Using PyTorch Lightning with Tune

PyTorch Lightning is a framework which brings structure into training PyTorch models. It aims to avoid boilerplate code, so you don’t have to write the same training loops all over again when building a new model.

../../_images/pytorch_lightning_full.png

The main abstraction of PyTorch Lightning is the LightningModule class, which should be extended by your application. There is a great post on how to transfer your models from vanilla PyTorch to Lightning.

The class structure of PyTorch Lightning makes it very easy to define and tune model parameters. This tutorial will show you how to use Tune to find the best set of parameters for your application on the example of training a MNIST classifier. Notably, the LightningModule does not have to be altered at all for this - so you can use it plug and play for your existing models, assuming their parameters are configurable!

Note

To run this example, you will need to install the following:

$ pip install ray torch torchvision pytorch-lightning

PyTorch Lightning classifier for MNIST

Let’s first start with the basic PyTorch Lightning implementation of an MNIST classifier. This classifier does not include any tuning code at this point.

Our example builds on the MNIST example from the blog post we talked about earlier.

First, we run some imports:

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

And then there is the Lightning model adapted from the blog post. Note that we left out the test set validation and made the model parameters configurable through a config dict that is passed on initialization. Also, we specify a data_dir where the MNIST data will be stored. Lastly, we added a new metric, the validation accuracy, to the logs.

class LightningMNISTClassifier(pl.LightningModule):
    """
    This has been adapted from
    https://towardsdatascience.com/from-pytorch-to-pytorch-lightning-a-gentle-introduction-b371b7caaf09
    """

    def __init__(self, config, data_dir=None):
        super(LightningMNISTClassifier, self).__init__()

        self.data_dir = data_dir or os.getcwd()

        self.layer_1_size = config["layer_1_size"]
        self.layer_2_size = config["layer_2_size"]
        self.lr = config["lr"]
        self.batch_size = config["batch_size"]

        # mnist images are (1, 28, 28) (channels, width, height)
        self.layer_1 = torch.nn.Linear(28 * 28, self.layer_1_size)
        self.layer_2 = torch.nn.Linear(self.layer_1_size, self.layer_2_size)
        self.layer_3 = torch.nn.Linear(self.layer_2_size, 10)

    def forward(self, x):
        batch_size, channels, width, height = x.size()
        x = x.view(batch_size, -1)

        x = self.layer_1(x)
        x = torch.relu(x)

        x = self.layer_2(x)
        x = torch.relu(x)

        x = self.layer_3(x)
        x = torch.log_softmax(x, dim=1)

        return x

    def cross_entropy_loss(self, logits, labels):
        return F.nll_loss(logits, labels)

    def accuracy(self, logits, labels):
        _, predicted = torch.max(logits.data, 1)
        correct = (predicted == labels).sum().item()
        accuracy = correct / len(labels)
        return torch.tensor(accuracy)

    def training_step(self, train_batch, batch_idx):
        x, y = train_batch
        logits = self.forward(x)
        loss = self.cross_entropy_loss(logits, y)
        accuracy = self.accuracy(logits, y)

        logs = {"ptl/train_loss": loss, "ptl/train_accuracy": accuracy}
        return {"loss": loss, "log": logs}

    def validation_step(self, val_batch, batch_idx):
        x, y = val_batch
        logits = self.forward(x)
        loss = self.cross_entropy_loss(logits, y)
        accuracy = self.accuracy(logits, y)

        return {"val_loss": loss, "val_accuracy": accuracy}

    def validation_epoch_end(self, outputs):
        avg_loss = torch.stack([x["val_loss"] for x in outputs]).mean()
        avg_acc = torch.stack([x["val_accuracy"] for x in outputs]).mean()
        logs = {"ptl/val_loss": avg_loss, "ptl/val_accuracy": avg_acc}

        return {
            "val_loss": avg_loss,
            "val_accuracy": avg_acc,
            "log": logs
        }

    @staticmethod
    def download_data(data_dir):
        transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.1307, ), (0.3081, ))
        ])
        return MNIST(data_dir, train=True, download=True, transform=transform)

    def prepare_data(self):
        mnist_train = self.download_data(self.data_dir)

        self.mnist_train, self.mnist_val = random_split(
            mnist_train, [55000, 5000])

    def train_dataloader(self):
        return DataLoader(self.mnist_train, batch_size=int(self.batch_size))

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

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


def train_mnist(config):
    model = LightningMNISTClassifier(config)
    trainer = pl.Trainer(max_epochs=10, show_progress_bar=False)

    trainer.fit(model)

And that’s it! You can now run train_mnist(config) to train the classifier, e.g. like so:

config = {
    "layer_1_size": 128,
    "layer_2_size": 256,
    "lr": 1e-3,
    "batch_size": 64
}
train_mnist(config)

Tuning the model parameters

The parameters above should give you a good accuracy of over 90% already. However, we might improve on this simply by changing some of the hyperparameters. For instance, maybe we get an even higher accuracy if we used a larger batch size.

Instead of guessing the parameter values, let’s use Tune to systematically try out parameter combinations and find the best performing set.

First, we need some additional imports:

import shutil
from functools import partial
from tempfile import mkdtemp
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning.utilities.cloud_io import load as pl_load
from ray import tune
from ray.tune import CLIReporter
from ray.tune.schedulers import ASHAScheduler, PopulationBasedTraining
from ray.tune.integration.pytorch_lightning import TuneReportCallback, \
    TuneReportCheckpointCallback

Talking to Tune with a PyTorch Lightning callback

PyTorch Lightning introduced Callbacks that can be used to plug custom functions into the training loop. This way the original LightningModule does not have to be altered at all. Also, we could use the same callback for multiple modules.

Ray Tune comes with ready-to-use PyTorch Lightning callbacks. To report metrics back to Tune after each validation epoch, we will use the TuneReportCallback:

from ray.tune.integration.pytorch_lightning import TuneReportCallback
callback = TuneReportCallback({
    "loss": "avg_val_loss",
    "mean_accuracy": "avg_val_accuracy"
}, on="validation_end")

This callback will take the avg_val_loss and avg_val_accuracy values from the PyTorch Lightning trainer and report them to Tune as the loss and mean_accuracy, respectively.

Adding the Tune training function

Then we specify our training function. Note that we added the data_dir as a parameter here to avoid that each training run downloads the full MNIST dataset. Instead, we want to access a shared data location.

We are also able to specify the number of epochs to train each model, and the number of GPUs we want to use for training. We also create a TensorBoard logger that writes logfiles directly into Tune’s root trial directory - if we didn’t do that PyTorch Lightning would create subdirectories, and each trial would thus be shown twice in TensorBoard, one time for Tune’s logs, and another time for PyTorch Lightning’s logs.

def train_mnist_tune(config, data_dir=None, num_epochs=10, num_gpus=0):
    model = LightningMNISTClassifier(config, data_dir)
    trainer = pl.Trainer(
        max_epochs=num_epochs,
        gpus=num_gpus,
        logger=TensorBoardLogger(
            save_dir=tune.get_trial_dir(), name="", version="."),
        progress_bar_refresh_rate=0,
        callbacks=[
            TuneReportCallback(
                {
                    "loss": "val_loss",
                    "mean_accuracy": "val_accuracy"
                },
                on="validation_end")
        ])

    trainer.fit(model)

Sharing the data

All our trials are using the MNIST data. To avoid that each training instance downloads their own MNIST dataset, we download it once and share the data_dir between runs.

data_dir = mkdtemp(prefix="mnist_data_")
LightningMNISTClassifier.download_data(data_dir)

We also delete this data after training to avoid filling up our disk or memory space.

    progress_reporter=reporter,

Configuring the search space

Now we configure the parameter search space. We would like to choose between three different layer and batch sizes. The learning rate should be sampled uniformly between 0.0001 and 0.1. The tune.loguniform() function is syntactic sugar to make sampling between these different orders of magnitude easier, specifically we are able to also sample small values.

config = {
    "layer_1_size": tune.choice([32, 64, 128]),
    "layer_2_size": tune.choice([64, 128, 256]),
    "lr": tune.loguniform(1e-4, 1e-1),
    "batch_size": tune.choice([32, 64, 128]),
}

Selecting a scheduler

In this example, we use an Asynchronous Hyperband scheduler. This scheduler decides at each iteration which trials are likely to perform badly, and stops these trials. This way we don’t waste any resources on bad hyperparameter configurations.

scheduler = ASHAScheduler(
    metric="loss",
    mode="min",
    max_t=num_epochs,
    grace_period=1,
    reduction_factor=2)

Changing the CLI output

We instantiate a CLIReporter to specify which metrics we would like to see in our output tables in the command line. This is optional, but can be used to make sure our output tables only include information we would like to see.

reporter = CLIReporter(
    parameter_columns=["layer_1_size", "layer_2_size", "lr", "batch_size"],
    metric_columns=["loss", "mean_accuracy", "training_iteration"])

Passing constants to the train function

The data_dir, num_epochs and num_gpus we pass to the training function are constants. To avoid including them as non-configurable parameters in the config specification, we can use functools.partial to wrap around the training function.

partial(
    train_mnist_tune,
    data_dir=data_dir,
    num_epochs=num_epochs,
    num_gpus=gpus_per_trial),

Training with GPUs

We can specify how many resources Tune should request for each trial. This also includes GPUs.

PyTorch Lightning takes care of moving the training to the GPUs. We already made sure that our code is compatible with that, so there’s nothing more to do here other than to specify the number of GPUs we would like to use:

    resources_per_trial={

Please note that in the current state of PyTorch Lightning, training on fractional GPUs or multiple GPUs requires some workarounds. We will address these in a separate tutorial - for now this example works with no or exactly one GPU.

Putting it together

Lastly, we need to start Tune with tune.run().

The full code looks like this:

def tune_mnist_asha(num_samples=10, num_epochs=10, gpus_per_trial=0):
    data_dir = mkdtemp(prefix="mnist_data_")
    LightningMNISTClassifier.download_data(data_dir)

    config = {
        "layer_1_size": tune.choice([32, 64, 128]),
        "layer_2_size": tune.choice([64, 128, 256]),
        "lr": tune.loguniform(1e-4, 1e-1),
        "batch_size": tune.choice([32, 64, 128]),
    }

    scheduler = ASHAScheduler(
        metric="loss",
        mode="min",
        max_t=num_epochs,
        grace_period=1,
        reduction_factor=2)

    reporter = CLIReporter(
        parameter_columns=["layer_1_size", "layer_2_size", "lr", "batch_size"],
        metric_columns=["loss", "mean_accuracy", "training_iteration"])

    tune.run(
        partial(
            train_mnist_tune,
            data_dir=data_dir,
            num_epochs=num_epochs,
            num_gpus=gpus_per_trial),
        resources_per_trial={
            "cpu": 1,
            "gpu": gpus_per_trial
        },
        config=config,
        num_samples=num_samples,
        scheduler=scheduler,
        progress_reporter=reporter,
        name="tune_mnist_asha")

    shutil.rmtree(data_dir)

In the example above, Tune runs 10 trials with different hyperparameter configurations. An example output could look like so:

  +------------------------------+------------+-------+----------------+----------------+-------------+--------------+----------+-----------------+----------------------+
  | Trial name                   | status     | loc   |   layer_1_size |   layer_2_size |          lr |   batch_size |     loss |   mean_accuracy |   training_iteration |
  |------------------------------+------------+-------+----------------+----------------+-------------+--------------+----------+-----------------+----------------------|
  | train_mnist_tune_63ecc_00000 | TERMINATED |       |            128 |             64 | 0.00121197  |          128 | 0.120173 |       0.972461  |                   10 |
  | train_mnist_tune_63ecc_00001 | TERMINATED |       |             64 |            128 | 0.0301395   |          128 | 0.454836 |       0.868164  |                    4 |
  | train_mnist_tune_63ecc_00002 | TERMINATED |       |             64 |            128 | 0.0432097   |          128 | 0.718396 |       0.718359  |                    1 |
  | train_mnist_tune_63ecc_00003 | TERMINATED |       |             32 |            128 | 0.000294669 |           32 | 0.111475 |       0.965764  |                   10 |
  | train_mnist_tune_63ecc_00004 | TERMINATED |       |             32 |            256 | 0.000386664 |           64 | 0.133538 |       0.960839  |                    8 |
  | train_mnist_tune_63ecc_00005 | TERMINATED |       |            128 |            128 | 0.0837395   |           32 | 2.32628  |       0.0991242 |                    1 |
  | train_mnist_tune_63ecc_00006 | TERMINATED |       |             64 |            128 | 0.000158761 |          128 | 0.134595 |       0.959766  |                   10 |
  | train_mnist_tune_63ecc_00007 | TERMINATED |       |             64 |             64 | 0.000672126 |           64 | 0.118182 |       0.972903  |                   10 |
  | train_mnist_tune_63ecc_00008 | TERMINATED |       |            128 |             64 | 0.000502428 |           32 | 0.11082  |       0.975518  |                   10 |
  | train_mnist_tune_63ecc_00009 | TERMINATED |       |             64 |            256 | 0.00112894  |           32 | 0.13472  |       0.971935  |                    8 |
  +------------------------------+------------+-------+----------------+----------------+-------------+--------------+----------+-----------------+----------------------+

As you can see in the training_iteration column, trials with a high loss (and low accuracy) have been terminated early. The best performing trial used layer_1_size=128, layer_2_size=64, lr=0.000502428 and batch_size=32.

Using Population Based Training to find the best parameters

The ASHAScheduler terminates those trials early that show bad performance. Sometimes, this stops trials that would get better after more training steps, and which might eventually even show better performance than other configurations.

Another popular method for hyperparameter tuning, called Population Based Training, instead perturbs hyperparameters during the training run. Tune implements PBT, and we only need to make some slight adjustments to our code.

Adding checkpoints to the PyTorch Lightning module

First, we need to introduce another callback to save model checkpoints. Since Tune requires a call to tune.report() after creating a new checkpoint to register it, we will use a combined reporting and checkpointing callback:

from ray.tune.integration.pytorch_lightning import TuneReportCheckpointCallback
callback = TuneReportCheckpointCallback(
    metrics={"loss": "val_loss", "mean_accuracy": "val_accuracy"},
    filename="checkpoint",
    on="validation_end")

The checkpoint value is the name of the checkpoint file within the checkpoint directory.

We also include checkpoint loading in our training function:

def train_mnist_tune_checkpoint(config,
                                checkpoint_dir=None,
                                data_dir=None,
                                num_epochs=10,
                                num_gpus=0):
    trainer = pl.Trainer(
        max_epochs=num_epochs,
        gpus=num_gpus,
        logger=TensorBoardLogger(
            save_dir=tune.get_trial_dir(), name="", version="."),
        progress_bar_refresh_rate=0,
        callbacks=[
            TuneReportCheckpointCallback(
                metrics={
                    "loss": "val_loss",
                    "mean_accuracy": "val_accuracy"
                },
                filename="checkpoint",
                on="validation_end")
        ])
    if checkpoint_dir:
        # Currently, this leads to errors:
        # model = LightningMNISTClassifier.load_from_checkpoint(
        #     os.path.join(checkpoint, "checkpoint"))
        # Workaround:
        ckpt = pl_load(
            os.path.join(checkpoint_dir, "checkpoint"),
            map_location=lambda storage, loc: storage)
        model = LightningMNISTClassifier._load_model_state(ckpt, config=config)
        trainer.current_epoch = ckpt["epoch"]
    else:
        model = LightningMNISTClassifier(config=config, data_dir=data_dir)

    trainer.fit(model)

Configuring and running Population Based Training

We need to call Tune slightly differently:

def tune_mnist_pbt(num_samples=10, num_epochs=10, gpus_per_trial=0):
    data_dir = mkdtemp(prefix="mnist_data_")
    LightningMNISTClassifier.download_data(data_dir)

    config = {
        "layer_1_size": tune.choice([32, 64, 128]),
        "layer_2_size": tune.choice([64, 128, 256]),
        "lr": 1e-3,
        "batch_size": 64,
    }

    scheduler = PopulationBasedTraining(
        time_attr="training_iteration",
        metric="loss",
        mode="min",
        perturbation_interval=4,
        hyperparam_mutations={
            "lr": lambda: tune.loguniform(1e-4, 1e-1).func(None),
            "batch_size": [32, 64, 128]
        })

    reporter = CLIReporter(
        parameter_columns=["layer_1_size", "layer_2_size", "lr", "batch_size"],
        metric_columns=["loss", "mean_accuracy", "training_iteration"])

    tune.run(
        partial(
            train_mnist_tune_checkpoint,
            data_dir=data_dir,
            num_epochs=num_epochs,
            num_gpus=gpus_per_trial),
        resources_per_trial={
            "cpu": 1,
            "gpu": gpus_per_trial
        },
        config=config,
        num_samples=num_samples,
        scheduler=scheduler,
        progress_reporter=reporter,
        name="tune_mnist_pbt")

    shutil.rmtree(data_dir)

Instead of passing tune parameters to the config dict, we start with fixed values, though we are also able to sample some of them, like the layer sizes. Additionally, we have to tell PBT how to perturb the hyperparameters. Note that the layer sizes are not tuned right here. This is because we cannot simply change layer sizes during a training run - which is what would happen in PBT.

An example output could look like this:

+-----------------------------------------+------------+-------+----------------+----------------+-----------+--------------+-----------+-----------------+----------------------+
| Trial name                              | status     | loc   |   layer_1_size |   layer_2_size |        lr |   batch_size |      loss |   mean_accuracy |   training_iteration |
|-----------------------------------------+------------+-------+----------------+----------------+-----------+--------------+-----------+-----------------+----------------------|
| train_mnist_tune_checkpoint_85489_00000 | TERMINATED |       |            128 |            128 | 0.001     |           64 | 0.108734  |        0.973101 |                   10 |
| train_mnist_tune_checkpoint_85489_00001 | TERMINATED |       |            128 |            128 | 0.001     |           64 | 0.093577  |        0.978639 |                   10 |
| train_mnist_tune_checkpoint_85489_00002 | TERMINATED |       |            128 |            256 | 0.0008    |           32 | 0.0922348 |        0.979299 |                   10 |
| train_mnist_tune_checkpoint_85489_00003 | TERMINATED |       |             64 |            256 | 0.001     |           64 | 0.124648  |        0.973892 |                   10 |
| train_mnist_tune_checkpoint_85489_00004 | TERMINATED |       |            128 |             64 | 0.001     |           64 | 0.101717  |        0.975079 |                   10 |
| train_mnist_tune_checkpoint_85489_00005 | TERMINATED |       |             64 |             64 | 0.001     |           64 | 0.121467  |        0.969146 |                   10 |
| train_mnist_tune_checkpoint_85489_00006 | TERMINATED |       |            128 |            256 | 0.00064   |           32 | 0.053446  |        0.987062 |                   10 |
| train_mnist_tune_checkpoint_85489_00007 | TERMINATED |       |            128 |            256 | 0.001     |           64 | 0.129804  |        0.973497 |                   10 |
| train_mnist_tune_checkpoint_85489_00008 | TERMINATED |       |             64 |            256 | 0.0285125 |          128 | 0.363236  |        0.913867 |                   10 |
| train_mnist_tune_checkpoint_85489_00009 | TERMINATED |       |             32 |            256 | 0.001     |           64 | 0.150946  |        0.964201 |                   10 |
+-----------------------------------------+------------+-------+----------------+----------------+-----------+--------------+-----------+-----------------+----------------------+

As you can see, each sample ran the full number of 10 iterations. All trials ended with quite good parameter combinations and showed relatively good performances. In some runs, the parameters have been perturbed. And the best configuration even reached a mean validation accuracy of 0.987062!

In summary, PyTorch Lightning Modules are easy to extend to use with Tune. It just took us importing one or two callbacks and a small wrapper function to get great performing parameter configurations.