clearml-docs/docs/clearml_data.md
2021-09-01 12:48:30 +03:00

17 KiB

title
ClearML Data

In Machine Learning, you are very likely dealing with a gargantuan amount of data that you need to put in a dataset, which you then need to be able to share, reproduce, and track.

ClearML Data Management solves two important challenges:

  • Accessibility - Making data easily accessible from every machine,
  • Versioning - Linking data and experiments for better traceability.

We believe Data is not code. It should not be stored in a git tree, because progress on datasets is not always linear. Moreover, it can be difficult and inefficient to find on a git tree the commit associated with a certain version of a dataset.

A clearml-data dataset is a collection of files, stored on a central storage location (S3 / GS / Azure / Network Storage). Datasets can be set up to inherit from other datasets, so data lineages can be created, and users can track when and how their data changes.

Dataset changes are stored using differentiable storage, meaning a version will store the change-set from its previous dataset parents.

Local copies of datasets are always cached, so the same data never needs to be downloaded twice. When a dataset is pulled it will automatically pull all parent datasets and merge them into one output folder for you to work with.

ClearML Data offers two interfaces:

  • clearml-data - CLI utility for creating, uploading, and managing datasets.
  • clearml.Dataset - A python interface for creating, retrieving, managing, and using datasets.

Setup

clearml-data comes built-in with our clearml python package! Just check out the getting started guide for more info!

Workflow

Below is an example of a workflow using ClearML Data's command line tool to create a dataset and inegrating the dataset into code using ClearML Data's python interface.

Creating a Dataset

Using the clearml-data CLI, users can create datasets using the following commands:

clearml-data create --project dataset_example --name initial_version
clearml-data add --files data_folder
clearml-data close

The commands will do the following:

  1. Start a Data Processing Task called "initial_version" in the "dataset_example" project

  2. The CLI will return a unique ID for the dataset

  3. All the files from the "data_folder" folder will be added to the dataset and uploaded by default to the ClearML server.

  4. The dataset will be finalized, making it immutable and ready to be consumed.

:::note clearml-data is stateful and remembers the last created dataset so there's no need to specify a specific dataset ID unless we want to work on another dataset. :::

Using a Dataset

Now in our python code, we can access and use the created dataset from anywhere:

from clearml import Dataset

local_path = Dataset.get(dataset_id='dataset_id_from_previous_command').get_local_copy()

We have all our files in the same folder structure under local_path, it is that simple!

The next step is to set the dataset_id as a parameter for our code and voilà! We can now train on any dataset we have in the system.

CLI Options

It's possible to manage datasets (create / modify / upload / delete) with the clearml-data command line tool.

Creating a Dataset

clearml-data create --project <project_name> --name <dataset_name> --parents <existing_dataset_id>`

Creates a new dataset.

Parameters

Name Description Optional
name Dataset's name No
project Dataset's project No
parents IDs of the dataset's parents. The dataset inherits all of its parents' content. Multiple parents can be entered, but they are merged in the order they were entered Yes
tags Dataset user tags. The dataset can be labeled, which can be useful for organizing datasets Yes

:::important clearml-data works in a stateful mode so once a new dataset is created, the following commands do not require the --id flag. :::


Add Files

clearml-data add --id <dataset_id> --files <filenames/folders_to_add>

It's possible to add individual files or complete folders.

Parameters

Name Description Optional
id Dataset's ID. Default: previously created / accessed dataset Yes
files Files / folders to add. Wildcard selection is supported, for example: ~/data/*.jpg ~/data/json No
dataset-folder Dataset base folder to add the files to in the dataset. Default: dataset root Yes
non-recursive Disable recursive scan of files Yes
verbose Verbose reporting Yes

Remove Files

clearml-data remove --id <dataset_id_to_remove_from> --files <filenames/folders_to_remove>

Parameters

Name Description Optional
id Dataset's ID. Default: previously created / accessed dataset Yes
files Files / folders to remove (wildcard selection is supported, for example: ~/data/*.jpg ~/data/json). Notice: file path is the path within the dataset, not the local path. No
non-recursive Disable recursive scan of files Yes
verbose Verbose reporting Yes

Upload Dataset Content

clearml-data upload [--id <dataset_id>] [--storage <upload_destination>]

Uploads added files to ClearML Server by default. It's possible to specify a different storage medium by entering an upload destination, such as s3://bucket, gs://, azure://, /mnt/shared/.

Parameters

Name Description Optional
id Dataset's ID. Default: previously created / accessed dataset Yes
storage Remote storage to use for the dataset files. Default: files_server Yes
verbose Verbose reporting Yes

Finalize Dataset

clearml-data close --id <dataset_id>

Finalizes the dataset and makes it ready to be consumed. It automatically uploads all files that were not previously uploaded. Once a dataset is finalized, it can no longer be modified.

Parameters

Name Description Optional
id Dataset's ID. Default: previously created / accessed dataset Yes
storage Remote storage to use for the dataset files. Default: files_server Yes
disable-upload Disable automatic upload when closing the dataset Yes
verbose Verbose reporting Yes

Sync Local Folder

clearml-data sync [--id <dataset_id] --folder <folder_location>  [--parents '<parent_id>']`

This option syncs a folder's content with ClearML. It is useful in case a user has a single point of truth (i.e. a folder) which updates from time to time.

Once an update should be reflected into ClearML's system, users can call clearml-data sync, create a new dataset, enter the folder, and the changes (either file addition, modification and removal) will be reflected in ClearML.

This command also uploads the data and finalizes the dataset automatically.

Parameters

Name Description Optional
id Dataset's ID. Default: previously created / accessed dataset Yes
folder Local folder to sync. Wildcard selection is supported, for example: ~/data/*.jpg ~/data/json No
storage Remote storage to use for the dataset files. Default: files_server Yes
parents IDs of the dataset's parents (i.e. merge all parents). All modifications made to the folder since the parents were synced will be reflected in the dataset Yes
project If creating a new dataset, specify the dataset's project name Yes
name If creating a new dataset, specify the dataset's name Yes
tags Dataset user tags Yes
skip-close Do not auto close dataset after syncing folders Yes
verbose Verbose reporting Yes

List Dataset Content

clearml-data list [--id <dataset_id>]

Parameters

Name Description Optional
id Dataset ID whose contents will be shown (alternatively, use project / name combination). Default: previously accessed dataset Yes
project Specify dataset project name (if used instead of ID, dataset name is also required) Yes
name Specify dataset name (if used instead of ID, dataset project is also required) Yes
filter Filter files based on folder / wildcard. Multiple filters are supported. Example: folder/date_*.json folder/sub-folder Yes
modified Only list file changes (add / remove / modify) introduced in this version Yes

Delete Dataset

clearml-data delete [--id <dataset_id_to_delete>]

Deletes an entire dataset from ClearML. This can also be used to delete a newly created dataset.

This does not work on datasets with children.

Parameters

Name Description Optional
id ID of dataset to be deleted. Default: previously created / accessed dataset that hasn't been finalized yet Yes
force Force dataset deletion even if other dataset versions depend on it Yes

Search for a Dataset

clearml-data search [--name <name>] [--project <project_name>] [--tags <tag>]

Lists all datasets in the system that match the search request.

Datasets can be searched by project, name, ID, and tags.

Parameters

Name Description Optional
ids A list of dataset IDs
project The project name of the datasets
name A dataset name or a partial name to filter datasets by
tags A list of dataset user tags

Compare Two Datasets

clearml-data compare [--source SOURCE] [--target TARGET] 

Compare two datasets (target vs. source). The command returns a comparison summary that looks like this:

Comparison summary: 4 files removed, 3 files modified, 0 files added

Parameters

Name Description Optional
source Source dataset id (used as baseline) No
target Target dataset id (compare against the source baseline dataset) No
verbose Verbose report all file changes (instead of summary) Yes

Merge Datasets

clearml-data squash --name NAME --ids [IDS [IDS ...]] 

Squash (merge) multiple datasets into a single dataset version.

Parameters

Name Description Optional
name Create squashed dataset name No
ids Source dataset IDs to squash (merge down) No
storage Remote storage to use for the dataset files. Default: files_server Yes
verbose Verbose report all file changes (instead of summary) Yes

Verify Dataset

clearml-data verify [--id ID] [--folder FOLDER] 

Verify that the dataset content matches the data from the local source.

Parameters

Name Description Optional
id Specify dataset ID. Default: previously created/accessed dataset Yes
folder Specify dataset local copy (if not provided the local cache folder will be verified) Yes
filesize If True, only verify file size and skip hash checks (default: false) Yes
verbose Verbose report all file changes (instead of summary) Yes

Get a Dataset

clearml-data get [--id ID] [--copy COPY] [--link LINK] [--overwrite]

Get a local copy of a dataset. By default, you get a read only cached folder, but you can get a mutable copy by using the --copy flag.

Parameters

Name Description Optional
id Specify dataset ID. Default: previously created / accessed dataset Yes
copy Get a writable copy of the dataset to a specific output folder Yes
link Create a soft link (not supported on Windows) to a read-only cached folder containing the dataset Yes
overwrite If True, overwrite the target folder Yes
verbose Verbose report all file changes (instead of summary) Yes

Publish a Dataset

clearml-data publish --id ID

Publish the dataset for public use. The dataset must be finalized before it is published.

Parameters

Name Description Optional
id The dataset task id to be published. No

Python API

It's also possible to manage a dataset using ClearML Data's python interface.

All API commands should be imported with:

from clearml import Dataset

See all API commands in the Dataset reference page.

Tutorials

Take a look at the ClearML Data tutorials: