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.
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
The commands will do the following:
-
Start a Data Processing Task called "initial_version" in the "dataset_example" project
-
The CLI will return a unique ID for the dataset
-
All the files from the "data_folder" folder will be added to the dataset and uploaded by default to the ClearML server.
:::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.
Setup
clearml-data comes built-in with our clearml python package! Just check out the getting started guide for more info!
Usage
CLI
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 |  |
| project | Dataset's project | |
| 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 |  |
| tags | Dataset user tags. The dataset can be labeled, which can be useful for organizing datasets | |
:::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 to Dataset
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 | |
| files | Files / folders to add. Wildcard selection is supported, for example: ~/data/*.jpg ~/data/json |
|
| dataset-folder | Dataset base folder to add the files to in the dataset. Default: dataset root | |
| non-recursive | Disable recursive scan of files | |
| verbose | Verbose reporting | |
Remove Files From Dataset
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 | |
| 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. |
|
| non-recursive | Disable recursive scan of files | |
| verbose | Verbose reporting | |
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 | |
| storage | Remote storage to use for the dataset files. Default: files_server | |
| disable-upload | Disable automatic upload when closing the dataset | |
| verbose | Verbose reporting | |
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 | |
| storage | Remote storage to use for the dataset files. Default: files_server | |
| verbose | Verbose reporting | |
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 | |
| folder | Local folder to sync. Wildcard selection is supported, for example: ~/data/*.jpg ~/data/json |
|
| storage | Remote storage to use for the dataset files. Default: files_server | |
| 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 | |
| project | If creating a new dataset, specify the dataset's project name | |
| name | If creating a new dataset, specify the dataset's name | |
| tags | Dataset user tags | |
| skip-close | Do not auto close dataset after syncing folders | |
| verbose | Verbose reporting | |
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 | |
| project | Specify dataset project name (if used instead of ID, dataset name is also required) | |
| name | Specify dataset name (if used instead of ID, dataset project is also required) | |
| filter | Filter files based on folder / wildcard. Multiple filters are supported. Example: folder/date_*.json folder/sub-folder |
|
| modified | Only list file changes (add / remove / modify) introduced in this version | |
Delete a 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 | |
| force | Force dataset deletion even if other dataset versions depend on it | |
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 | |
Python API
All API commands should be imported with
from clearml import Dataset
Dataset.get(dataset_id=DS_ID).get_local_copy()
Returns a path to dataset in cache, and downloads it if it is not already in cache.
Parameters
| Name | Description | Optional |
|---|---|---|
| use_soft_links | If True, use soft links. Default: False on Windows, True on Posix systems | |
| raise_on_error | If True, raise exception if dataset merging failed on any file | |
Dataset.get(dataset_id=DS_ID).get_mutable_local_copy()
Downloads the dataset to a specific folder (non-cached). If the folder already has contents, specify whether to overwrite its contents with the dataset contents.
Parameters
| Name | Description | Optional |
|---|---|---|
| target_folder | Local target folder for the writable copy of the dataset | |
| overwrite | If True, recursively delete the contents of the target folder before creating a copy of the dataset. If False (default) and target folder contains files, raise exception or return None | |
| raise_on_error | If True, raise exception if dataset merging failed on any file | |
Dataset.create()
Create a new dataset.
Parent datasets can be specified, and the new dataset inherits all of its parent's content. Multiple dataset parents can be listed. Merging of parent datasets is done based on the list's order, where each parent can override overlapping files in the previous parent dataset.
Parameters
| Name | Description | Optional |
|---|---|---|
| dataset_name | Name of the new dataset | |
| dataset_project | The project containing the dataset. If not specified, infer project name from parent datasets. If there is no parent dataset, then this value is required | |
| parent_datasets | Expand a parent dataset by adding / removing files | |
| use_current_task | If True, the dataset is created on the current Task. Default: False | |
Dataset.add_files()
Add files or folder into the current dataset.
Parameters
| Name | Description | Optional |
|---|---|---|
| path | Add a folder / file to the dataset | |
| wildcard | Add only a specific set of files based on wildcard matching. Wildcard matching can be a single string or a list of wildcards, for example: ~/data/*.jpg, ~/data/json |
|
| local_base_folder | Files will be located based on their relative path from local_base_folder | |
| dataset_path | Where in the dataset the folder / files should be located | |
| recursive | If True, match all wildcard files recursively | |
| verbose | If True, print to console files added / modified | |
Dataset.upload()
Start file uploading, the function returns when all files are uploaded.
Parameters
| Name | Description | Optional |
|---|---|---|
| show_progress | If True, show upload progress bar | |
| verbose | If True, print verbose progress report | |
| output_url | Target storage for the compressed dataset (default: file server). Examples: s3://bucket/data, gs://bucket/data , azure://bucket/data, /mnt/share/data |
|
| compression | Compression algorithm for the Zipped dataset file (default: ZIP_DEFLATED) | |
Dataset.finalize()
Closes the dataset and marks it as Completed. After a dataset has been closed, it can no longer be modified. Before closing a dataset, its files must first be uploaded.
Parameters
| Name | Description | Optional |
|---|---|---|
| verbose | If True, print verbose progress report | |
| raise_on_error | If True, raise exception if dataset finalizing failed | |