Gerenciar e fazer queries em dados visuais de forma eficiente com o MongoDB Atlas Vector Search e o FiftyOne

1
pip install fiftyone

1
import fiftyone as fo
2
import fiftyone.zoo as foz
3
4
## load dataset from zoo
5
dataset = foz.load_zoo_dataset("quickstart")
6
7
## launch the app
8
session = fo.launch_app(dataset)

1
from fiftyone import ViewField as F
2
3
view = dataset.filter_labels(
4
    "predictions",
5
    (F("label") == "airplane") & (F("confidence") > 0.8)
6
)

1
# view the dataset and summary
2
print(view)

1
Dataset:     quickstart
2
Media type:  image
3
Num samples: 14
4
Sample fields:
5
    id:           fiftyone.core.fields.ObjectIdField
6
    filepath:     fiftyone.core.fields.StringField
7
    tags:         fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
8
    metadata:     fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.ImageMetadata)
9
    ground_truth: fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)
10
    uniqueness:   fiftyone.core.fields.FloatField
11
    predictions:  fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)
12
View stages:
13
    1. FilterLabels(field='predictions', filter={'$and': [{...}, {...}]}, only_matches=True, trajectories=False)

1
## Inspect the MongoDB agg pipeline
2
print(view._pipeline())

1
[{'$addFields': {'predictions.detections': {'$filter': {'input': '$predictions.detections',
2
     'cond': {'$and': [{'$eq': ['$$this.label', 'airplane']},
3
       {'$gt': ['$$this.confidence', 0.8]}]}}}}},
4
 {'$match': {'$expr': {'$gt': [{'$size': {'$ifNull': ['$predictions.detections',
5
        []]}},
6
     0]}}}]

1
# Sort by the number of objects in the `ground_truth` field
2
3
stage = fo.Mongo([
4
    {
5
        "$addFields": {
6
            "_sort_field": {
7
                "$size": {"$ifNull": ["$ground_truth.detections", []]}
8
            }
9
        }
10
    },
11
    {"$sort": {"_sort_field": -1}},
12
    {"$project": {"_sort_field": False}},
13
])
14
view = dataset.add_stage(stage)

1
export FIFTYONE_DATABASE_NAME=fiftyone
2
export FIFTYONE_DATABASE_URI='mongodb+srv://$USERNAME:$PASSWORD@fiftyone.XXXXXX.mongodb.net/?retryWrites=true&w=majority'

1
export FIFTYONE_BRAIN_DEFAULT_SIMILARITY_BACKEND=mongodb

1
import fiftyone.brain as fob
2
fob.compute_similarity(
3
    dataset,
4
    model="clip-vit-base32-torch", ### Use a CLIP model
5
    brain_key="your_key",
6
    embeddings='clip_embeddings',
7
)

1
## get ID of third sample
2
query = dataset.skip(2).first().id
3
4
## get 25 most similar images
5
view = dataset.sort_by_similarity(query, k=25, brain_key="your_key")
6
session = fo.launch_app(view)

1
query = dataset.first().id
2
3
# shuffle, 
4
# then vector search against 1st sample, 
5
# finally skip top 5 restuls
6
view = dataset.sort_by_similarity(query, k = 20).skip(5)

1
query = "animals"
2
3
view = dataset.sort_by_similarity(query, k = 25)
4
session = fo.launch_app(view)