Adding support for applying Lightning Flash models to datasets

[ehofesmann]

Adds support for applying Lightning Flash models to datasets via SampleCollection.apply_model() and SampleCollection.compute_embeddings().

This integration requires flash==0.4.0 or later.

This branch doubles as the release branch for fiftyone==0.10.0.

Predictions example:

from flash.image import ObjectDetector
import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart").exclude_fields("predictions").clone()

model = ObjectDetector.load_from_checkpoint("https://flash-weights.s3.amazonaws.com/object_detection_model.pt")

view = dataset.take(5)
view.apply_model(model, label_field="predictions")

session = fo.launch_app(view=view)

Embeddings example:

from flash.image import ImageEmbedder
import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset("quickstart")

model = ImageEmbedder(backbone="resnet101", embedding_dim=128)

view = dataset.take(5)
embeddings = view.compute_embeddings(model)

print(embeddings.shape)

See Lightning-Universe/lightning-flash#360 for many more examples!

[benjaminpkane]

Released