Edit on GitHub

import

Download a file or directory tracked by another DVC or Git repository into the workspace, and track it (an import .dvc file is created).

See also our dvc.api.open() Python API function.

Synopsis

usage: dvc import [-h] [-q | -v] [--file <filename>]
                  [-o <path>] [--rev <commit>]
                  [--no-exec | --no-download]
                  [-j <number>] [--desc <text>]
                  [--type <str>] [--label <str>] [--meta key=value]
                  url path

positional arguments:
  url              Location of DVC or Git repository to download from
  path             Path to a file or directory within the repository

Description

Provides an easy way to reuse files or directories tracked in any DVC repository (e.g. datasets, intermediate results, ML models) or Git repository (e.g. code, small image/other files). dvc import downloads the target file or directory (found at path in url), and tracks it in the local project. This makes it possible to update the import later, if the data source has changed (see dvc update).

See dvc list for a way to browse repository contents to find files or directories to import.

Note that dvc get corresponds to the first step this command performs (just download the data).

The imported data is cached, and linked (or copied) to the current working directory with its original file name e.g. data.txt (or to a location provided with --out). An import .dvc file is created in the same location e.g. data.txt.dvc – similar to using dvc add after downloading the data.

DVC won't push data imported from other DVC repos to remote storage. dvc pull will download from the original source.

The url argument specifies the address of the DVC or Git repository containing the data source. Both HTTP and SSH protocols are supported (e.g. [user@]server:project.git). url can also be a local file system path (including the current project e.g. .).

The path argument specifies a file or directory to download (paths inside tracked directories are supported). It should be relative to the root of the repo (absolute paths are supported when url is local). Note that DVC-tracked targets must be found in a dvc.yaml or .dvc file of the repo.

See dvc import-url to download and track data from other supported locations such as S3, SSH, HTTP, etc.

.dvc files support references to data in an external DVC repository (hosted on a Git server). In such a .dvc file, the deps field specifies the url and data path, and the outs field contains the corresponding local path in the workspace. It records enough metadata about the imported data to enable DVC to efficiently determine whether the local copy is out of date.

To actually version the data, git add (and git commit) the import .dvc file.

⚠️ Relevant notes and limitation:

  • Source DVC repos should have a default DVC remote containing the target data for this command to work.
  • The only exception to the above requirement is for local repos, where DVC will try to copy the data from its cache first.
  • Limited support for chained imports is available (importing data that was itself imported into the source repo from another one).
  • Note that dvc repro doesn't check or update import .dvc files (see dvc freeze), use dvc update to bring the import up to date from the data source.

Options

  • -o <path>, --out <path> - specify a path to the desired location in the workspace to place the downloaded file or directory (instead of using the current working directory).

  • --file <filename> - specify a path and/or file name for the .dvc file created by this command (e.g. --file stages/stage.dvc). This overrides the default file name: <file>.dvc, where <file> is the desired file name of the imported data (out).

  • --rev <commit> - commit hash, branch or tag name, etc. (any Git revision) of the repository to download the file or directory from. The latest commit in master (tip of the default branch) is used by default when this option is not specified.

    Note that this adds a rev field in the import .dvc file that fixes it to the revision. This can impact the behavior of dvc update (see the Importing and updating fixed revisions example below).

  • --no-exec - create the import .dvc file without accessing url (assumes that the data source is valid). This is useful if you need to define the project imports quickly, and import the data later (use dvc update to finish the operation(s)).

  • --no-download - create the import .dvc file including the source data information (repository URL and version) but without downloading the associated data. This is useful if you need track changes in remote data without using local storage space (yet). The data can be downloaded later using dvc pull. File version can be updated using dvc update --no-download.

  • -j <number>, --jobs <number> - parallelism level for DVC to download data from the remote. The default value is 4 * cpu_count(). Using more jobs may speed up the operation. Note that the default value can be set in the source repo using the jobs config option of dvc remote modify.

  • --desc <text> - user description of the data.

  • --type <str> - user-assigned type of the data.

  • --label <text> - user-assigned label(s) to add to the data.

  • --meta key=value - custom metadata to add to the data.

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

  • -q, --quiet - do not write anything to standard output. Exit with 0 if no problems arise, otherwise 1.

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

Examples

A simple case for this command is to import a dataset from an external DVC repository, such as our get started example repo.

$ dvc import git@github.com:iterative/example-get-started \
             data/data.xml
Importing 'data/data.xml (git@github.com:iterative/example-get-started)'
-> 'data.xml'

In contrast with dvc get, this command doesn't just download the data file, but it also creates an import .dvc file with a link to the data source (as explained in the description above). (This .dvc file can later be used to update the import.) Check data.xml.dvc:

md5: 7de90e7de7b432ad972095bc1f2ec0f8
frozen: true
wdir: .
deps:
  - path: data/data.xml
    repo:
      url: git@github.com:iterative/example-get-started
      rev_lock: 6c73875a5f5b522f90b5afa9ab12585f64327ca7
outs:
  - md5: a304afb96060aad90176268345e10355
    path: data.xml
    cache: true

Several of the values above are pulled from the original .dvc file in the external DVC repository. The url and rev_lock subfields under repo are used to save the origin and version of the dependency, respectively.

Example: Importing and updating fixed revisions

To import a specific version of a file/directory, we may use the --rev option:

$ dvc import --rev cats-dogs-v1 \
             git@github.com:iterative/dataset-registry.git \
             use-cases/cats-dogs
Importing
'use-cases/cats-dogs (git@github.com:iterative/dataset-registry.git)'
-> 'cats-dogs'

When using this option, the import .dvc file will also have a rev subfield under repo:

deps:
  - path: use-cases/cats-dogs
    repo:
      url: git@github.com:iterative/dataset-registry.git
      rev: cats-dogs-v1
      rev_lock: 0547f5883fb18e523e35578e2f0d19648c8f2d5c

If rev is a Git branch or tag (where the underlying commit changes), the data source may have updates at a later time. To bring it up to date if so (and update rev_lock in the .dvc file), simply use dvc update <stage>.dvc. If rev is a specific commit hash (does not change), dvc update without options will not have an effect on the import .dvc file. You may force-update it to a different commit with dvc update --rev:

$ dvc update --rev cats-dogs-v2 cats-dogs.dvc

In the above example, the value for rev in the new .dvc file will be master (a branch) so it will be able update normally going forward.

Example: Data registry

If you take a look at our dataset registry project, you'll see that it's organized into different directories such as tutorials/versioning and use-cases/, and these contain .dvc files that track different datasets. Given this simple structure, its data files can be easily shared among several other projects using dvc get and dvc import. For example:

$ dvc get https://github.com/iterative/dataset-registry \
          tutorials/versioning/data.zip

Used in our versioning tutorial

Or

$ dvc import git@github.com:iterative/dataset-registry.git \
             use-cases/cats-dogs

dvc import provides a better way to incorporate data files tracked in external DVC repositories because it saves the connection between the current project and the source repo. This means that enough information is recorded in an import .dvc file in order to reproduce downloading of this same data version in the future, where and when needed. This is achieved with the repo field, for example (matching the import command above):

frozen: true
deps:
  - path: use-cases/cats-dogs
    repo:
      url: git@github.com:iterative/dataset-registry.git
      rev_lock: 0547f5883fb18e523e35578e2f0d19648c8f2d5c
outs:
  - md5: b6923e1e4ad16ea1a7e2b328842d56a2.dir
    path: cats-dogs
    cache: true

See a full explanation in our Data Registry use case.

Example: Importing from any Git repository

You can even import files from plain Git repos that are not DVC repositories. For example, let's import a dataset from GSA's data repo:

$ dvc import git@github.com:GSA/data \
           enterprise-architecture/it-standards.csv
Importing ...

Note that Git-tracked files can be imported from DVC repos as well.

The file is imported, and along with it, an import .dvc file is created. Check it-standards.csv.dvc:

deps:
  - path: enterprise-architecture/it-standards.csv
    repo:
      url: git@github.com:GSA/data
      rev_lock: af6a1feb542dc05b4d3e9c80deb50e6596876e5f
outs:
  - md5: 7e6de779a1ab286745c808f291d2d671
    path: it-standards.csv

The url and rev_lock subfields under repo are used to save the origin and version of the dependency, respectively.

Example: Chained imports

DVC supports importing data that was itself imported into the source repo, as long as all the repos in the import chain (and their default remote storage) are accessible from the final destination repo.

Consider an example with 3 DVC repos (A, B, and C). DVC repo /repo/a contains a data.csv file tracked with dvc add:

/repo/a
├── data.csv
└── data.csv.dvc

In repo B, we import data.csv from A and into a subdirectory:

$ dvc import /repo/a data.csv --out training/data.csv

Project B may of course contain other files unique to itself, for example:

/repo/b
└── training
    ├── data.csv
    ├── data.csv.dvc
    ├── labels
    │   ├── test.txt
    │   └── truth.txt
    └── labels.dvc

Notice that the training/labels directory (not an import) is also tracked in B separately.

If we examine training/data.csv.dvc, we can see that that the import source is repo A (/repo/a):

deps:
  - path: data.csv
    repo:
      url: /repo/a
      rev_lock: 32ab3ddc8a0b5cbf7ed8cb252f93915a34b130eb
outs:
  - md5: acbd18db4cc2f85cedef654fccc4a4d8
    size: 3234523
    path: data.csv

Now lets imagine that we run the following command in our third repo, C:

$ dvc import /repo/b training

This will result in the following directory structure, which contains a chained import and a regular one:

/repo/c
├── training
│   ├── data.csv
│   └── labels
│       ├── test.txt
│       └── truth.txt
└── training.dvc
  • training/data.csv is imported from A into B into C
  • training/labels/ is imported from B into C directly

However, training.dvc only references repo B (/repo/b):

deps:
  - path: training
    repo:
      url: /repo/b
      rev_lock: 15136ed84b59468b68fd66b8141b41c5be682ced
outs:
  - md5: e784c380dd9aa9cb13fbe22e62d7b2de.dir
    size: 27
    nfiles: 3
    path: training

Each time that we dvc import or dvc update* training/ into C (or even dvc pull it) DVC will first look up the contents of training in B and notice that training/data.csv is itself an import. It will then resolve the chain as needed (finding data.csv in A).

*Note that when running dvc update training from repo C, DVC will only check whether or not training/ changed in repo B. So if data.csv has only changed in A, training/data.csv won't be updated in C until dvc update training/data.csv has been run in B.

This means both repos A and B must be reachable when dvc import runs in repo C, otherwise the import chain resolution would fail.

The default DVC remotes for all repos in the import chain must also be accessible (repo C needs to have all the appropriate credentials).