Enriching your data and using it¶

In order to improve a model you need to train it on the most relevant data points - the ones that represent use cases in which the model is underperforming. Focusing on these relevant use cases is a hard task, especially when dealing with large datasets. This is why Data Engine provides the ability to intuitively query and create new subsets of your data, so you will be able to zoom in on these use cases and improve your model effectively.

To query and generate training-ready data sets from your connected data source, you will first need to enrich it by adding custom metadata that’s relevant to your use case. Data Engine enrichments can include metadata, annotations, and model predictions. There are 3 ways to upload enrichments using the DagsHub Client:

The Pandas Dataframe interface
The metadata context interface
Dictionary assignments on datapoints

1. Enriching with Pandas Dataframe interface¶

To enrich datasource with an existing data frame, use the following command:

ds.upload_metadata_from_dataframe(dataFrame, path_column="path")

path_column supports either an index or the name of the column, which specifies the path of the data file you want to attach metadata to. If not specified, the first dataframe column will be assumed by default. All other columns in the data frame are considered as metadata columns.

For example:

# Create a pandas dataframe
columns = ["path", "squirrel_detected"]
data = [["data/dataPoint1.png", True],
        ["data/dataPoint2.png", False]]

df = pd.DataFrame(data, columns=columns)

# Enriching data with the data frame
ds.upload_metadata_from_dataframe(df, path_column="path")

2. Enriching with metadata context interface¶

To Enrich your data source with the metadata context interface, you can use the metadata_context with the following commands:

with ds.metadata_context() as ctx:
    metadata = {
        "key_1": value_1,
        "key_2": value_2,
    }

    # To add enrichments to a single data point within the data source:
    ctx.update_metadata("path/to/datapoint.png", metadata)

    # To add enrichments to several data points within the data source:
    ctx.update_metadata(["path/to/datapoint1.png", "path/to/datapoint2.png"], metadata)

Where path/to/datapoint*.png is the path relative to the root of the datasource.

For example, for our baby Yoda segmentation repo:

with ds.metadata_context() as metadata_context:
    metadata = {
        "episode": 5,
        "has_baby_yoda": True,
    }

    # Enriching one data point
    metadata_context.update_metadata("images/005.jpg", metadata)

    # Enriching several data points 
    metadata_context.update_metadata(["images/006.jpg", "images/007.jpg"], metadata)

In our baby Yoda segmentation repository - where the data is located at repo://simon/baby-yoda-segmentor/data - the data point path will be images/005.jpg

3. Enriching with with dictionary-like assignment¶

datapoints = ds.all()
datapoints["images/005.jpg"]["episode"] = 5
datapoints["images/005.jpg"].save()

save() should be called for each datapoint eventually for changes to be commited.

when looping over many datapoints the preferred method (to avoid many network writes) is to work within a metadata context block, but dictionary syntax can still be used:

with ds.metadata_context() as ctx:
    for dp in datapoints:
        dp["episode"] = 4

note that in the above example save() is omitted as a commit is done once the context is exited.

Supported metadata types:

Int
Float
Boolean
String

Blobs (bytes). For example:

with ds.metadata_context() as ctx:
    with open(”file1_depth.png”, “rb”) as f:
        ctx.update_metadata(”file1”, {”depth_image”: f.read()})

Deleting datapoints and metadata¶

To delete a datapoint from a datasource, simply run:

datapoints["images/005.jpg"].delete()

Or, delete a list of datapoints from a datasource:

datapoints = ds.all()
ds.delete_datapoints(datapoints)

Note

Deleted datapoints will no longer show up in queries.
Does not delete the datapoint's file, only removing the data from the datasource.
You can still query deleted datapoints and associated metadata with versioned queries whose time is before deletion time.
You can re-add deleted datapoints to the datasource.
Datasource scanning will not add this datapoint back.

Delete one or more metadata fields from a datapoint:

datapoints["images/005.jpg"].delete_metadata("episode")
# or
datapoints["images/005.jpg"].delete_metadata(["episode", "annotation"])

Or, delete fields at the source level, from multiple datapoints:

datapoints = ds.all()
ds.delete_metadata_from_datapoints(datapoints, ["episode", "annotation"])

Note

Deleted values can be accessed using versioned query with time set before the deletion.

Viewing and using datasource content¶

The easiest way to display your datapoints is ds.head().dataframe or ds.all().dataframe, though you can leave out the .dataframe to get the actual QueryResult object.

You can fetch the data points (set of filenames + their enrichments) that exist in a data source. To display data points within a datasource, use the following commands:

# Fetch the first 100 data points 
ds.head()

# Fetch the first 42 data points 
ds.head(42)

# Fetch all data points
ds.all()

The returned object will include the fetched datapoints and their metadata.

If you want to do sophisticated computations, aggregations, etc. on the fetched metadata, you can easily do that by transforming the fetched datapoints to a Pandas DataFrame. Each row will represent a datapoint, with one column specifying its file path and the other columns showing its metadata values.

df = ds.head().dataframe

Downloading datapoint files¶

To download the raw data files referenced in the path field of the fetched query results, use the download_files function:

ds.all().download_files(target_directory="path")

If target_directory is not specified, data points will be downloaded to the ~/dagshub/datasets/<user>/<repo>/<datasource_id> directory.

If you have another column that includes valid paths inside you’re repository, for example if you have images that you use as labels, and are also stored in your project, you can download them using:

ds.all().download_files(target_directory="path", path_field="label_path")

Where path_field is the name of the column that has the paths you’d like to download.

Download files directly from your storage bucket¶

If your data source is a cloud object storage bucket, which you connected to DagsHub, then you can download raw data files directly from your bucket. This can lead to massive savings of time and data transfer costs, so we recommend doing that. Use the following example to enable bucket downloaders:

from dagshub.common.download import enable_s3_bucket_downloader, enable_gcs_bucket_downloader

# S3
enable_s3_bucket_downloader()
# GCS
enable_gcs_bucket_downloader()

# You can also use a custom client, if the default auth is not set up
import boto3

client = boto3.client("s3", endpoint_url=...)
enable_s3_bucket_downloader(client)

For this to work, you need to be authenticated with your cloud provider and have proper permissions. Usually, that means downloading the appropriate cloud vendor SDK (aws or gcloud). If you don’t use bucket downloaders, everything will still work - you will download the raw files from dagshub.com using your DagsHub credentials.

Downloading Blob fields¶

Blob fields are not downloaded by default, to download the data, use the following function:

df = ds.all().get_blob_fields(”binary_1”, “binary_2”, load_into_memory = True)

Working with Annotation Columns¶

Data Engine treats certain data types, like annotations uniquely, enabling you to send them to be annotated, or visualize them.

If you created the annotation with DagsHub Annotations, then this will be automatically taken care of.

However, if you import annotations from a different place, and convert them into the correct label studio format, you'll need to mark them as metadata type annotation. To do this simply run:

from dagshub.data_engine import datasources
ds = datasources.get('<user_name>/<repo_name>', '<datasource_name>')

ds.metadata_field("<your annotation column name>").set_annotation().apply()

Then you'll be able to see and use the annotations as necessary.