API Usage Guide

The ManyLatents programmatic API enables workflow integration and in-memory data chaining.

Quick Start

from manylatents.api import run

# Single algorithm run
result = run(
    data='swissroll',
    algorithms={'latent': {'_target_': 'manylatents.algorithms.latent.pca.PCAModule', 'n_components': 10}}
)

embeddings = result['embeddings']  # numpy array
scores = result['scores']          # dict of metrics

Chained Workflows

Chain multiple algorithms by passing the output of one as input to another:

from manylatents.api import run

# Step 1: Initial dimensionality reduction
result1 = run(
    data='swissroll',
    algorithms={'latent': {'_target_': 'manylatents.algorithms.latent.pca.PCAModule', 'n_components': 50}}
)

# Step 2: Chain to another algorithm
result2 = run(
    input_data=result1['embeddings'],
    algorithms={'latent': {'_target_': 'manylatents.algorithms.latent.umap.UMAPModule', 'n_components': 2}}
)

final_embeddings = result2['embeddings']

Note: Embeddings are automatically converted to numpy arrays by the evaluation system.

Available Algorithms

Dimensionality Reduction

• PCA: manylatents.algorithms.latent.pca.PCAModule
• t-SNE: manylatents.algorithms.latent.tsne.TSNEModule
• UMAP: manylatents.algorithms.latent.umap.UMAPModule
• PHATE: manylatents.algorithms.latent.phate.PHATEModule

API Reference

run(input_data=None, **overrides)

Execute a dimensionality reduction algorithm.

Parameters:

• input_data (np.ndarray, optional): In-memory data array of shape (n_samples, n_features). If provided, this data is used instead of loading from a dataset.
• **overrides: Configuration overrides (e.g., data='swissroll', algorithms={...})

Returns:

Dictionary with keys: - embeddings: Computed embeddings (numpy array) - label: Labels from dataset (if available) - metadata: Run metadata dictionary - scores: Evaluation metrics (if enabled)

Examples:

# Using a built-in dataset
result = run(data='swissroll', algorithms={'latent': 'pca'})

# Using in-memory data
import numpy as np
my_data = np.random.randn(1000, 100).astype(np.float32)
result = run(input_data=my_data, algorithms={'latent': 'pca'})

# Chaining algorithms
result2 = run(input_data=result['embeddings'], algorithms={'latent': 'umap'})

Advanced Usage

Custom Configuration

result = run(
    data='swissroll',
    algorithms={
        'latent': {
            '_target_': 'manylatents.algorithms.latent.umap.UMAPModule',
            'n_components': 2,
            'n_neighbors': 15,
            'min_dist': 0.1
        }
    }
)

Disabling Features for Speed

# Disable W&B logging
result = run(data='swissroll', algorithms={'latent': 'pca'}, debug=True)

# Skip evaluation metrics
result = run(data='swissroll', algorithms={'latent': 'pca'}, metrics=None)

Data Format Requirements

• Input: numpy.ndarray with dtype float32 or float64
• Shape: (n_samples, n_features)
• Output: numpy array (tensor conversion is handled automatically)

Multi-Step Example

from manylatents.api import run

# Progressive dimensionality reduction
steps = [
    ('PCA 100D', 'manylatents.algorithms.latent.pca.PCAModule', 100),
    ('PCA 50D', 'manylatents.algorithms.latent.pca.PCAModule', 50),
    ('UMAP 2D', 'manylatents.algorithms.latent.umap.UMAPModule', 2),
]

# Initial data
current_data = run(
    data='swissroll',
    algorithms={'latent': {'_target_': steps[0][1], 'n_components': steps[0][2]}}
)

# Chain subsequent steps
for name, target, n_comp in steps[1:]:
    print(f"Running {name}...")
    current_data = run(
        input_data=current_data['embeddings'],
        algorithms={'latent': {'_target_': target, 'n_components': n_comp}}
    )

print(f"Final shape: {current_data['embeddings'].shape}")

Implementation Details

The in-memory data pipeline uses: - PrecomputedDataModule: Accepts data parameter for numpy arrays - InMemoryDataset: Wraps arrays in LatentOutputs format - Compatible with all ManyLatents metrics, callbacks, and visualizations

Troubleshooting

Common Issues

"PrecomputedDataModule requires either a 'path' or 'data' argument"

Provide either data='dataset_name' OR input_data=array, not both or neither.

See Also

• Testing Guide
• Metrics