Tip

This tutorial is available as a Jupyter notebook.

🛑 Early Stopping

In this tutorial, we’re going to learn about how to perform early stopping in Composer using callbacks.

In Composer, callbacks modify trainer behavior and are called at the relevant Events in the training loop. This tutorial focuses on two callbacks, the EarlyStopper and ThresholdStopper, both of which halt training early depending on different criteria.

Recommended Background

This tutorial assumes that you’re generally familiar with techniques such as early stopping.

You should also be comfortable with the material in the Getting Started tutorial before you embark on this slightly more advanced tutorial.

Finally, you’ll probably have a better intuition for how the demonstrated features work if you brush up on Composer’s event-driven design in our Welcome Tour.

Tutorial Goals and Concepts Covered

The goal of this tutorial is to demonstrate a basic training run using one of our callbacks to control if/when training stops before the maximum training duration.

We’ll demonstrate:

• Basic training setup with Composer

• Using the EarlyStopper, and

• Using the ThresholdStopper

A comprehensive overview of Composer callbacks is outside the scope of this tutorial, but this should introduce you to some useful tools and give you a sense for the ways callbacks can be used to modify training behavior.

Let’s get started!

## Setup

In this tutorial, we’ll train a ComposerModel and halt training for criteria that we’ll set. We’ll use the same basic setup as in the Getting Started tutorial. If you want to better understand the details of the setup, that’s a good place to review.

Install Composer

First, install Composer if you haven’t already:

%pip install mosaicml
# To install from source instead of the last release, comment the command above and uncomment the following one.
# %pip install git+https://github.com/mosaicml/composer.git

Seed

Next, we’ll set the seed for reproducibility:

from composer.utils.reproducibility import seed_all

seed_all(42)

Dataloader Setup

Next, instantiate the training and evaluation datasets for CIFAR10

import torch.utils.data
from torchvision import datasets, transforms

data_directory = "./data"

# Normalization constants
mean = (0.507, 0.487, 0.441)
std = (0.267, 0.256, 0.276)

batch_size = 1024

cifar10_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize(mean, std)])

train_dataset = datasets.CIFAR10(data_directory, train=True, download=True, transform=cifar10_transforms)
eval_dataset = datasets.CIFAR10(data_directory, train=False, download=True, transform=cifar10_transforms)

# Setting shuffle=False to allow for easy overfitting in this example
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=False)
eval_dataloader = torch.utils.data.DataLoader(eval_dataset, batch_size=batch_size, shuffle=False)

Model, Optimizer, Scheduler, and Evaluator Setup

Finally, set up the model, optimizer, scheduler, and an evaluator.

from composer import models
from composer.optim import DecoupledSGDW, LinearWithWarmupScheduler
from composer.core import Evaluator

model = models.composer_resnet_cifar(model_name='resnet_56', num_classes=10)

optimizer = DecoupledSGDW(
    model.parameters(), # Model parameters to update
    lr=0.05, # Peak learning rate
    momentum=0.9,
    weight_decay=2.0e-3 # If this looks large, it's because its not scaled by the LR as in non-decoupled weight decay
)

lr_scheduler = LinearWithWarmupScheduler(
    t_warmup="1ep", # Warm up over 1 epoch
    alpha_i=1.0, # Flat LR schedule achieved by having alpha_i == alpha_f
    alpha_f=1.0
)

evaluator = Evaluator(
    dataloader = eval_dataloader,
    label = "eval",
    metric_names = ['Accuracy']
)

## EarlyStopper

The EarlyStopper callback tracks a particular training or evaluation metric and stops training if the metric does not improve within a given time interval.

The callback takes the following parameters:

• monitor: The name of the metric to track

• dataloader_label: This string identifies which dataloader the metric belongs to. By default, the train dataloader is labeled train, and the evaluation dataloader is labeled eval. (These names can be customized via the train_dataloader_label in the Trainer or the label argument of the Evaluator, respectively.)

• patience: The interval of the time that the callback will wait before stopping training if the metric is not improving. You can use integers to specify the number of epochs or provide a Time string—e.g., "50ba" or "2ep" for 50 batches and 2 epochs, respectively.

• min_delta: If non-zero, the change in the tracked metric over the patience window must be at least this large.

• comp: A comparison operator can be provided to measure the change in the monitored metric. The comparison operator will be called like comp(current_value, previous_best)

See the API Reference for more information.

Here, we’ll use our callback to track the Accuracy metric over one epoch on the test dataset:

from composer.callbacks import EarlyStopper

early_stopper = EarlyStopper(monitor="Accuracy", dataloader_label="eval", patience=1)

Now that we have our callback, we can instantiate a Composer trainer and train:

from composer.trainer import Trainer

# Early stopping should stop training before we reach 100 epochs!
train_epochs = "100ep"

trainer = Trainer(
    model=model,
    train_dataloader=train_dataloader,
    eval_dataloader=evaluator,
    max_duration=train_epochs,
    optimizers=optimizer,
    schedulers=lr_scheduler,
    callbacks=[early_stopper],    # Instruct the trainer to use our early stopping callback
    train_subset_num_batches=10,  # Only training on a subset of the data to trigger the callback sooner
)

# Train!
trainer.fit()

## ThresholdStopper

The ThresholdStopper callback is similar to the EarlyStopper, but it halts training when the metric crosses a threshold set in the ThresholdStopper callback.

This callback takes the following parameters:

• monitor, dataloader_label, and comp: Same as the EarlyStopper callback

• threshold: The float threshold that dictates when to halt training.

• stop_on_batch: If True, training can halt in the middle of an epoch, rather than just add the end.

We will reuse the same setup for the ThresholdStopper example.

from composer.callbacks import ThresholdStopper

threshold_stopper = ThresholdStopper("Accuracy", "eval", threshold=0.3)

# Threshold stopping should stop training before we reach 100 epochs!
train_epochs = "100ep"

trainer = Trainer(
    model=model,
    train_dataloader=train_dataloader,
    eval_dataloader=evaluator,
    max_duration=train_epochs,
    optimizers=optimizer,
    schedulers=lr_scheduler,
    callbacks=[threshold_stopper],  # Instruct the trainer to use our threshold stopper callback
    train_subset_num_batches=10,    # Only training on a subset of the data to trigger the callback sooner
)

# Train!
trainer.fit()

What next?

You’ve now seen how to implement early stopping in Composer using our EarlyStopper and ThresholdStopper callbacks.

To dig deeper into Composer callbacks check out the docs and our API references.

In addition, please continue to explore our tutorials! Here’s a couple suggestions:

• Continue learning about other Composer features like automatic gradient accumulation and automatic restarting from checkpoints

• Give your model life after training with Composer’s export for inference tools

Come get involved with MosaicML!

We’d love for you to get involved with the MosaicML community in any of these ways:

Star Composer on GitHub

Help make others aware of our work by starring Composer on GitHub.

Join the MosaicML Slack

Head on over to the MosaicML slack to join other ML efficiency enthusiasts. Come for the paper discussions, stay for the memes!

Contribute to Composer

Is there a bug you noticed or a feature you’d like? File an issue or make a pull request!