The lightweight PyTorch wrapper for high-performance AI research. Scale your models, not the boilerplate.

---

Website • Key Features • How To Use • Docs • Examples • Community • Grid AI • License

*Codecov is > 90%+ but build delays may show less

---

PyTorch Lightning is just organized PyTorch

Lightning disentangles PyTorch code to decouple the science from the engineering.

---

Lightning Design Philosophy

Lightning structures PyTorch code with these principles:

Lightning forces the following structure to your code which makes it reusable and shareable:

• Research code (the LightningModule).
• Engineering code (you delete, and is handled by the Trainer).
• Non-essential research code (logging, etc... this goes in Callbacks).
• Data (use PyTorch DataLoaders or organize them into a LightningDataModule).

Once you do this, you can train on multiple-GPUs, TPUs, CPUs and even in 16-bit precision without changing your code!

Get started with our 2 step guide

---

Continuous Integration

Lightning is rigorously tested across multiple GPUs, TPUs CPUs and against major Python and PyTorch versions.

Current build statuses

System / PyTorch ver.	1.6 (min. req.)	1.7	1.8 (LTS)	1.9 (latest)	1.10 (nightly)
Conda py3.7 [linux]
Linux py3.7 [GPUs**]	-	-		-	-
Linux py3.7 [TPUs***]	-	-		-	-
Linux py3.{6,7,8,9}		-	-		-
OSX py3.{6,7,8,9}		-	-		-
Windows py3.{6,7,8,9}		-	-		-

• ** tests run on two NVIDIA P100
• *** tests run on Google GKE TPUv2/3
• TPU py3.7 means we support Colab and Kaggle env.

---

How To Use

Step 0: Install

Simple installation from PyPI

pip install pytorch-lightning

Other installation options

Install with optional dependencies

pip install pytorch-lightning['extra']

Conda

conda install pytorch-lightning -c conda-forge

Install stable 1.4.x

the actual status of 1.4 [stable] is following:

Install future release from the source

pip install git+https://github.com/PytorchLightning/pytorch-lightning.git@release/1.4.x --upgrade

Install bleeding-edge - future 1.5

Install nightly from the source (no guarantees)

pip install https://github.com/PyTorchLightning/pytorch-lightning/archive/master.zip

or from testing PyPI

pip install -iU https://test.pypi.org/simple/ pytorch-lightning

Step 1: Add these imports

import os
import torch
from torch import nn
import torch.nn.functional as F
from torchvision.datasets import MNIST
from torch.utils.data import DataLoader, random_split
from torchvision import transforms
import pytorch_lightning as pl

Step 2: Define a LightningModule (nn.Module subclass)

A LightningModule defines a full system (ie: a GAN, autoencoder, BERT or a simple Image Classifier).

class LitAutoEncoder(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.encoder = nn.Sequential(nn.Linear(28 * 28, 128), nn.ReLU(), nn.Linear(128, 3))
        self.decoder = nn.Sequential(nn.Linear(3, 128), nn.ReLU(), nn.Linear(128, 28 * 28))

    def forward(self, x):
        # in lightning, forward defines the prediction/inference actions
        embedding = self.encoder(x)
        return embedding

    def training_step(self, batch, batch_idx):
        # training_step defines the train loop. It is independent of forward
        x, y = batch
        x = x.view(x.size(0), -1)
        z = self.encoder(x)
        x_hat = self.decoder(z)
        loss = F.mse_loss(x_hat, x)
        self.log("train_loss", loss)
        return loss

    def configure_optimizers(self):
        optimizer = torch.optim.Adam(self.parameters(), lr=1e-3)
        return optimizer

Note: Training_step defines the training loop. Forward defines how the LightningModule behaves during inference/prediction.

Step 3: Train!

dataset = MNIST(os.getcwd(), download=True, transform=transforms.ToTensor())
train, val = random_split(dataset, [55000, 5000])

autoencoder = LitAutoEncoder()
trainer = pl.Trainer()
trainer.fit(autoencoder, DataLoader(train), DataLoader(val))

Advanced features

Lightning has over 40+ advanced features designed for professional AI research at scale.

Here are some examples:

Highlighted feature code snippets

# 8 GPUs
# no code changes needed
trainer = Trainer(max_epochs=1, gpus=8)

# 256 GPUs
trainer = Trainer(max_epochs=1, gpus=8, num_nodes=32)

Train on TPUs without code changes

# no code changes needed
trainer = Trainer(tpu_cores=8)

16-bit precision

# no code changes needed
trainer = Trainer(precision=16)

Experiment managers

from pytorch_lightning import loggers

# tensorboard
trainer = Trainer(logger=TensorBoardLogger("logs/"))

# weights and biases
trainer = Trainer(logger=loggers.WandbLogger())

# comet
trainer = Trainer(logger=loggers.CometLogger())

# mlflow
trainer = Trainer(logger=loggers.MLFlowLogger())

# neptune
trainer = Trainer(logger=loggers.NeptuneLogger())

# ... and dozens more

EarlyStopping

es = EarlyStopping(monitor="val_loss")
trainer = Trainer(callbacks=[es])

Checkpointing

checkpointing = ModelCheckpoint(monitor="val_loss")
trainer = Trainer(callbacks=[checkpointing])

Export to torchscript (JIT) (production use)

# torchscript
autoencoder = LitAutoEncoder()
torch.jit.save(autoencoder.to_torchscript(), "model.pt")

Export to ONNX (production use)

# onnx
with tempfile.NamedTemporaryFile(suffix=".onnx", delete=False) as tmpfile:
    autoencoder = LitAutoEncoder()
    input_sample = torch.randn((1, 64))
    autoencoder.to_onnx(tmpfile.name, input_sample, export_params=True)
    os.path.isfile(tmpfile.name)

Pro-level control of training loops (advanced users)

For complex/professional level work, you have optional full control of the training loop and optimizers.

class LitAutoEncoder(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.automatic_optimization = False

    def training_step(self, batch, batch_idx):
        # access your optimizers with use_pl_optimizer=False. Default is True
        opt_a, opt_b = self.optimizers(use_pl_optimizer=True)

        loss_a = ...
        self.manual_backward(loss_a, opt_a)
        opt_a.step()
        opt_a.zero_grad()

        loss_b = ...
        self.manual_backward(loss_b, opt_b, retain_graph=True)
        self.manual_backward(loss_b, opt_b)
        opt_b.step()
        opt_b.zero_grad()

---

Advantages over unstructured PyTorch

• Models become hardware agnostic
• Code is clear to read because engineering code is abstracted away
• Easier to reproduce
• Make fewer mistakes because lightning handles the tricky engineering
• Keeps all the flexibility (LightningModules are still PyTorch modules), but removes a ton of boilerplate
• Lightning has dozens of integrations with popular machine learning tools.
• Tested rigorously with every new PR. We test every combination of PyTorch and Python supported versions, every OS, multi GPUs and even TPUs.
• Minimal running speed overhead (about 300 ms per epoch compared with pure PyTorch).

---

Examples

Hello world

• MNIST hello world

Contrastive Learning

• BYOL
• CPC v2
• Moco v2
• SIMCLR

NLP

• GPT-2
• BERT

Reinforcement Learning

• DQN
• Dueling-DQN
• Reinforce

Vision

• GAN

Classic ML

• Logistic Regression
• Linear Regression

---

Community

The lightning community is maintained by

• 10+ core contributors who are all a mix of professional engineers, Research Scientists, and Ph.D. students from top AI labs.
• 480+ active community contributors.

Want to help us build Lightning and reduce boilerplate for thousands of researchers? Learn how to make your first contribution here

Lightning is also part of the PyTorch ecosystem which requires projects to have solid testing, documentation and support.

Asking for help

If you have any questions please:

1. Read the docs.
2. Search through existing Discussions, or add a new question
3. Join our slack.

Funding

We're venture funded to make sure we can provide around the clock support, hire a full-time staff, attend conferences, and move faster through implementing features you request.

---

Grid AI

Grid AI is our platform for training models at scale on the cloud!

Sign up for our FREE community Tier here

To use grid, take your regular command:

python my_model.py --learning_rate 1e-6 --layers 2 --gpus 4

And change it to use the grid train command:

grid train --grid_gpus 4 my_model.py --learning_rate 'uniform(1e-6, 1e-1, 20)' --layers '[2, 4, 8, 16]'

The above command will launch (20 * 4) experiments each running on 4 GPUs (320 GPUs!) - by making ZERO changes to your code.