PyTorch Lightning

PyTorch Lightning provides a lightweight wrapper for organizing your PyTorch code and easily adding advanced features such as distributed training and 16-bit precision. W&B provides a lightweight wrapper for logging your ML experiments. But you don’t need to combine the two yourself: Weights & Biases is incorporated directly into the PyTorch Lightning library via the WandbLogger.

Integrate with Lightning

• PyTorch Logger
• Fabric Logger

from lightning.pytorch.loggers import WandbLogger
from lightning.pytorch import Trainer

wandb_logger = WandbLogger(log_model="all")
trainer = Trainer(logger=wandb_logger)

wandb.log({"accuracy":0.99, "trainer/global_step": step})

import lightning as L
from wandb.integration.lightning.fabric import WandbLogger

wandb_logger = WandbLogger(log_model="all")
fabric = L.Fabric(loggers=[wandb_logger])
fabric.launch()
fabric.log_dict({"important_metric": important_metric})

Sign up and create an API key

An API key authenticates your machine to W&B. You can generate an API key from your user profile.

1. Click your user profile icon in the upper right corner.
2. Select User Settings, then scroll to the API Keys section.
3. Click Reveal. Copy the displayed API key. To hide the API key, reload the page.

Install the wandb library and log in

To install the wandb library locally and log in:

• Command Line
• Python
• Python notebook

pip install wandb

import wandb
wandb.login()

!pip install wandb

import wandb
wandb.login()

Use PyTorch Lightning’s WandbLogger

PyTorch Lightning has multiple WandbLogger classes to log metrics and model weights, media, and more.

• PyTorch
• Fabric

To integrate with Lightning, instantiate the WandbLogger and pass it to Lightning’s Trainer or Fabric.

• PyTorch Logger
• Fabric Logger

trainer = Trainer(logger=wandb_logger)

fabric = L.Fabric(loggers=[wandb_logger])
fabric.launch()
fabric.log_dict({
    "important_metric": important_metric
})

Common logger arguments

Below are some of the most used parameters in WandbLogger. Review the PyTorch Lightning documentation for details about all logger arguments.

• PyTorch
• Fabric

Log your hyperparameters

• PyTorch Logger
• Fabric Logger

class LitModule(LightningModule):
    def __init__(self, *args, **kwarg):
        self.save_hyperparameters()

wandb_logger.log_hyperparams(
    {
        "hyperparameter_1": hyperparameter_1,
        "hyperparameter_2": hyperparameter_2,
    }
)

Log additional config parameters

# 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, parameter histogram and model topology

You can pass your model object to wandblogger.watch() to monitor your models’s gradients and parameters as you train. See the PyTorch Lightning WandbLogger documentation

Log metrics

• PyTorch Logger
• Fabric Logger

import torch
from torch.nn import Linear, CrossEntropyLoss, functional as F
from torch.optim import Adam
from torchmetrics.functional import accuracy
from lightning.pytorch import LightningModule


class My_LitModule(LightningModule):
    def __init__(self, n_classes=10, n_layer_1=128, n_layer_2=256, lr=1e-3):
        """method used to define the model parameters"""
        super().__init__()

        # mnist images are (1, 28, 28) (channels, width, height)
        self.layer_1 = Linear(28 * 28, n_layer_1)
        self.layer_2 = Linear(n_layer_1, n_layer_2)
        self.layer_3 = Linear(n_layer_2, n_classes)

        self.loss = CrossEntropyLoss()
        self.lr = lr

        # save hyper-parameters to self.hparams (auto-logged by W&B)
        self.save_hyperparameters()

    def forward(self, x):
        """method used for inference input -> output"""

        # (b, 1, 28, 28) -> (b, 1*28*28)
        batch_size, channels, width, height = x.size()
        x = x.view(batch_size, -1)

        # let's do 3 x (linear + relu)
        x = F.relu(self.layer_1(x))
        x = F.relu(self.layer_2(x))
        x = self.layer_3(x)
        return x

    def training_step(self, batch, batch_idx):
        """needs to return a loss from a single batch"""
        _, loss, acc = self._get_preds_loss_accuracy(batch)

        # Log loss and metric
        self.log("train_loss", loss)
        self.log("train_accuracy", acc)
        return loss

    def validation_step(self, batch, batch_idx):
        """used for logging metrics"""
        preds, loss, acc = self._get_preds_loss_accuracy(batch)

        # Log loss and metric
        self.log("val_loss", loss)
        self.log("val_accuracy", acc)
        return preds

    def configure_optimizers(self):
        """defines model optimizer"""
        return Adam(self.parameters(), lr=self.lr)

    def _get_preds_loss_accuracy(self, batch):
        """convenience function since train/valid/test steps are similar"""
        x, y = batch
        logits = self(x)
        preds = torch.argmax(logits, dim=1)
        loss = self.loss(logits, y)
        acc = accuracy(preds, y)
        return preds, loss, acc

import lightning as L
import torch
import torchvision as tv
from wandb.integration.lightning.fabric import WandbLogger
import wandb

fabric = L.Fabric(loggers=[wandb_logger])
fabric.launch()

model = tv.models.resnet18()
optimizer = torch.optim.SGD(model.parameters(), lr=lr)
model, optimizer = fabric.setup(model, optimizer)

train_dataloader = fabric.setup_dataloaders(
    torch.utils.data.DataLoader(train_dataset, batch_size=batch_size)
)

model.train()
for epoch in range(num_epochs):
    for batch in train_dataloader:
        optimizer.zero_grad()
        loss = model(batch)
        loss.backward()
        optimizer.step()
        fabric.log_dict({"loss": loss})

Log the min/max of a metric

Using wandb’s define_metric function you can define whether you’d like your W&B summary metric to display the min, max, mean or best value for that metric. If define_metric _ isn’t used, then the last value logged with appear in your summary metrics. See the define_metric reference docs here and the guide here for more.

To tell W&B to keep track of the max validation accuracy in the W&B summary metric, call wandb.define_metric only once, at the beginning of training:

• PyTorch Logger
• Fabric Logger

class My_LitModule(LightningModule):
    ...

    def validation_step(self, batch, batch_idx):
        if trainer.global_step == 0:
            wandb.define_metric("val_accuracy", summary="max")

        preds, loss, acc = self._get_preds_loss_accuracy(batch)

        # Log loss and metric
        self.log("val_loss", loss)
        self.log("val_accuracy", acc)
        return preds

wandb.define_metric("val_accuracy", summary="max")
fabric = L.Fabric(loggers=[wandb_logger])
fabric.launch()
fabric.log_dict({"val_accuracy": val_accuracy})

Checkpoint a model

To save model checkpoints as W&B Artifacts, use the Lightning ModelCheckpoint callback and set the log_model argument in the WandbLogger.

• PyTorch Logger
• Fabric Logger

trainer = Trainer(logger=wandb_logger, callbacks=[checkpoint_callback])

fabric = L.Fabric(loggers=[wandb_logger], callbacks=[checkpoint_callback])

The latest and best aliases are automatically set to easily retrieve a model checkpoint from a W&B Artifact:

# 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"

• Via Logger
• Via wandb

# download checkpoint locally (if not already cached)
wandb_logger.download_artifact(checkpoint_reference, artifact_type="model")

# download checkpoint locally (if not already cached)
run = wandb.init(project="MNIST")
artifact = run.use_artifact(checkpoint_reference, type="model")
artifact_dir = artifact.download()

• PyTorch Logger
• Fabric Logger

# load checkpoint
model = LitModule.load_from_checkpoint(Path(artifact_dir) / "model.ckpt")

# Request the raw checkpoint
full_checkpoint = fabric.load(Path(artifact_dir) / "model.ckpt")

model.load_state_dict(full_checkpoint["model"])
optimizer.load_state_dict(full_checkpoint["optimizer"])

The model checkpoints you log are viewable through the W&B Artifacts UI, and include the full model lineage (see an example model checkpoint in the UI here).

To bookmark your best model checkpoints and centralize them across your team, you can link them to the W&B Model Registry.

Here you can organize your best models by task, manage model lifecycle, facilitate easy tracking and auditing throughout the ML lifecyle, and automate downstream actions with webhooks or jobs.

Log images, text, and more

The WandbLogger has log_image, log_text and log_table methods for logging media.

You can also directly call wandb.log or trainer.logger.experiment.log to log other media types such as Audio, Molecules, Point Clouds, 3D Objects and more.

• Log Images
• Log Text
• Log Tables

# 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"])

# using .log in the trainer
trainer.logger.experiment.log(
    {"samples": [wandb.Image(img, caption=caption) for (img, caption) in my_images]},
    step=current_trainer_global_step,
)

# data should be a list of lists
columns = ["input", "label", "prediction"]
my_data = [["cheese", "english", "english"], ["fromage", "french", "spanish"]]

# using columns and data
wandb_logger.log_text(key="my_samples", columns=columns, data=my_data)

# using a pandas DataFrame
wandb_logger.log_text(key="my_samples", dataframe=my_dataframe)

# log a W&B Table that has a text caption, an image and audio
columns = ["caption", "image", "sound"]

# data should be a list of lists
my_data = [
    ["cheese", wandb.Image(img_1), wandb.Audio(snd_1)],
    ["wine", wandb.Image(img_2), wandb.Audio(snd_2)],
]

# log the Table
wandb_logger.log_table(key="my_samples", columns=columns, data=data)

You can use Lightning’s Callbacks system to control when you log to Weights & Biases via the WandbLogger, in this example we log a sample of our validation images and predictions:

import torch
import wandb
import lightning.pytorch as pl
from lightning.pytorch.loggers import WandbLogger

# or
# from wandb.integration.lightning.fabric import WandbLogger


class LogPredictionSamplesCallback(Callback):
    def on_validation_batch_end(
        self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx
    ):
        """Called when the validation batch ends."""

        # `outputs` comes from `LightningModule.validation_step`
        # which corresponds to our model predictions in this case

        # Let's log 20 sample image predictions from the first batch
        if batch_idx == 0:
            n = 20
            x, y = batch
            images = [img for img in x[:n]]
            captions = [
                f"Ground Truth: {y_i} - Prediction: {y_pred}"
                for y_i, y_pred in zip(y[:n], outputs[:n])
            ]

            # Option 1: log images with `WandbLogger.log_image`
            wandb_logger.log_image(key="sample_images", images=images, caption=captions)

            # Option 2: log images and predictions as a W&B Table
            columns = ["image", "ground truth", "prediction"]
            data = [
                [wandb.Image(x_i), y_i, y_pred] or x_i,
                y_i,
                y_pred in list(zip(x[:n], y[:n], outputs[:n])),
            ]
            wandb_logger.log_table(key="sample_table", columns=columns, data=data)


trainer = pl.Trainer(callbacks=[LogPredictionSamplesCallback()])

Use multiple GPUs with Lightning and W&B

PyTorch Lightning has Multi-GPU support through their DDP Interface. However, PyTorch Lightning’s design requires you to be careful about how you instantiate our GPUs.

Lightning assumes that each GPU (or Rank) in your training loop must be instantiated in exactly the same way - with the same initial conditions. However, only rank 0 process gets access to the wandb.run object, and for non-zero rank processes: wandb.run = None. This could cause your non-zero processes to fail. Such a situation can put you in a deadlock because rank 0 process will wait for the non-zero rank processes to join, which have already crashed.

For this reason, be careful about how we set up your training code. The recommended way to set it up would be to have your code be independent of the wandb.run object.

class MNISTClassifier(pl.LightningModule):
    def __init__(self):
        super(MNISTClassifier, self).__init__()

        self.model = nn.Sequential(
            nn.Flatten(),
            nn.Linear(28 * 28, 128),
            nn.ReLU(),
            nn.Linear(128, 10),
        )

        self.loss = nn.CrossEntropyLoss()

    def forward(self, x):
        return self.model(x)

    def training_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self.forward(x)
        loss = self.loss(y_hat, y)

        self.log("train/loss", loss)
        return {"train_loss": loss}

    def validation_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self.forward(x)
        loss = self.loss(y_hat, y)

        self.log("val/loss", loss)
        return {"val_loss": loss}

    def configure_optimizers(self):
        return torch.optim.Adam(self.parameters(), lr=0.001)


def main():
    # Setting all the random seeds to the same value.
    # This is important in a distributed training setting.
    # Each rank will get its own set of initial weights.
    # If they don't match up, the gradients will not match either,
    # leading to training that may not converge.
    pl.seed_everything(1)

    train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True, num_workers=4)
    val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False, num_workers=4)

    model = MNISTClassifier()
    wandb_logger = WandbLogger(project="<project_name>")
    callbacks = [
        ModelCheckpoint(
            dirpath="checkpoints",
            every_n_train_steps=100,
        ),
    ]
    trainer = pl.Trainer(
        max_epochs=3, gpus=2, logger=wandb_logger, strategy="ddp", callbacks=callbacks
    )
    trainer.fit(model, train_loader, val_loader)

Examples

You can follow along in a video tutorial with a Colab here.

Frequently Asked Questions

How does W&B integrate with Lightning?

The core integration is based on the Lightning loggers API, which lets you write much of your logging code in a framework-agnostic way. Loggers are passed to the Lightning Trainer and are triggered based on that API’s rich hook-and-callback system. This keeps your research code well-separated from engineering and logging code.

What does the integration log without any additional code?

We’ll save your model checkpoints to W&B, where you can view them or download them for use in future runs. We’ll also capture system metrics, like GPU usage and network I/O, environment information, like hardware and OS information, code state (including git commit and diff patch, notebook contents and session history), and anything printed to the standard out.

What if I need to use wandb.run in my training setup?

You need to expand the scope of the variable you need to access yourself. In other words, make sure that the initial conditions are the same on all processes.

if os.environ.get("LOCAL_RANK", None) is None:
    os.environ["WANDB_DIR"] = wandb.run.dir

If they are, you can use os.environ["WANDB_DIR"] to set up the model checkpoints directory. This way, any non-zero rank process can access wandb.run.dir.