Running Experiments

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

Running experiment commands and pipelines

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.).

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.

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. dvc 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.

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.

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
...
train:
  valid_pct: 0.1
  arch: shufflenet_v2_x2_0
  img_size: 256
  batch_size: 8
  fine_tune_args:
    epochs: 8
    base_lr: 0.01
...

$ dvc exp run --set-param train.fine_tune_args.base_lr=0.001
...

$ dvc exp run -S train.img_size=1024 -S train.batch_size=512  # set multiple params
...

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 queue a couple experiments:

$ dvc exp run --queue -S train.fine_tune_args.base_lr=0.001
Queueing with overrides '{'params.yaml': ['train.fine_tune_args.base_lr=0.001']}'.
Queued experiment 'blowy-pail' for future execution.
$ dvc exp run --queue -S train.fine_tune_args.base_lr=0.002
Queueing with overrides '{'params.yaml': ['train.fine_tune_args.base_lr=0.002']}'.
Queued experiment 'nubby-gram' for future execution.

Alternatively, we can setup a hyperparameter grid search:

$ dvc exp run --queue \
    -S train.arch='resnet18,shufflenet_v2_x2_0' \
    -S 'train.fine_tune_args.base_lr=range(0.001, 0.01, 0.001)'
Queueing with overrides '{'params.yaml': ['train.arch=resnet18', 'train.fine_tune_args.base_lr=0.001']}'.
Queued experiment 'bijou-chis' for future execution.
Queueing with overrides '{'params.yaml': ['train.arch=resnet18', 'train.fine_tune_args.base_lr=0.002']}'.
Queued experiment 'color-meal' for future execution.
Queueing with overrides '{'params.yaml': ['train.arch=resnet18', 'train.fine_tune_args.base_lr=0.003']}'.
Queued experiment 'fusil-chin' for future execution.
...
Queueing with overrides '{'params.yaml': ['train.arch=shufflenet_v2_x2_0', 'train.fine_tune_args.base_lr=0.001']}'.
Queued experiment 'lumpy-jato' for future execution.
Queueing with overrides '{'params.yaml': ['train.arch=shufflenet_v2_x2_0', 'train.fine_tune_args.base_lr=0.002']}'.
Queued experiment 'gypsy-wino' for future execution.
Queueing with overrides '{'params.yaml': ['train.arch=shufflenet_v2_x2_0', 'train.fine_tune_args.base_lr=0.003']}'.
...

Run them all with dvc queue start:

$ dvc queue start
...

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).

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