PyTorch Lightning

The PyTorch Keras for ML researchers. More control. Less boilerplate.

Simple installation from PyPI

pip install pytorch-lightning  

Docs

View the docs here

What is it?

Lightning is a very lightweight wrapper on PyTorch. This means you don't have to learn a new library. It defers core training and validation logic to you and automates the rest. It guarantees tested, correct, modern best practices for the automated parts.

Why do I want to use lightning?

When starting a new project the last thing you want to do is recode a training loop, multi-cluster training, 16-bit precision, early-stopping, model loading/saving, when to validate, etc... You're likely to spend a long time ironing out all the bugs without even getting to the core of your research.

With lightning, you guarantee those parts of your code work so you can focus on what the meat of the research: The data and the training/validation loop logic.

Don't worry about training on multiple gpus or speeding up your code, lightning will do that for you!

---

README Table of Contents

• How do I use it
• What lightning automates
• Tensorboard integration
• Lightning features
• Demos
• Tutorials
• Contributing
• Bleeding edge install
• Lightning Design Principles
• FAQ

---

How do I do use it?

To use lightning do 2 things:

1. Define a LightningModel

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

import pytorch_lightning as pl

class CoolModel(pl.LightningModule):

    def __init__(self):
        super(CoolModel, self).__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 my_loss(self, y_hat, y):
        return F.cross_entropy(y_hat, y)

    def training_step(self, batch, batch_nb):
        x, y = batch
        y_hat = self.forward(x)
        return {'loss': self.my_loss(y_hat, y)}

    def validation_step(self, batch, batch_nb):
        x, y = batch
        y_hat = self.forward(x)
        return {'val_loss': self.my_loss(y_hat, y)}

    def validation_end(self, outputs):
        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
        return {'avg_val_loss': avg_loss}

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

    @pl.data_loader
    def tng_dataloader(self):
        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

    @pl.data_loader
    def val_dataloader(self):
        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

    @pl.data_loader
    def test_dataloader(self):
        return DataLoader(MNIST(os.getcwd(), train=True, download=True, transform=transforms.ToTensor()), batch_size=32)

2. Fit with a trainer

from pytorch_lightning import Trainer
from test_tube import Experiment

model = CoolModel()
exp = Experiment(save_dir=os.getcwd())

# train on cpu using only 10% of the data (for demo purposes)
trainer = Trainer(experiment=exp, max_nb_epochs=1, train_percent_check=0.1)

# train on 4 gpus
# trainer = Trainer(experiment=exp, max_nb_epochs=1, gpus=[0, 1, 2, 3])

# train on 32 gpus across 4 nodes (make sure to submit appropriate SLURM job)
# trainer = Trainer(experiment=exp, max_nb_epochs=1, gpus=[0, 1, 2, 3, 4, 5, 6, 7], nb_gpu_nodes=4)

# train (1 epoch only here for demo)
trainer.fit(model)

# view tensorflow logs 
print('View tensorboard logs by running\ntensorboard --logdir %s' % os.getcwd())
print('and going to http://localhost:6006 on your browser')

What does lightning control for me?

Everything in gray!
You define the blue parts using the LightningModule interface:

# what to do in the training loop
def training_step(self, data_batch, batch_nb):

# what to do in the validation loop
def validation_step(self, data_batch, batch_nb):

# how to aggregate validation_step outputs
def validation_end(self, outputs):

# and your dataloaders
def tng_dataloader():
def val_dataloader():
def test_dataloader():

Could be as complex as seq-2-seq + attention

# define what happens for training here
def training_step(self, data_batch, batch_nb):
    x, y = data_batch
    
    # define your own forward and loss calculation
    hidden_states = self.encoder(x)
     
    # even as complex as a seq-2seq + 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, data_batch, batch_nb):    
    x, y = data_batch
    
    # or as basic as a CNN classification
    out = self.forward(x)
    loss = my_loss(out, y)
    return {'loss': loss} 

And you also decide how to collate the output of all validation steps

def validation_end(self, outputs):
    """
    Called at the end of validation to aggregate outputs
    :param outputs: list of individual outputs of each validation step
    :return:
    """
    val_loss_mean = 0
    val_acc_mean = 0
    for output in outputs:
        val_loss_mean += output['val_loss']
        val_acc_mean += output['val_acc']

    val_loss_mean /= len(outputs)
    val_acc_mean /= len(outputs)
    tqdm_dic = {'val_loss': val_loss_mean.item(), 'val_acc': val_acc_mean.item()}
    return tqdm_dic

Tensorboard

Lightning is fully integrated with tensorboard.

Lightning also adds a text column with all the hyperparameters for this experiment.

Simply note the path you set for the Experiment

from test_tube import Experiment
from pytorch-lightning import  Trainer

exp = Experiment(save_dir='/some/path')
trainer = Trainer(experiment=exp)
...

And run tensorboard from that dir

tensorboard --logdir /some/path     

Lightning automates all of the following (each is also configurable):

Checkpointing

• Model saving
• Model loading
• Restoring training session

Computing cluster (SLURM)

• Running grid search on a cluster
• Walltime auto-resubmit

Debugging

• Fast dev run
• Inspect gradient norms
• Log GPU usage
• Make model overfit on subset of data
• Print the parameter count by layer
• Print which gradients are nan

Distributed training

• 16-bit mixed precision
• Multi-GPU
• Multi-node
• Single GPU
• Self-balancing architecture

Experiment Logging

• Display metrics in progress bar
• Log metric row every k batches
• Process position
• Tensorboard support
• Save a snapshot of all hyperparameters
• Snapshot code for a training run
• Write logs file to csv every k batches

Training loop

• Accumulate gradients
• Force training for min or max epochs
• Force disable early stop
• Gradient Clipping
• Hooks
• Learning rate scheduling
• Use multiple optimizers (like GANs)
• Set how much of the training set to check (1-100%)

Validation loop

• Check validation every n epochs
• Hooks
• Set how much of the validation set to check
• Set how much of the test set to check
• Set validation check frequency within 1 training epoch
• Set the number of validation sanity steps

Demo

# install lightning
pip install pytorch-lightning

# clone lightning for the demo
git clone https://github.com/williamFalcon/pytorch-lightning.git
cd pytorch-lightning
cd examples/new_project_templates/

# all of the following demos use the SAME model to show no modification needs to be made to your code

# train on cpu 
python single_cpu_template.py

# train on multiple-gpus 
python single_gpu_node_template.py --gpus "0,1"

# train on 32 gpus on a cluster (run on a SLURM managed cluster)
python multi_node_cluster_template.py --nb_gpu_nodes 4 --gpus '0,1,2,3,4,5,6,7'

Tutorials

• Basic Lightning use
• 9 key speed features in Pytorch-Lightning
• SLURM, multi-node training with Lightning

---

FAQ

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 unecessary abstractions!

Are there plans to support Python 2?
Nope.

Contributing

Welcome to the PyTorch Lightning community! We're building the most advanced research platform on the planet to implement the latest, best practices that the amazing PyTorch team rolls out!

Lightning Design Principles

We encourage all sorts of contributions you're interested in adding! When coding for lightning, please follow these principles.

No PyTorch interference

We don't want to add any abstractions on top of pure PyTorch. This gives researchers all the control they need without having to learn yet another framework.

Simple Internal Code

It's useful for users to look at the code and understand very quickly what's happening. Many users won't be engineers. Thus we need to value clear, simple code over condensed ninja moves. While that's super cool, this isn't the project for that :)

Force User Decisions To Best Practices

There are 1,000 ways to do something. However, something eventually becomes standard practice that everyone does. Thus we pick one way of doing it and force everyone to do it this way. A good example is accumulated gradients. There are many ways to implement, we just pick one and force users to use that one. A bad forced decision would be to make users use a specific library to do something.

When something becomes a best practice, we add it to the framework. This likely looks like code in utils or in the model file that everyone keeps adding over and over again across projects. When this happens, bring that code inside the trainer and add a flag for it.

Simple External API

What makes sense to you may not make sense to others. Create an issue with an API change suggestion and validate that it makes sense for others. Treat code changes how you treat a startup: validate that it's a needed feature, then add if it makes sense for many people.

Gain User Trust

As a researcher you can't have any part of your code going wrong. So, make thorough tests that ensure an implementation of a new trick or subbtle change is correct.

Bug fixes:

1. Submit a github issue.
2. Fix it.
3. Submit a PR!

New Features:

1. Submit a github issue.
2. We'll agree on the feature scope.
3. Submit a PR! (with updated docs and tests 🙃).

Bleeding edge

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

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