Build and train PyTorch models and connect them to the ML lifecycle using Lightning App templates, without handling DIY infrastructure, cost management, scaling, and other headaches.
Go to file
William Falcon e9fca35039 added val loop options 2019-06-27 13:47:19 -04:00
docs added val loop options 2019-06-27 13:47:19 -04:00
pytorch_lightning renamed options 2019-06-27 11:27:11 -04:00
.gitignore updated args 2019-06-25 19:42:15 -04:00
COPYING Add src, docs and other important folders 2019-04-03 22:16:02 +05:30
MANIFEST.in Fix pip install too 2019-04-03 22:47:55 +05:30
README.md added docs page 2019-06-26 20:03:39 -04:00
mkdocs.yml added lightning model docs 2019-06-27 10:04:24 -04:00
pyproject.toml Fix pip install too 2019-04-03 22:47:55 +05:30
requirements.txt initial commit 2019-03-30 20:50:32 -04:00
setup.cfg Fix pip install too 2019-04-03 22:47:55 +05:30
setup.py release v0.11 2019-06-26 18:44:59 -04:00
update.sh beta release to pypi 2019-03-31 15:26:23 -04:00

README.md

Pytorch Lightning

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

PyPI version

pip install pytorch-lightning    

Docs

View the docs here

What is it?

Keras is too abstract for researchers. Lightning abstracts the full training loop but gives you control in the critical points.

Why do I want to use lightning?

Because you want to use best practices and get gpu training, multi-node training, checkpointing, mixed-precision, etc... for free.

To use lightning do 2 things:

  1. Define a trainer (which will run ALL your models).
  2. Define a model.

What are some key lightning features?

  • Automatic training loop
# define what happens for training here
def training_step(self, data_batch, batch_nb):
    x, y = data_batch
    out = self.forward(x)
    loss = my_loss(out, y)
    return {'loss': loss} 
  • Automatic validation loop
# define what happens for validation here
def validation_step(self, data_batch, batch_nb):    x, y = data_batch
    out = self.forward(x)
    loss = my_loss(out, y)
    return {'loss': loss} 
  • Automatic early stopping
callback = EarlyStopping(...)
Trainer(early_stopping=callback)
  • Learning rate annealing
# anneal at 100 and 200 epochs
Trainer(lr_scheduler_milestones=[100, 200])
  • 16 bit precision training (must have apex installed)
Trainer(use_amp=True, amp_level='O2')
  • multi-gpu training
# train on 4 gpus
Trainer(gpus=[0, 1, 2, 3])
  • Automatic checkpointing
# do 3 things:
# 1
Trainer(checkpoint_callback=ModelCheckpoint)

# 2 return what to save in a checkpoint
def get_save_dict(self):
    return {'state_dict': self.state_dict()}


# 3 use the checkpoint to reset your model state
def load_model_specific(self, checkpoint):
    self.load_state_dict(checkpoint['state_dict'])
  • Log all details of your experiment (model params, code snapshot, etc...)
from test_tube import Experiment

exp = Experiment(...)
Trainer(experiment=exp)
  • Run grid-search on cluster
from test_tube import Experiment, SlurmCluster, HyperOptArgumentParser

def training_fx(hparams, cluster, _):
    # hparams are local params
    model = MyModel()
    trainer = Trainer(...)
    trainer.fit(model)

# grid search number of layers
parser = HyperOptArgumentParser(strategy='grid_search')
parser.opt_list('--layers', default=5, type=int, options=[1, 5, 10, 20, 50])
hyperparams = parser.parse_args()

cluster = SlurmCluster(hyperparam_optimizer=hyperparams)
cluster.optimize_parallel_cluster_gpu(training_fx)

Demo

# install lightning
pip install pytorch-lightning

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

# run demo (on cpu)
python fully_featured_trainer.py

Without changing the model AT ALL, you can run the model on a single gpu, over multiple gpus, or over multiple nodes.

# run a grid search on two gpus
python fully_featured_trainer.py --gpus "0;1"

# run single model on multiple gpus
python fully_featured_trainer.py --gpus "0;1" --interactive