Machine Learning: Groups

You will learn how to organize your ML assets in groups with this tutorial.

In this tutorial, you will learn how to organize your Machine Learning (ML) assets under a single, logical group. These assets can be datasets, models (from different ML frameworks solving the same problem), or any other TileDB asset that you want to bundle under the same semantics. Storing ML datasets alongside ML models in a registry for experimentation offers several advantages:

• Enhanced reproducibility
• Version consistency
• Efficient experimentation
• Iterative model development
• Data versioning and lineage
• Collaboration and knowledge sharing
• Complete experiment documentation
• Reduced dependencies on external resources

Note

The following tutorial assumes that you have successfully completed the tutorials for ML dataset ingestion and ML model ingestion in TileDB.

Create an ML group

You can create a group that will store your ML assets:

group = tempfile.mkdtemp("mnist_experiment")
tiledb.group_create(group)
tdb_grp = tiledb.Group(group, "w")
tdb_grp.close()

You can create also sub-groups according to your use case. Users have the flexibility to bundle multiple models and datasets, from the same or different frameworks, into a group. This group can encompass elements that address or correspond to the same application, solution, algorithm, or problem statement. For this tutorial, you will be creating two subgroups. This will allow you to store your datasets and models under the same group, which can serve as a directory for storing all your experiment’s related assets, following a logical hierarchy.

with tiledb.Group(group, "w") as grp:
    subgrp_datasets_path = "datasets"
    subgrp_models_path = "models"

    # you can also use `tiledb.group_create()` to create groups
    tiledb.group_create(subgrp_datasets_path)
    tiledb.group_create(subgrp_models_path)

    grp.add(subgrp_datasets_path)
    grp.add(subgrp_models_path)

Add a dataset array to a sub-group

tdb_data_grp = tiledb.Group(subgrp_datasets_path, "w")
tdb_data_grp.add(dataset, name="mnist_dataset")
tdb_data_grp.close()

Add a model array to a sub-group

tdb_model_grp = tiledb.Group(subgrp_models_path, "w")
tdb_model_grp.add(tf_model_array, name="tf_mnist_model")
tdb_model_grp.add(pytorch_model_array, name="pytorch_mnist_model")
tdb_model_grp.close()

Iterate group members

tdb_grp.open("r")

# show a directory structure of the groups
print(tdb_grp)

# iterate through the group members
for i in range(0, len(tdb_grp)):
    print(f"URI: {tdb_grp[i].uri}, Type: {tdb_grp[i].type}")

# remember to close the array
tdb_grp.close()

tmps6lrj65rmnist_experiment GROUP
|-- datasets GROUP
|------ mnist_dataset ARRAY
|-- models GROUP
|------ tf_mnist_model ARRAY
|------ pytorch_mnist_model ARRAY

URI: file:///Users/konstantinostsitsimpikos/tiledb_dev/arrayloader-benchmarks/models, Type: <class 'tiledb.group.Group'>
URI: file:///Users/konstantinostsitsimpikos/tiledb_dev/arrayloader-benchmarks/datasets, Type: <class 'tiledb.group.Group'>

Delete groups

import shutil

with tiledb.Group(subgrp_datasets_path, "m") as subgrp:
    # recursively delete the subgroup and array
    subgrp.delete(recursive=True)
    # make sure to delete the subgroup folder and array folder
    shutil.rmtree(subgrp_datasets_path)
    shutil.rmtree(subgrp_models_path)
    shutil.rmtree(dataset)
    shutil.rmtree(tf_model_array)
    shutil.rmtree(pytorch_model_array)


with tiledb.Group(group, "m") as grp:
    # delete group folder contents
    grp.delete()
    # make sure to delete the group folder
    shutil.rmtree(group)

For more details on TileDB groups, visit Arrays Tutorials: Groups.