🏃‍♀️ Run Name

The run_name is a string used to name a specific training run. Naming your training run has many benefits. Namely, you can more easily group and keep track of metrics, checkpoints, and other training artifacts. In addition, your run_name will show up in many places as you utilize Composer.

Run Name Creation

The run_name argument is an optional argument to the Trainer.

There are two ways to get a run_name.

You create your own run_name and pass it to the trainer, like so:

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

from composer import Trainer
from composer.models import mnist_model

transform = transforms.Compose([transforms.ToTensor()])
dataset = datasets.MNIST("data", train=True, download=True, transform=transform)
train_dataloader = DataLoader(dataset, batch_size=128)
run_name = 'my-cool-run-name'

trainer = Trainer(
    model=mnist_model(num_classes=10),
    train_dataloader=train_dataloader,
    max_duration="2ep",
    run_name=run_name,
)
trainer.fit()

You can instead let the trainer create a run_name for you. The one created for you is the timestamp followed by a coolname; e.g. ‘1657932618-infrared-ferret’

How the Run Name is Used

This run_name will be added as an attribute to State and it is used by various other pieces of the composer infrastructure as described below.

The run_name is often used in the Composer as placeholder in a format string. This means that if a string is specified to name a file, like '{run_name}-foo-bar' for example, then that placeholder will get filled in by the actual run_name at runtime, so the file will actually be named ‘1657932618-infrared-ferret-foo-bar’.

Run Names in Checkpoint Saving

In checkpoint saving you can use the run_name as a placeholder in a format string to name the folders and checkpoints locally and in the cloud if you are uploading your checkpoints using Weights and Biases or an RemoteUploaderDownloader.

See CheckpointSaver for more information on specifying the arguments for files and folder names with the run_name when creating a Trainer object.

Run Names in Logging

In addition to checkpointing, loggers also use the run_name for default logging.

Experiment Tracking Loggers

• The TensorboardLogger will save all the logs for a run to a folder called run_name and the name of each run in the Tensorboard GUI will be run_name.

• The run_name you specify will be used by the WandBLogger as the run name for Weights and Biases.

• The run_name you specify will be used by the CometMLLogger as the run name for your Comet experiment.

Object Store Logger

The RemoteUploaderDownloader will often use the run_name as part of how it names objects.

File Logger

The run_name is also used in the FileLogger as the default name for the file the FileLogger logs to is '{run_name}/logs-rank{rank}.txt'

See Logging for more information.

Run Names in Profiling

The profiling tools for the training also save profiling files to folders named after the run_name. See Performance Profiling for more information.