Look here every month for great discussions and technical Q&A's from our users and core development team.
Here are some Q&A’s from our Discord channel that we think are worth sharing.
dvc metrics diff [a_rev] [b_rev], substituting hashes, branches, or tags for [a_rev] and [b_rev]. But what if I wanted to see the metrics for a list of tags?
DVC has a built in function for this! You can use
dvc metrics show with the
$ dvc metrics show -T
to list the metrics for all tagged experiments.
Also, we have a couple of relevant discussions going on in our GitHub repo about handling experiments and hyperparameter tuning. Feel free to join the discussion and let us know what kind of support would help you most.
.dvcfile? In particular, I’d like to save summary statistics (e.g., mean, minimum, and maximum) about my data.
One simple way to keep metadata in a
.dvc file is by using the
meta entry is a
key:value pair (for example,
name: Jean-Luc). The
meta field can be manually added or written programmatically, but note that if
.dvc file is overwritten (perhaps by
dvc add, or
dvc import) these values will not be preserved. You can read more about this
in our docs.
Another approach would be to track the statistics of your dataset in a metric file, just as you might track performance metrics of a model. For a tutorial on using DVC metrics please see our docs.
ERROR: unexpected error - /my-folder is not a git repository. What’s going on?
This is a consequence of new support we’ve added for monorepos with the
dvc init --subdir functionality
(see more here), which lets
there be multiple DVC projects within a single Git repository. Now, if a DVC
repository doesn’t contain a
.git directory, DVC expects the
no_scm flag to
be present in
.dvc/config and raises an error if not. For example, one of our
users reported this when using DVC to pull files into a Docker container that
didn’t have Git initialized (for more about using DVC without Git,
see our docs).
You can fix this by running
dvc config core.no_scm true (you could include
this command in the script that creates Docker images). Alternately, you could
.git in your Docker container, but this is not advisable for all
We are currently working to add graceful error-handling for this particular issue so stay tuned.
dvc repro has a flag that should help here. You can use the
--force flag to reproduce the pipeline even when no changes in the
dependencies (for example, a training datset tracked by DVC) have been found. So
if you had a hypoethetical DVC pipeline whose final process was
you could run
dvc repro -f deploy.dvc to rerun the whole pipeline.
.dvcfiles corresponding to different remotes?
Yes, one project directory can contain datasets from several different DVC
remotes. Specifically, DVC has functions
dvc import and
dvc get that emulate
the experience of using a package manager for grabbing datasets from external
sources. You can use
dvc import or
dvc get to access any number of datasets
that are dependencies in a given project. For more on this,
see our tutorial on data registries.
DVC doesn’t collect any information about your data (or code, or models, for that matter). You may have noticed that DVC collects Anonymized Usage Analytics, which users may opt out of. The data we collect is extremely limited and anonymized, as it is collected mainly for the purpose of prioritizing bugs and feature development based on DVC usage. For example, we collect info about your operating system, DVC version, and installation method (the complete list of collected features is here).
Many of our users work with sensitive or private data, and we’ve developed DVC with such scenarios in mind from day one.
Increasingly, DVC is being used not to just to version and manage machine learning projects, but as part of MLOps, practices for combining data science and software engineering. As MLOps is a fairly new discipline, standards and references aren’t yet solidified. So while there isn’t (yet) a standard recipe for using DVC in MLOps projects, we can point you to a few architectures we like, and which have been reported in sufficient detail to recreate.
First, DVC can be used to detect events (such as dataset changes) in a CI/CD system that traditional version control systems might not be able to. An excellent and thorough blog by Danilo Sato et al. explores using DVC in this way, as part of a CI/CD system that retrains a model automatically when changes in the dataset are detected.
Second, DVC can be used to support model training on cloud GPUs, particularly as a tool for pushing and pulling files (such as datasets and trained models) between cloud computing instances, DVC repositories, and other environments. This architecture was the subject of a recent blog by Marcel Mikl and Bert Besser. Their report describes the cloud computing setup and continuous integration pipeline quite well.
If you develop your own architecture for using DVC in MLOps, please keep us posted. We’ll be eager to learn from your experience. Also, keep an eye on our blog in the next few months. We’re rolling out some new tools with a focus on MLOps!