Adding batch image processing to Vision API

[daspecster]

The amount of data that Vision can return can be pretty deep and processing more than one image at a time adds another level to that.

Here's my proposed approach to work with multiple images with multiple feature types.

        >>> images = client.batch()
        >>> for i in range(10):
        ...     images.add_image(client.image(source_uri='gs://%s.jpg' % (i,)))
        >>> images.detect_faces()
        {'0.jpg': [], '1.jpg': [],}
        >>>
        >>> # With multiple feature types.
        >>> features = [Feature(FeatureTypes.FACE_DETECTION, 5),
        ...             Feature(FeatureTypes.LOGO_DETECTION, 3)]
        >>> for i in range(10):
        ...     image = client.image(source_uri='gs://%s.jpg' % (i,))
        ...     images.add_image(image=image, features=features)
        >>> images.detect()

With a context manager it could look something like this...

        >>> features = [Feature(FeatureTypes.FACE_DETECTION, 5),
        ...             Feature(FeatureTypes.LOGO_DETECTION, 3)]
        >>> with client.images() as images:
        ...     for i in range(10):
        ...         image = client.image(source_uri='gs://%s.jpg' % (i,))
        ...         images.add_image(image=image, features=features)
        ...     results = images.detect()

[dhermes]

So client.images() is like a batch? Why not name it as such? What is the equivalent backend method that you'd be using? (Or would you just be sending N requests?)

[daspecster]

It will actually be one request with N vision request objects, so I could change it to batch().

The GAPIC method is called batch_annotate_images.

[dhermes]

SGTM. As for the interface of the batch, you should check out the datastore batch / accompanying docs. It commits on __exit__, but we also document an explicit public interface people can use to add mutations to the batch and to manually commit or rollback the batch (rollback is datastore transaction specific).

[tseaver]

I definitely like having the context manager be "sugar" over an explicit imperative surface.

[bw4sz]

Is there a benefit to running batch if we are asking for the same feature. Would I expect performance increase by running batch versus looping through N requests?