule`, do the following. Example:: .. code-block:: python self.logger.experiment.some_wandb_function() """ if self._experiment is None: if self._offline: os.environ["WANDB_MODE"] = "dryrun" attach_id = getattr(self, "_attach_id", None) if wandb.run is not None: # wandb process already created in this instance rank_zero_warn( "There is a wandb run already in progress and newly created instances of `WandbLogger` will reuse" " this run. If this is not desired, call `wandb.finish()` before instantiating `WandbLogger`." ) self._experiment = wandb.run elif attach_id is not None and hasattr(wandb, "_attach"): # attach to wandb process referenced self._experiment = wandb._attach(attach_id) else: # create new wandb process self._experiment = wandb.init(**self._wandb_init) # define default x-axis if isinstance(self._experiment, Run) and getattr( self._experiment, "define_metric", None ): self._experiment.define_metric("trainer/global_step") self._experiment.define_metric( "*", step_metric="trainer/global_step", step_sync=True ) self._experiment._label(repo="lightning_fabric_logger") # pylint: disable=protected-access with telemetry.context(run=self._experiment) as tel: tel.feature.lightning_fabric_logger = True return self._experiment def watch( self, model: nn.Module, log: str = "gradients", log_freq: int = 100, log_graph: bool = True, ) -> None: self.experiment.watch(model, log=log, log_freq=log_freq, log_graph=log_graph) @override @rank_zero_only def log_hyperparams(self, params: Union[Dict[str, Any], Namespace]) -> None: # type: ignore[override] params = _convert_params(params) params = _sanitize_callable_params(params) self.experiment.config.update(params, allow_val_change=True) @override @rank_zero_only def log_metrics( self, metrics: Mapping[str, float], step: Optional[int] = None ) -> None: assert rank_zero_only.rank == 0, "experiment tried to log from global_rank != 0" metrics = _add_prefix(metrics, self._prefix, self.LOGGER_JOIN_CHAR) if step is not None: self.experiment.log(dict(metrics, **{"trainer/global_step": step})) else: self.experiment.log(metrics) @rank_zero_only def log_table( self, key: str, columns: Optional[List[str]] = None, data: Optional[List[List[Any]]] = None, dataframe: Any = None, step: Optional[int] = None, ) -> None: """Log a Table containing any object type (text, image, audio, video, molecule, html, etc). Can be defined either with `columns` and `data` or with `dataframe`. """ metrics = {key: wandb.Table(columns=columns, data=data, dataframe=dataframe)} self.log_metrics(metrics, step) @rank_zero_only def log_text( self, key: str, columns: Optional[List[str]] = None, data: Optional[List[List[str]]] = None, dataframe: Any = None, step: Optional[int] = None, ) -> None: """Log text as a Table. Can be defined either with `columns` and `data` or with `dataframe`. """ self.log_table(key, columns, data, dataframe, step) @rank_zero_only def log_html( self, key: str, htmls: List[Any], step: Optional[int] = None, **kwargs: Any ) -> None: """Log html files. Optional kwargs are lists passed to each html (ex: inject). """ if not isinstance(htmls, list): raise TypeError(f'Expected a list as "htmls", found {type(htmls)}') n = len(htmls) for k, v in kwargs.items(): if len(v) != n: raise ValueError(f"Expected {n} items but only found {len(v)} for {k}") kwarg_list = [{k: kwargs[k][i] for k in kwargs} for i in range(n)] metrics = { key: [wandb.Html(html, **kwarg) for html, kwarg in zip(htmls, kwarg_list)] } self.log_metrics(metrics, step) # type: ignore[arg-type] @rank_zero_only def log_image( self, key: str, images: List[Any], step: Optional[int] = None, **kwargs: Any ) -> None: """Log images (tensors, numpy arrays, PIL Images or file paths). Optional kwargs are lists passed to each image (ex: caption, masks, boxes). """ if not isinstance(images, list): raise TypeError(f'Expected a list as "images", found {type(images)}') n = len(images) for k, v in kwargs.items(): if len(v) != n: raise ValueError(f"Expected {n} items but only found {len(v)} for {k}") kwarg_list = [{k: kwargs[k][i] for k in kwargs} for i in range(n)] metrics = { key: [wandb.Image(img, **kwarg) for img, kwarg in zip(images, kwarg_list)] } self.log_metrics(metrics, step) # type: ignore[arg-type] @rank_zero_only def log_audio( self, key: str, audios: List[Any], step: Optional[int] = None, **kwargs: Any ) -> None: r"""Log audios (numpy arrays, or file paths). Args: key: The key to be used for logging the audio files audios: The list of audio file paths, or numpy arrays to be logged step: The step number to be used for logging the audio files \**kwargs: Optional kwargs are lists passed to each ``Wandb.Audio`` instance (ex: caption, sample_rate). Optional kwargs are lists passed to each audio (ex: caption, sample_rate). """ if not isinstance(audios, list): raise TypeError(f'Expected a list as "audios", found {type(audios)}') n = len(audios) for k, v in kwargs.items(): if len(v) != n: raise ValueError(f"Expected {n} items but only found {len(v)} for {k}") kwarg_list = [{k: kwargs[k][i] for k in kwargs} for i in range(n)] metrics = { key: [ wandb.Audio(audio, **kwarg) for audio, kwarg in zip(audios, kwarg_list) ] } self.log_metrics(metrics, step) # type: ignore[arg-type] @rank_zero_only def log_video( self, key: str, videos: List[Any], step: Optional[int] = None, **kwargs: Any ) -> None: """Log videos (numpy arrays, or file paths). Args: key: The key to be used for logging the video files videos: The list of video file paths, or numpy arrays to be logged step: The step number to be used for logging the video files **kwargs: Optional kwargs are lists passed to each Wandb.Video instance (ex: caption, fps, format). Optional kwargs are lists passed to each video (ex: caption, fps, format). """ if not isinstance(videos, list): raise TypeError(f'Expected a list as "videos", found {type(videos)}') n = len(videos) for k, v in kwargs.items(): if len(v) != n: raise ValueError(f"Expected {n} items but only found {len(v)} for {k}") kwarg_list = [{k: kwargs[k][i] for k in kwargs} for i in range(n)] metrics = { key: [ wandb.Video(video, **kwarg) for video, kwarg in zip(videos, kwarg_list) ] } self.log_metrics(metrics, step) # type: ignore[arg-type] @property @override def save_dir(self) -> Optional[str]: """Gets the save directory. Returns: The path to the save directory. """ return self._save_dir @property @override def name(self) -> Optional[str]: """The project name of this experiment. Returns: The name of the project the current experiment belongs to. This name is not the same as `wandb.Run`'s name. To access wandb's internal experiment name, use ``logger.experiment.name`` instead. """ return self._project @property @override def version(self) -> Optional[str]: """Gets the id of the experiment. Returns: The id of the experiment if the experiment exists else the id given to the constructor. """ # don't create an experiment if we don't have one return self._experiment.id if self._experiment else self._id @property def log_dir(self) -> Optional[str]: """Gets the save directory. Returns: The path to the save directory. """ return self.save_dir @property def group_separator(self) -> str: """Return the default separator used by the logger to group the data into subfolders.""" return self.LOGGER_JOIN_CHAR @property def root_dir(self) -> Optional[str]: """Return the root directory. Return the root directory where all versions of an experiment get saved, or `None` if the logger does not save data locally. """ return self.save_dir.parent if self.save_dir else None def log_graph(self, model: Module, input_array: Optional[Tensor] = None) -> None: """Record model graph. Args: model: the model with an implementation of ``forward``. input_array: input passes to `model.forward` This is a noop function and does not perform any operation. """ return @override def after_save_checkpoint(self, checkpoint_callback: "ModelCheckpoint") -> None: # log checkpoints as artifacts if ( self._log_model == "all" or self._log_model is True and checkpoint_callback.save_top_k == -1 ): # TODO: Replace with new Fabric Checkpoints system self._scan_and_log_pytorch_checkpoints(checkpoint_callback) elif self._log_model is True: self._checkpoint_callback = checkpoint_callback @staticmethod @rank_zero_only def download_artifact( artifact: str, save_dir: Optional[_PATH] = None, artifact_type: Optional[str] = None, use_artifact: Optional[bool] = True, ) -> str: """Downloads an artifact from the wandb server. Args: artifact: The path of the artifact to download. save_dir: The directory to save the artifact to. artifact_type: The type of artifact to download. use_artifact: Whether to add an edge between the artifact graph. Returns: The path to the downloaded artifact. """ if wandb.run is not None and use_artifact: artifact = wandb.run.use_artifact(artifact) else: api = wandb.Api() artifact = api.artifact(artifact, type=artifact_type) save_dir = None if save_dir is None else os.fspath(save_dir) return artifact.download(root=save_dir) def use_artifact( self, artifact: str, artifact_type: Optional[str] = None ) -> "Artifact": """Logs to the wandb dashboard that the mentioned artifact is used by the run. Args: artifact: The path of the artifact. artifact_type: The type of artifact being used. Returns: wandb Artifact object for the artifact. """ return self.experiment.use_artifact(artifact, type=artifact_type) @override @rank_zero_only def save(self) -> None: """Save log data.""" self.experiment.log({}, commit=True) @override @rank_zero_only def finalize(self, status: str) -> None: if self._log_checkpoint_on == "success" and status != "success": # Currently, checkpoints only get logged on success return # log checkpoints as artifacts if ( self._checkpoint_callback and self._experiment is not None and self._log_checkpoint_on in ["success", "all"] ): self._scan_and_log_pytorch_checkpoints(self._checkpoint_callback) def _scan_and_log_pytorch_checkpoints( self, checkpoint_callback: "ModelCheckpoint" ) -> None: from lightning.pytorch.loggers.utilities import _scan_checkpoints # get checkpoints to be saved with associated score checkpoints = _scan_checkpoints(checkpoint_callback, self._logged_model_time) # log iteratively all new checkpoints for t, p, s, _ in checkpoints: metadata = { "score": s.item() if isinstance(s, Tensor) else s, "original_filename": Path(p).name, checkpoint_callback.__class__.__name__: { k: getattr(checkpoint_callback, k) for k in [ "monitor", "mode", "save_last", "save_top_k", "save_weights_only", "_every_n_train_steps", ] # ensure it does not break if `ModelCheckpoint` args change if hasattr(checkpoint_callback, k) }, } if not self._checkpoint_name: self._checkpoint_name = f"model-{self.experiment.id}" artifact = wandb.Artifact( name=self._checkpoint_name, type="model", metadata=metadata ) artifact.add_file(p, name="model.ckpt") aliases = ( ["latest", "best"] if p == checkpoint_callback.best_model_path else ["latest"] ) self.experiment.log_model(artifact, aliases=aliases) # remember logged models - timestamp needed in case filename didn't change (lastkckpt or custom name) self._logged_model_time[p] = t