PyTorch Lightning

The lightweight PyTorch wrapper for ML researchers. Scale your models. Write less boilerplate.

---

Trending contributors

---

Continuous Integration

System / PyTorch ver.	1.3 (min. req.)	1.4	1.5 (latest)
Conda py3.7 [linux]
Linux py3.7 [GPU]	-	-
Linux py3.7 [TPU]	-	-
Linux py3.6 / py3.7 / py3.8		-
OSX py3.6 / py3.7	-
Windows py3.6 / py3.7 / py3.8			-

Simple installation from PyPI

pip install pytorch-lightning

From Conda

conda install pytorch-lightning -c conda-forge

Docs

• master
• stable
• 0.8.4
• 0.8.3
• 0.8.1
• 0.7.6
• 0.7.5
• 0.7.3
• 0.7.1
• 0.6.0
• 0.5.3

Refactoring your PyTorch code + benefits + full walk-through

Demo

Here's a minimal example without a validation or test loop.

# this is just a plain nn.Module with some structure

class LitClassifier(pl.LightningModule):

    def __init__(self):
        super().__init__()
        self.l1 = torch.nn.Linear(28 * 28, 10)

    def forward(self, x):
        return torch.relu(self.l1(x.view(x.size(0), -1)))

    def training_step(self, batch, batch_nb):
        x, y = batch
        loss = F.cross_entropy(self(x), y)
        tensorboard_logs = {'train_loss': loss}
        return {'loss': loss, 'log': tensorboard_logs}

    def configure_optimizers(self):
        return torch.optim.Adam(self.parameters(), lr=0.02)

# train!
train_loader = DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

model = LitClassifier()
trainer = pl.Trainer(gpus=8, precision=16)    
trainer.fit(model, train_loader) 

Other examples:
MNIST hello world
GAN
BERT
DQN
MNIST on TPUs

What is it?

READ THIS QUICK START PAGE

Lightning is a way to organize your PyTorch code to decouple the science code from the engineering. It's more of a PyTorch style-guide than a framework.

In Lightning, you organize your code into 3 distinct categories:

1. Research code (goes in the LightningModule).
2. Engineering code (you delete, and is handled by the Trainer).
3. Non-essential research code (logging, etc... this goes in Callbacks).

Here's an example of how to refactor your research code into a LightningModule.

The rest of the code is automated by the Trainer!

Testing Rigour

All the automated code by the Trainer is tested rigorously with every new PR.

For every PR we test all combinations of:

• PyTorch 1.3, 1.4, 1.5
• Python 3.6, 3.7, 3.8
• Linux, OSX, Windows
• Multiple GPUs

How does performance compare with vanilla PyTorch?
We have tests to ensure we get the EXACT same results in under 600 ms difference per epoch. In reality, lightning adds about a 300 ms overhead per epoch.
Check out the parity tests here.

Overall, Lightning guarantees rigorously tested, correct, modern best practices for the automated parts.

How flexible is it?

As you see, you're just organizing your PyTorch code - there's no abstraction.

And for the stuff that the Trainer abstracts out, you can override any part you want to do things like implement your own distributed training, 16-bit precision, or even a custom backward pass.

For example, here you could do your own backward pass without worrying about GPUs, TPUs or 16-bit since we already handle it.

class LitModel(LightningModule):
    def optimizer_step(self, current_epoch, batch_idx, optimizer, optimizer_idx,
                     second_order_closure=None, on_tpu=False, using_native_amp=False, using_lbfgs=False):
      optimizer.step()

    def optimizer_zero_grad(self, current_epoch, batch_idx, optimizer, opt_idx):
      optimizer.zero_grad()

For anything else you might need, we have an extensive callback system you can use to add arbitrary functionality not implemented by our team in the Trainer.

Who is Lightning for?

• Professional researchers
• Ph.D. students
• Corporate production teams

If you're just getting into deep learning, we recommend you learn PyTorch first! Once you've implemented a few models, come back and use all the advanced features of Lightning :)

What does lightning control for me?

Everything in Blue! This is how lightning separates the science (red) from engineering (blue).

How much effort is it to convert?

If your code is not a huge mess you should be able to organize it into a LightningModule in less than 1 hour. If your code IS a mess, then you needed to clean up anyhow ;)

Check out this step-by-step guide.
Or watch this video.

Starting a new project?

Use our seed-project aimed at reproducibility!

Why do I want to use lightning?

Although your research/production project might start simple, once you add things like GPU AND TPU training, 16-bit precision, etc, you end up spending more time engineering than researching. Lightning automates AND rigorously tests those parts for you.

Support

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

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

---

README Table of Contents

• How do I use it
• What lightning automates
• Tensorboard integration
• Lightning features
• Examples
• Tutorials
• Asking for help
• Contributing
• Bleeding edge install
• Lightning Design Principles
• Lightning team
• FAQ

---

Realistic example

Here's how you would organize a realistic PyTorch project into Lightning.

The LightningModule defines a system such as seq-2-seq, GAN, etc... It can ALSO define a simple classifier.

In summary, you:

1. Define a LightningModule

    class LitSystem(pl.LightningModule):

        def __init__(self):
            super().__init__()
            # not the best model...
            self.l1 = torch.nn.Linear(28 * 28, 10)

        def forward(self, x):
            return torch.relu(self.l1(x.view(x.size(0), -1)))

        def training_step(self, batch, batch_idx):
            ...

2. Fit it with a Trainer

from pytorch_lightning import Trainer

model = LitSystem()

# most basic trainer, uses good defaults
trainer = Trainer()
trainer.fit(model)

Check out the COLAB demo here

What types of research works?

Anything! Remember, that this is just organized PyTorch code. The Training step defines the core complexity found in the training loop.

Could be as complex as a seq2seq

# define what happens for training here
def training_step(self, batch, batch_idx):
    x, y = batch

    # define your own forward and loss calculation
    hidden_states = self.encoder(x)

    # even as complex as a seq-2-seq + attn model
    # (this is just a toy, non-working example to illustrate)
    start_token = '<SOS>'
    last_hidden = torch.zeros(...)
    loss = 0
    for step in range(max_seq_len):
        attn_context = self.attention_nn(hidden_states, start_token)
        pred = self.decoder(start_token, attn_context, last_hidden)
        last_hidden = pred
        pred = self.predict_nn(pred)
        loss += self.loss(last_hidden, y[step])

    #toy example as well
    loss = loss / max_seq_len
    return {'loss': loss}

Or as basic as CNN image classification

# define what happens for validation here
def validation_step(self, batch, batch_idx):
    x, y = batch

    # or as basic as a CNN classification
    out = self(x)
    loss = my_loss(out, y)
    return {'loss': loss}

And without changing a single line of code, you could run on CPUs

trainer = Trainer(max_epochs=1)

Or GPUs

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

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

Or TPUs

# Distributes TPU core training
trainer = Trainer(tpu_cores=8)

# Single TPU core training
trainer = Trainer(tpu_cores=[1])

When you're done training, run the test accuracy

trainer.test()

Visualization

Lightning has out-of-the-box integration with the popular logging/visualizing frameworks

• Tensorboard
• MLFlow
• Neptune.ai
• Comet.ml
• Wandb
• ...

Lightning automates 40+ parts of DL/ML research

• GPU training
• Distributed GPU (cluster) training
• TPU training
• EarlyStopping
• Logging/Visualizing
• Checkpointing
• Experiment management
• Full list here

Running speed

Migrating to lightning does not mean compromising on speed! You can expect an overhead of about 300 ms per epoch compared with pure PyTorch.

Examples

Check out this awesome list of research papers and implementations done with Lightning.

• Contextual Emotion Detection (DoubleDistilBert)
• Generative Adversarial Network
• Hyperparameter optimization with Optuna
• Image Inpainting using Partial Convolutions
• MNIST on TPU
• NER (transformers, TPU, huggingface)
• NeuralTexture (CVPR)
• Recurrent Attentive Neural Process
• Siamese Nets for One-shot Image Recognition
• Speech Transformers
• Transformers transfer learning (Huggingface)
• Transformers text classification
• VAE Library of over 18+ VAE flavors
• Finetune BERT, RoBERTa etc on QA Datasets like SQuAD
• Pytorch-Lightning + Microsoft NNI with Docker

Tutorials

Check out our introduction guide to get started. Or jump straight into our tutorials.

---

Asking for help

Welcome to the Lightning community!

If you have any questions, feel free to:

1. read the docs.
2. Search through the issues.
3. Ask on stackoverflow with the tag pytorch-lightning.
4. Join our slack.

---

FAQ

How do I use Lightning for rapid research?
Here's a walk-through

Why was Lightning created?
Lightning has 3 goals in mind:

1. Maximal flexibility while abstracting out the common boilerplate across research projects.
2. Reproducibility. If all projects use the LightningModule template, it will be much much easier to understand what's going on and where to look! It will also mean every implementation follows a standard format.
3. Democratizing PyTorch power-user features. Distributed training? 16-bit? know you need them but don't want to take the time to implement? All good... these come built into Lightning.

How does Lightning compare with Ignite and fast.ai?
Here's a thorough comparison.

Is this another library I have to learn?
Nope! We use pure Pytorch everywhere and don't add unnecessary abstractions!

Are there plans to support Python 2?
Nope.

Are there plans to support virtualenv?
Nope. Please use anaconda or miniconda.

conda activate my_env
pip install pytorch-lightning

Custom installation

Bleeding edge

If you can't wait for the next release, install the most up to date code with:

• using GIT (locally clone whole repo with full history)

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

• using instant zip (last state of the repo without git history)

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

Any release installation

You can also install any past release 0.X.Y from this repository:

pip install https://github.com/PytorchLightning/pytorch-lightning/archive/0.X.Y.zip --upgrade

Lightning team

Leads

• William Falcon (williamFalcon) (Lightning founder)
• Jirka Borovec (Borda) (ghost :)
• Ethan Harris (ethanwharris) (Torchbearer founder)
• Matthew Painter (MattPainter01) (Torchbearer founder)
• Justus Schock (justusschock) (Former Core Member PyTorch Ignite)

Core Maintainers

• Nick Eggert (neggert)
• Jeff Ling (jeffling)
• Jeremy Jordan (jeremyjordan)
• Tullie Murrell (tullie)
• Adrian Wälchli (awaelchli)
• Nicki Skafte (skaftenicki)

---

Funding

Building open-source software with only a few part-time people is hard! We've secured funding to make sure we can hire a full-time staff, attend conferences, and move faster through implementing features you request.

Our goal is to build an incredible research platform and a big supportive community. Many open-source projects have gone on to fund operations through things like support and special help for big corporations!

If you are one of these corporations, please feel free to reach out to will@pytorchlightning.ai!

BibTeX

If you want to cite the framework feel free to use this (but only if you loved it 😊):

@article{falcon2019pytorch,
  title={PyTorch Lightning},
  author={Falcon, WA},
  journal={GitHub. Note: https://github.com/PyTorchLightning/pytorch-lightning Cited by},
  volume={3},
  year={2019}
}