FiftyOne Brain

0.21.5 · active · verified Thu Apr 16

The FiftyOne Brain extends the FiftyOne ecosystem with powerful machine learning capabilities for data curation and model analysis. It provides features like visual similarity search, text-based querying, finding unique and representative samples, detecting media quality issues, and identifying annotation mistakes. The library is actively developed, with frequent releases.

Common errors

```
ValueError: Brain method run with key 'my_brain_key' already exists
```
cause You are attempting to compute a FiftyOne Brain method (e.g., `compute_visualization`, `compute_similarity`) using a `brain_key` that is already associated with an existing run on the dataset.

fix Delete the old run before recomputing: `dataset.delete_brain_run('my_brain_key')`, or provide a new, unique `brain_key` for the current run.
```
ModuleNotFoundError: No module named 'umap'
```
cause You are trying to use the UMAP dimensionality reduction method in `fob.compute_visualization` but the `umap-learn` package is not installed.

fix Install the required package: `pip install umap-learn`.
```
AttributeError: module 'fiftyone.brain' has no attribute 'compute_similarity' (or other Brain functions)
```
cause The `fiftyone-brain` package might not be correctly installed, or an outdated version of `fiftyone` is in use where the brain components are not properly linked or exposed.

fix Ensure `fiftyone-brain` is installed: `pip install fiftyone-brain`. If you have `fiftyone` installed, verify `pip show fiftyone-brain` shows it. Sometimes reinstalling `fiftyone` (`pip install --upgrade fiftyone`) can resolve component linking issues.

Warnings

gotcha When re-running a FiftyOne Brain method like `compute_visualization` or `compute_similarity`, if you use the same `brain_key` as a previous run, it will raise a `ValueError`.
Fix: Before re-running, either delete the existing run with `dataset.delete_brain_run("your_brain_key")` or provide a new, unique `brain_key`.
gotcha The `compute_visualization` method, when using `method="umap"`, requires the `umap-learn` package to be installed separately.
Fix: Install the `umap-learn` library: `pip install umap-learn`.
breaking `compute_visualization` with `TSNE` had compatibility issues with `scikit-learn` versions older than 1.7.0.
Fix: Upgrade `fiftyone-brain` to v0.21.4 or newer, or ensure `scikit-learn` is at least 1.7.0 if using an older `fiftyone-brain` version.
gotcha If a `brain_key` is not explicitly provided when computing a Brain method, the run will not be persisted on your dataset and will not be accessible via the FiftyOne App.
Fix: Always provide a descriptive `brain_key` when calling `compute_visualization`, `compute_similarity`, etc., to ensure the results are saved and accessible.
gotcha To use certain embedding models with `compute_similarity` (e.g., 'clip-vit-base32-torch'), you might need to install additional FiftyOne plugins or packages like `fiftyone-embeddings`.
Fix: Check the FiftyOne documentation for specific model requirements. Often, `pip install fiftyone-embeddings` or related model-specific packages are needed.

Install

pip install fiftyone-brain Install fiftyone-brain
pip install fiftyone Install FiftyOne (includes Brain)

Imports

fiftyone.brain
wrong:
```
import fiftyone_brain as fb
```
correct:
```
import fiftyone.brain as fob
```
The common convention is to import the brain subpackage from `fiftyone` and alias it as `fob`.
fiftyone
```
import fiftyone as fo
```
Often used in conjunction with FiftyOne Brain features.
fiftyone.zoo
```
import fiftyone.zoo as foz
```
Commonly used to load sample datasets for analysis with FiftyOne Brain.

Quickstart

This quickstart demonstrates how to load a dataset, compute a low-dimensional visualization of its embeddings using UMAP, and then compute a visual similarity index, which are core functionalities of FiftyOne Brain. The visualization can then be explored interactively in the FiftyOne App.

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

# Load a sample dataset
dataset = foz.load_zoo_dataset("quickstart", max_samples=100)

# Compute a 2D visualization of the dataset embeddings
print("Computing visualization...")
results = fob.compute_visualization(
    dataset,
    brain_key="quickstart_viz",
    method="umap" # Requires `pip install umap-learn`
)

print(f"Visualization results stored under brain key: {results.key}")

# You can then launch the FiftyOne App to view the visualization
# session = fo.launch_app(dataset)
# session.wait()

# Compute visual similarity index
print("Computing similarity index...")
fob.compute_similarity(
    dataset,
    brain_key="quickstart_similarity",
    model="clip-vit-base32-torch" # Requires `pip install fiftyone-embeddings`
)

print(f"Similarity index computed for brain key: quickstart_similarity")

view raw JSON →