Dvc: pkg: deploy model or dataset

[dmpetrov]

Currently dvc pkg install instantiates data artifacts in a current repository and cannot work outside of a repo.

In some cases, like production model deployment scenario, a repository is not needed. The only important thing is a data file content without any meta files (aka dvc-files).

We need to implement an option (like --flatten or --deploy) for this out-of-repository scenario.

Model deployment example:

$ mkdir /mnt/deploy
$ ls -a /mnt/deploy
.   ..
$ dvc pkg install --flatten -s model.pkl http://github.com/dmpetrov/new_tag_classifier /mnt/deploy/
$ ls -a /mnt/deploy
.   ..    model.pkl 

Download a dataset example:

$ dvc pkg install --flatten http://github.com/dmpetrov/imagenet .
$ ls -a
.   ..    imagenet/ 
$ du -sh imagenet
118G

Possible option names:

1. --flatten, -f - it might conflict with the current -f FILE, --file FILE.
2. --deploy, -d

[snagacarl]

For your command dvc pkg install --flatten -s model.pkl http://github.com/dmpetrov/new_tag_classifier /mnt/deploy/ how would the command specify the specific version of the model.pkl? Should it be included in the Git URL (http://github.com/dmpetrov/new_tag_classifier#v1.0) or maybe via arguments (--tag v1.0 or --branch prod)?

[dmpetrov]

Hi @snagacarl
The command was not released yet :) pease expect API changes

@efiop is working on the implementation which might change API.
I’d expect to have version in arguments. Version in URL might be a good addition to this.

[efiop]

Fixed by #2012 . Example command (can be ran outside of dvc repo):

dvc pkg get https://github.com/iterative/dvc scripts/innosetup/dvc.ico

Importing the whole package will be handled by https://github.com/iterative/dvc/issues/2140