Running Experiments

We explain how to execute DVC Experiments, setting their parameters, queueing them for future execution, running them in parallel, among other details.

If this is the first time you are introduced into data science experimentation, you may want to check the basics in Get Started: Experiments first.

`dvc.yaml` files

DVC relies on dvc.yaml files that contain the commands to run the experiment(s). These files codify pipelines that specify one or more stages of the experiment workflow (code, dependencies, outputs, etc.).

See Get Started: Data Pipelines for an intro to this topic.

Running the pipeline(s)

You can run the experiment pipelines using dvc exp run. It uses ./dvc.yaml (in the current directory) by default.

$ dvc exp run
...
Reproduced experiment(s): matte-vies

DVC observes the dependency graph between stages, so it only runs the ones with changed dependencies or outputs missing from the cache. You can limit this to certain reproduction targets or even single stages (--single-item flag).

DVC projects actually support more than one pipeline, in one or more dvc.yaml files. The --all-pipelines option lets you run them all at once.

dvc exp run is an experiment-specific alternative to dvc repro. dvc exp save can be used to capture experiments after executing ML processes manually.

Experiment results

The results of the last dvc exp run can be seen in the workspace. They are stored and tracked internally by DVC.

To display and compare multiple experiments along with their parameters and metrics, use dvc exp show or dvc exp diff. plots diff also accepts experiments as revisions. See Reviewing and Comparing Experiments for more details.

Use dvc exp apply to restore the results of any other experiment instead. See Bring experiment results to your workspace for more info.

Only files tracked by either Git or DVC are saved to the experiment. Untracked files cannot be restored.

Tuning (hyper)parameters

Parameters are any values used inside your code to tune modeling attributes, or that affect experiment results in any other way. For example, a random forest classifier may require a maximum depth value. Machine learning experimentation often involves defining and searching hyperparameter spaces to improve the resulting model metrics.

Your source code should read params from structured parameters files (params.yaml by default). Define them with the params field of dvc.yaml for DVC to track them. When a param value has changed, dvc exp run invalidates any stages that depend on it, and reproduces them.

See dvc params for more details.

You could manually edit a params file and run an experiment using those as inputs. Since this is a common sequence, the built-in option dvc exp run --set-param (-S) is provided as a shortcut. It takes an existing param name and value, and updates the file on-the-fly before execution.

$ cat params.yaml
model:
  learning_rate: 0.001
  units=64

$ dvc exp run --set-param model.learning_rate=0.0002
...

$ dvc exp run -S learning_rate=0.001 -S units=128  # set multiple params
...

See Hydra composition for more advanced configuration options via parameter overrides (change, append, or remove, or use "choice" sets and ranges).

The experiments queue

The --queue option of dvc exp run tells DVC to append an experiment for later execution. Nothing is actually run yet. Let's setup a simple hyperparameter grid search:

$ dvc exp run --queue -S units=10
Queued experiment '1cac8ca' for future execution.
$ dvc exp run --queue -S units=64
Queued experiment '23660bb' for future execution.
$ dvc exp run --queue -S units=128
Queued experiment '3591a5c' for future execution.
$ dvc exp run --queue -S units=256
Queued experiment '4109ead' for future execution.

How are experiments queued?

Queued experiments are managed using dvc-task and Celery.

Run them all with dvc queue start:

$ dvc queue start
...

In most cases, experiment tasks will be executed in the order that they were added to the queue (First In, First Out), but this is not guaranteed.

Their execution happens outside your workspace in temporary directories for isolation, so each experiment is derived from the workspace at the time it was queued.

Queued experiments are processed serially by default, but can be run in parallel by using more than one --jobs (to dvc queue start more than one worker).

Parallel runs (using --jobs > 1) are experimental and may be unstable. Make sure you're using number of jobs that your environment can handle (no more than the CPU cores).

Note that since queued experiments are run isolated from each other, common stages may be executed multiple times depending on the state of the run-cache at that time.

How are experiments isolated?

DVC creates a copy of the experiment's original workspace in .dvc/tmp/exps/ and runs it there. All workspaces share the single project cache, however.

💡 To isolate any experiment (without queuing it), you can use the --temp flag. This allows you to continue working while a long experiment runs, e.g.:

$ nohup dvc exp run --temp &
[1] 30473
nohup: ignoring input and appending output to 'nohup.out'

Note that Git-ignored files/dirs are excluded from queued/temp runs to avoid committing unwanted files into Git (e.g. once successful experiments are persisted). To include untracked files, stage them with git add first (before dvc exp run) and git reset them afterwards.

To clear the experiments queue and start over, use dvc queue remove --queued.

For more advanced grid searches, DVC supports complex config via Hydra composition.

Checkpoint experiments

To track successive steps in a longer or deeper experiment, you can register "checkpoints" from your code. These combine DVC Experiments with code logging. The latter can be achieved either with DVCLive, by using dvc.api.make_checkpoint() (Python code), or writing signal files (any programming language) following the same steps as make_checkpoint().

See Checkpoints to learn more about this feature.

Running checkpoint experiments is no different than running regular ones, e.g.:

$ dvc exp run -S param=value

All checkpoints registered at runtime will be preserved, even if the process gets interrupted (e.g. with Ctrl+C, or by an error). Without interruption, a "wrap-up" checkpoint will be added (if needed), so that changes to pipeline outputs don't remain in the workspace.

Subsequent uses of dvc exp run will resume from the latest checkpoint (using the latest cached versions of all outputs). To resume from a previous checkpoint (list them with dvc exp show), you must first dvc exp apply it before using dvc exp run. For --queue or --temp runs, use --rev to specify the checkpoint to resume from.

Alternatively, use --reset to start over (discards previous checkpoints and their outputs). This is useful for re-training ML models, for example.

Note that queuing an experiment that uses checkpoints implies --reset, unless a --rev is provided (refer to the previous section).