Adding SOTA Ensembling for vision tasks

[kartik4949]

🚀 Feature

So Ensemble a De-Facto for winning competitions and even in real world scenario it is used given enough time budget on inference

multi model ensembling

classification

better_prediction = flash.ensemble(models=[model1, model2], type='classification', method='some_sota_method')

object detection

# Weighted Boxes Fusion
better_prediction = flash.ensemble(models=[model1, model2], type='object_detection', method='WBF')

TTA support is possible too.

e.g

# classification.
model = ImageClassifier.load_from_checkpoint("https://flash-weights.s3.amazonaws.com/image_classification_model.pt")
config = {'transformations':['original', 'flip_lr']}
better_predictions = model.ensemble(type='TTA', task='classification', config=config})

# object detection
model
config = {'transformations : ['original', [1024, 1024], [1080, 1920]], 'fusion':'WBF'} #['original', 'resize', 'resize']
better_predictions = model.ensemble(type='TTA', task='detection', config=config) # fuse using WBF i.e weighted bounding box fusion (SOTA)

Motivation

Pitch

Alternatives

Additional context

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[edgarriba]

duplicated feature request #186

[kartik4949]

duplicated feature request #186

This issue was created in feb
While the referenced one was created in march hence this one is older one!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.