PyTorch Lightning

Setting up PyTorch Lightning and W&B

For this tutorial, we need PyTorch Lightning and Weights and Biases.

pip install lightning -q
pip install wandb -qU

import lightning.pytorch as pl

# your favorite machine learning tracking tool
from lightning.pytorch.loggers import WandbLogger

import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import random_split, DataLoader

from torchmetrics import Accuracy

from torchvision import transforms
from torchvision.datasets import CIFAR10

import wandb

Now you’ll need to log in to your wandb account.

wandb.login()

DataModule - The Data Pipeline we Deserve

DataModules are a way of decoupling data-related hooks from the LightningModule so you can develop dataset agnostic models.

It organizes the data pipeline into one shareable and reusable class. A datamodule encapsulates the five steps involved in data processing in PyTorch:

• Download / tokenize / process.
• Clean and (maybe) save to disk.
• Load inside Dataset.
• Apply transforms (rotate, tokenize, etc…).
• Wrap inside a DataLoader.

Learn more about datamodules here. Let’s build a datamodule for the Cifar-10 dataset.

class CIFAR10DataModule(pl.LightningDataModule):
    def __init__(self, batch_size, data_dir: str = './'):
        super().__init__()
        self.data_dir = data_dir
        self.batch_size = batch_size

        self.transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
        ])
        
        self.num_classes = 10
    
    def prepare_data(self):
        CIFAR10(self.data_dir, train=True, download=True)
        CIFAR10(self.data_dir, train=False, download=True)
    
    def setup(self, stage=None):
        # Assign train/val datasets for use in dataloaders
        if stage == 'fit' or stage is None:
            cifar_full = CIFAR10(self.data_dir, train=True, transform=self.transform)
            self.cifar_train, self.cifar_val = random_split(cifar_full, [45000, 5000])

        # Assign test dataset for use in dataloader(s)
        if stage == 'test' or stage is None:
            self.cifar_test = CIFAR10(self.data_dir, train=False, transform=self.transform)
    
    def train_dataloader(self):
        return DataLoader(self.cifar_train, batch_size=self.batch_size, shuffle=True)

    def val_dataloader(self):
        return DataLoader(self.cifar_val, batch_size=self.batch_size)

    def test_dataloader(self):
        return DataLoader(self.cifar_test, batch_size=self.batch_size)

Callbacks

A callback is a self-contained program that can be reused across projects. PyTorch Lightning comes with few built-in callbacks which are regularly used. Learn more about callbacks in PyTorch Lightning here.

Built-in Callbacks

In this tutorial, we will use Early Stopping and Model Checkpoint built-in callbacks. They can be passed to the Trainer.

Custom Callbacks

If you are familiar with Custom Keras callback, the ability to do the same in your PyTorch pipeline is just a cherry on the cake.

Since we are performing image classification, the ability to visualize the model’s predictions on some samples of images can be helpful. This in the form of a callback can help debug the model at an early stage.

class ImagePredictionLogger(pl.callbacks.Callback):
    def __init__(self, val_samples, num_samples=32):
        super().__init__()
        self.num_samples = num_samples
        self.val_imgs, self.val_labels = val_samples
    
    def on_validation_epoch_end(self, trainer, pl_module):
        # Bring the tensors to CPU
        val_imgs = self.val_imgs.to(device=pl_module.device)
        val_labels = self.val_labels.to(device=pl_module.device)
        # Get model prediction
        logits = pl_module(val_imgs)
        preds = torch.argmax(logits, -1)
        # Log the images as wandb Image
        trainer.logger.experiment.log({
            "examples":[wandb.Image(x, caption=f"Pred:{pred}, Label:{y}") 
                           for x, pred, y in zip(val_imgs[:self.num_samples], 
                                                 preds[:self.num_samples], 
                                                 val_labels[:self.num_samples])]
            })
        

LightningModule - Define the System

The LightningModule defines a system and not a model. Here a system groups all the research code into a single class to make it self-contained. LightningModule organizes your PyTorch code into 5 sections:

• Computations (__init__).
• Train loop (training_step)
• Validation loop (validation_step)
• Test loop (test_step)
• Optimizers (configure_optimizers)

One can thus build a dataset agnostic model that can be easily shared. Let’s build a system for Cifar-10 classification.

class LitModel(pl.LightningModule):
    def __init__(self, input_shape, num_classes, learning_rate=2e-4):
        super().__init__()
        
        # log hyperparameters
        self.save_hyperparameters()
        self.learning_rate = learning_rate
        
        self.conv1 = nn.Conv2d(3, 32, 3, 1)
        self.conv2 = nn.Conv2d(32, 32, 3, 1)
        self.conv3 = nn.Conv2d(32, 64, 3, 1)
        self.conv4 = nn.Conv2d(64, 64, 3, 1)

        self.pool1 = torch.nn.MaxPool2d(2)
        self.pool2 = torch.nn.MaxPool2d(2)
        
        n_sizes = self._get_conv_output(input_shape)

        self.fc1 = nn.Linear(n_sizes, 512)
        self.fc2 = nn.Linear(512, 128)
        self.fc3 = nn.Linear(128, num_classes)

        self.accuracy = Accuracy(task='multiclass', num_classes=num_classes)

    # returns the size of the output tensor going into Linear layer from the conv block.
    def _get_conv_output(self, shape):
        batch_size = 1
        input = torch.autograd.Variable(torch.rand(batch_size, *shape))

        output_feat = self._forward_features(input) 
        n_size = output_feat.data.view(batch_size, -1).size(1)
        return n_size
        
    # returns the feature tensor from the conv block
    def _forward_features(self, x):
        x = F.relu(self.conv1(x))
        x = self.pool1(F.relu(self.conv2(x)))
        x = F.relu(self.conv3(x))
        x = self.pool2(F.relu(self.conv4(x)))
        return x
    
    # will be used during inference
    def forward(self, x):
       x = self._forward_features(x)
       x = x.view(x.size(0), -1)
       x = F.relu(self.fc1(x))
       x = F.relu(self.fc2(x))
       x = F.log_softmax(self.fc3(x), dim=1)
       
       return x
    
    def training_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)
        
        # training metrics
        preds = torch.argmax(logits, dim=1)
        acc = self.accuracy(preds, y)
        self.log('train_loss', loss, on_step=True, on_epoch=True, logger=True)
        self.log('train_acc', acc, on_step=True, on_epoch=True, logger=True)
        
        return loss
    
    def validation_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)

        # validation metrics
        preds = torch.argmax(logits, dim=1)
        acc = self.accuracy(preds, y)
        self.log('val_loss', loss, prog_bar=True)
        self.log('val_acc', acc, prog_bar=True)
        return loss
    
    def test_step(self, batch, batch_idx):
        x, y = batch
        logits = self(x)
        loss = F.nll_loss(logits, y)
        
        # validation metrics
        preds = torch.argmax(logits, dim=1)
        acc = self.accuracy(preds, y)
        self.log('test_loss', loss, prog_bar=True)
        self.log('test_acc', acc, prog_bar=True)
        return loss
    
    def configure_optimizers(self):
        optimizer = torch.optim.Adam(self.parameters(), lr=self.learning_rate)
        return optimizer

Train and Evaluate

Now that we have organized our data pipeline using DataModule and model architecture+training loop using LightningModule, the PyTorch Lightning Trainer automates everything else for us.

The Trainer automates:

• Epoch and batch iteration
• Calling of optimizer.step(), backward, zero_grad()
• Calling of .eval(), enabling/disabling grads
• Saving and loading weights
• Weights and Biases logging
• Multi-GPU training support
• TPU support
• 16-bit training support

dm = CIFAR10DataModule(batch_size=32)
# To access the x_dataloader we need to call prepare_data and setup.
dm.prepare_data()
dm.setup()

# Samples required by the custom ImagePredictionLogger callback to log image predictions.
val_samples = next(iter(dm.val_dataloader()))
val_imgs, val_labels = val_samples[0], val_samples[1]
val_imgs.shape, val_labels.shape

model = LitModel((3, 32, 32), dm.num_classes)

# Initialize wandb logger
wandb_logger = WandbLogger(project='wandb-lightning', job_type='train')

# Initialize Callbacks
early_stop_callback = pl.callbacks.EarlyStopping(monitor="val_loss")
checkpoint_callback = pl.callbacks.ModelCheckpoint()

# Initialize a trainer
trainer = pl.Trainer(max_epochs=2,
                     logger=wandb_logger,
                     callbacks=[early_stop_callback,
                                ImagePredictionLogger(val_samples),
                                checkpoint_callback],
                     )

# Train the model 
trainer.fit(model, dm)

# Evaluate the model on the held-out test set ⚡⚡
trainer.test(dataloaders=dm.test_dataloader())

# Close wandb run
wandb.finish()

Final Thoughts

I come from the TensorFlow/Keras ecosystem and find PyTorch a bit overwhelming even though it’s an elegant framework. Just my personal experience though. While exploring PyTorch Lightning, I realized that almost all of the reasons that kept me away from PyTorch is taken care of. Here’s a quick summary of my excitement:

• Then: Conventional PyTorch model definition used to be all over the place. With the model in some model.py script and the training loop in the train.py file. It was a lot of looking back and forth to understand the pipeline.
• Now: The LightningModule acts as a system where the model is defined along with the training_step, validation_step, etc. Now it’s modular and shareable.
• Then: The best part about TensorFlow/Keras is the input data pipeline. Their dataset catalog is rich and growing. PyTorch’s data pipeline used to be the biggest pain point. In normal PyTorch code, the data download/cleaning/preparation is usually scattered across many files.
• Now: The DataModule organizes the data pipeline into one shareable and reusable class. It’s simply a collection of a train_dataloader, val_dataloader(s), test_dataloader(s) along with the matching transforms and data processing/downloads steps required.
• Then: With Keras, one can call model.fit to train the model and model.predict to run inference on. model.evaluate offered a good old simple evaluation on the test data. This is not the case with PyTorch. One will usually find separate train.py and test.py files.
• Now: With the LightningModule in place, the Trainer automates everything. One needs to just call trainer.fit and trainer.test to train and evaluate the model.
• Then: TensorFlow loves TPU, PyTorch…
• Now: With PyTorch Lightning, it’s so easy to train the same model with multiple GPUs and even on TPU.
• Then: I am a big fan of Callbacks and prefer writing custom callbacks. Something as trivial as Early Stopping used to be a point of discussion with conventional PyTorch.
• Now: With PyTorch Lightning using Early Stopping and Model Checkpointing is a piece of cake. I can even write custom callbacks.

🎨 Conclusion and Resources

I hope you find this report helpful. I will encourage to play with the code and train an image classifier with a dataset of your choice.

Here are some resources to learn more about PyTorch Lightning:

• Step-by-step walk-through - This is one of the official tutorials. Their documentation is really well written and I highly encourage it as a good learning resource.
• Use Pytorch Lightning with Weights & Biases - This is a quick colab that you can run through to learn more about how to use W&B with PyTorch Lightning.