Edit on GitHub

exp run

Run or resume a DVC experiment based on a DVC pipeline.

When called with no arguments, this is equivalent to dvc repro followed by dvc exp save.

Synopsis

usage: dvc exp run [-h] [-q | -v] [-f] [-i]
                   [-s] [-p] [-P] [-R]
                   [-n <name>] [-S [<filename>:]<override_pattern>]
                   [--queue] [--run-all] [-j <number>] [--temp]
                   [-r <experiment_rev>] [-C <path>]
                   [-m <message>]
                   [--downstream] [--force-downstream]
                   [--pull] [--dry] [--allow-missing]
                   [-k] [--ignore-errors]
                   [targets [targets ...]]

positional arguments:
  targets               Stages to reproduce. 'dvc.yaml' by default

Description

Executes and tracks experiments in your repository without polluting it with unnecessary Git commits, branches, directories, etc.

Only files tracked by either Git or DVC are saved to the experiment. See dvc exp save --include-untracked for an alternative.

dvc exp run has the same general behavior as dvc repro when it comes to targets and stage execution (restores the dependency graph, etc.).

This includes committing any changed data dependencies to the DVC cache when preparing the experiment, which can take some time.

Use the --set-param (-S) option as a shortcut to change parameter values on-the-fly before running the experiment.

It's possible to queue experiments for later execution with the --queue flag. Queued experiments can be run with dvc queue start and managed with other dvc queue commands.

See the Running Experiments guide for more details on these features and more.

Review your experiments with dvc exp show. Successful ones can be made persistent by restoring them via dvc exp branch or dvc exp apply and committing them to the Git repo. Unnecessary ones can be cleared with dvc exp remove.

Options

  • -S [<filename>:]<override_pattern>, --set-param [<filename>:]<override_pattern> - set the value of dvc params for this experiment. This will update the parameters file (params.yaml by default) before running the experiment. Use the optional [<filename>:] prefix to use a custom params file.

    Valid <override_pattern> values can be defined in Hydra's basic override syntax (see example). Hydra's choice and range sweep overrides are also supported, but these require the --queue flag to be provided as well (see example).

  • -n <name>, --name <name> - specify a unique name for this experiment. A default one will be generated otherwise, such as puffy-daks.

    The name of the experiment is exposed in env var DVC_EXP_NAME.

  • --temp - run this experiment outside your workspace (in .dvc/tmp/exps). Useful to continue working (e.g. in another terminal) while a long experiment runs.

  • --queue - place this experiment at the end of a line for future execution, but don't run it yet. Use dvc queue start to process the queue.

  • --run-all - run all queued experiments (see --queue) and outside your workspace (in .dvc/tmp/exps). Use -j to execute them in parallel.

    dvc exp run --run-all [--jobs] is now a shortcut for dvc queue start [--jobs] followed by dvc queue logs -f. The --run-all and --jobs options will be deprecated in a future DVC release.

  • -j <number>, --jobs <number> - run this number of queued experiments in parallel. Only has an effect along with --run-all. Defaults to 1 (the queue is processed serially).

  • -f, --force - reproduce pipelines even if no changes were found (same as dvc repro -f).

  • -C <path>, --copy-paths <path> - list of ignored or untracked paths to copy into the temp directory. Only used if --temp or --queue is specified.

  • -m <message>, --message <message> - custom message to use when saving the experiment. If not provided, dvc: commit experiment {hash} will be used.

  • -i, --interactive - ask for confirmation before reproducing each stage. The stage is only executed if the user types "y".

  • -s, --single-item - reproduce only a single stage by turning off the recursive search for changed dependencies. Multiple stages are executed (non-recursively) if multiple stage names are given as targets.

  • -p, --pipeline - reproduce the entire pipelines that the targets belong to. Use dvc dag <target> to show the parent pipeline of a target.

  • -P, --all-pipelines - reproduce all pipelines for all dvc.yaml files present in the DVC project. Specifying targets has no effects with this option, as all possible targets are already included.

  • -R, --recursive - looks for dvc.yaml files to reproduce in any directories given as targets, and in their subdirectories. If there are no directories among the targets, this option has no effect.

  • --downstream - only execute the stages after the given targets in their corresponding pipelines, including the target stages themselves. This option has no effect if targets are not provided.

  • --force-downstream - in cases like ... -> A (changed) -> B -> C it will reproduce A first and then B, even if B was previously executed with the same inputs from A (cached). To be precise, it reproduces all descendants of a changed stage or the stages following the changed stage, even if their direct dependencies did not change.

    It can be useful when we have a common dependency among all stages, and want to specify it only once (for stage A here). For example, if we know that all stages (A and below) depend on requirements.txt, we can specify it in A, and omit it in B and C.

    This is a way to force-execute stages without changes. This can also be useful for pipelines containing stages that produce non-deterministic (semi-random) outputs, where outputs can vary on each execution, meaning the cache cannot be trusted for such stages.

  • --pull - attempts to download missing data as needed. This includes (1) dependencies of stages to be run, (2) outputs of otherwise unchanged stages to be skipped, (3) [run cache] for stages to be checked out from cache (unless --no-run-cache is passed).

  • --allow-missing - skip stages with no other changes than missing data.

    In DVC>=3.0, --allow-missing will not skip data saved with DVC<3.0 because the hash type changed in DVC 3.0, which DVC considers a change to the data. To migrate data to the new hash type, run dvc cache migrate --dvc-files. See more information about upgrading from DVC 2.x to 3.0.

  • -k, --keep-going - Continue executing, skipping stages having dependencies on the failed stage. The other dependencies of the targets will still be executed.

  • --ignore-errors - Ignore all errors when executing the stages. Unlike --keep-going, stages having dependencies on the failed stage will be executed.

  • -h, --help - prints the usage/help message, and exits.

  • -q, --quiet - do not write anything to standard output. Exit with 0 if all stages are up to date or if all stages are successfully executed, otherwise exit with 1. The command defined in the stage is free to write output regardless of this flag.

  • -v, --verbose - displays detailed tracing information.

Examples

This example is based on our Get Started, where you can find the actual source code.

Clone the DVC repo and download the data it depends on:

$ git clone git@github.com:iterative/example-get-started.git
$ cd example-get-started
$ dvc pull

Let's also install the Python requirements:

We strongly recommend creating a virtual environment first.

$ pip install -r src/requirements.txt

Let's check the latest metrics of the project:

$ dvc metrics show
Path         avg_prec    roc_auc
scores.json  0.60405     0.9608

For this experiment, we want to see the results for a smaller dataset input, so let's limit the data to 20 MB and reproduce the pipeline with dvc exp run:

$ truncate --size=20M data/data.xml
$ dvc exp run
...
Reproduced experiment(s): puffy-daks
Experiment results have been applied to your workspace.

$ dvc metrics diff
Path         Metric    HEAD     workspace  Change
scores.json  avg_prec  0.60405  0.56103    -0.04302
scores.json  roc_auc   0.9608   0.94003    -0.02077

The dvc metrics diff command shows the difference in performance for the experiment we just ran (puffy-daks).

Example: Modify parameters on-the-fly

dvc exp run --set-param (-S) saves you the need to manually edit a params file (see dvc params) before running an experiment.

This option accepts Hydra's basic override syntax. For example, it can override (train.epochs=10), append (+train.weight_decay=0.01), or remove (~model.dropout) parameters:

dvc exp run -S 'prepare.split=0.1' -S 'featurize.max_features=100'
...

Note that you can modify multiple parameters at once in the same command.

By default, -S overwrites the values in params.yaml. To use another params file, add a <filename>: prefix. For example, let's append a new parameter to train_config.json:

$ dvc exp run -S 'train_config.json:+train.weight_decay=0.001'
...

$ dvc params diff --targets train_config.json
Path               Param                HEAD    workspace
train_config.json  train.weight_decay   -       0.001

Warnings

exp run --set-param (-S) doesn't update your dvc.yaml to start or stop tracking parameters. When appending or removing params, check if you need to update the params section accordingly.

Similarly, when using custom param files, check that these are defined in dvc.yaml.

Combining --set-param and --queue, we can perform a grid search for tuning hyperparameters.

DVC supports Hydra's syntax for choice and range sweeps to add multiple experiments to the queue. These can be used for multiple parameters at the same time, adding all combinations to the queue:


$ dvc exp run -S 'train.min_split=8,64' -S 'train.n_est=range(100,500,100)' --queue
Queueing with overrides '{'params.yaml': ['train.min_split=8', 'train.n_est=100']}'.
Queued experiment 'azure-ices' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=8', 'train.n_est=200']}'.
Queued experiment 'zingy-peri' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=8', 'train.n_est=300']}'.
Queued experiment 'jammy-feds' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=8', 'train.n_est=400']}'.
Queued experiment 'lowse-shay' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=64', 'train.n_est=100']}'.
Queued experiment 'brown-hugs' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=64', 'train.n_est=200']}'.
Queued experiment 'local-scud' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=64', 'train.n_est=300']}'.
Queued experiment 'alpha-neck' for future execution.
Queueing with overrides '{'params.yaml': ['train.min_split=64', 'train.n_est=400']}'.
Queued experiment 'algal-hood' for future execution.
$ dvc queue start
...

Example: Only pull pipeline data as needed.

You can combine the --pull and --allow-missing flags to reproduce a pipeline while only pulling the data that is actually needed to run the changed stages.

Given the pipeline used in example-get-started-experiments:

$ dvc dag
      +--------------------+
      | data/pool_data.dvc |
      +--------------------+
                 *
                 *
                 *
          +------------+
          | data_split |
          +------------+
           **         **
         **             **
        *                 **
  +-------+                 *
  | train |*                *
  +-------+ ****            *
      *         ***         *
      *            ****     *
      *                **   *
+-----------+         +----------+
| sagemaker |         | evaluate |
+-----------+         +----------+

If we are in a machine where all the data is missing:

$ dvc status
data_split:
        changed deps:
                deleted:            data/pool_data
        changed outs:
                not in cache:       data/test_data
                not in cache:       data/train_data
train:
        changed deps:
                deleted:            data/train_data
        changed outs:
                not in cache:       models/model.pkl
                not in cache:       models/model.pth
                not in cache:       results/train
evaluate:
        changed deps:
                deleted:            data/test_data
                deleted:            models/model.pkl
        changed outs:
                not in cache:       results/evaluate
sagemaker:
        changed deps:
                deleted:            models/model.pth
        changed outs:
                not in cache:       model.tar.gz
data/pool_data.dvc:
        changed outs:
                not in cache:       data/pool_data

We can modify the evaluate stage and DVC will only pull the necessary data to run that stage (models/model.pkl data/test_data/) while skipping the rest of the stages:

$ dvc exp run --pull --allow-missing
Reproducing experiment 'hefty-tils'
'data/pool_data.dvc' didn't change, skipping
Stage 'data_split' didn't change, skipping
Stage 'train' didn't change, skipping
Running stage 'evaluate':
...

See pull missing data in the user guide for more details.

Example: Include untracked or ignored paths

If your code relies on some paths that are intentionally untracked or ignored by Git, you can use -C/--copy-paths to ensure those files are accessible when you use the --temp or --queue flags:

$ dvc exp run --temp -C secrets.txt -C symlinked-directory

The paths will be copied to the temporary directory but will not be tracked, to prevent unintentional leaks.