website·Indexed 33 days ago
https://docs.3lc.aiDocumentation for 3LC, a machine learning ecosystem for managing datasets and training metrics. It includes a Python package for creating and modifying tables (such as semantic and instance segmentation), integrating with frameworks like PyTorch, PyTorch Lightning, and Detectron2, and a Hub Frontend based on Flask and Jinja2.
The 3LC Hub is a browser-based, no-code interface that unifies dataset management, model training, performance analysis, and retraining into a single data-centric workflow. It integrates three primary components:
tlc package is the Python entry point for 3LC, a tool used to analyze and improve machine learning models and datasets. It provides functionality to construct and read tlc.Table and tlc.Run objects, collect per-sample metrics, and serve data to the 3LC Dashboard via the Object Service. The public interface is defined by the names in tlc.__all__ and curated sub-namespaces, which follow semantic versioning.The tlc.integration.torch.samplers module provides several utility functions to create PyTorch-compatible samplers from tables based on weight columns:
create_weighted_sampler(): For probability-proportional sampling (with replacement).create_random_sampler(): For uniform random sampling.create_sequential_sampler(): For deterministic sequential iteration.create_repeat_by_weight_sampler(): For repeating rows based on weights to ensure distribution match while sampling every row (with weight > 1) in a single epoch.tlc.data_types package provides several specialized containers for different types of computer vision data:tlc.integration.super_gradients.datasets module provides SuperGradients-compatible dataset implementations that are populated directly from 3LC Tables. It includes specialized datasets for different computer vision tasks, specifically object detection and pose estimation.tlc.metrics.collectors package provides PyTorch implementations for collecting various types of model metrics. It includes specialized collectors for embeddings, segmentation, and custom functional metrics, all inheriting from a common base class.Custom sample types allow developers to round-trip arbitrary Python objects through 3LC Tables. A sample type converts data between two forms:
There are two primary storage tracks:
tlc.sample_types.SampleType and implement to_row() and from_row().tlc.sample_types.ExternalSampleType and implement save() and load().Note: While custom types allow data round-tripping, the Dashboard only displays raw stored data (URLs, numbers, lists) for custom schemas. Specialized Dashboard features (like image rendering or annotation editing) are only available for built-in schemas.
The tlc.integration.detectron2 package provides specialized tools for integrating 3LC with Detectron2, including:
detectron_metrics_collection_hook for standard Detectron2 metrics, metrics_collection_hook for 3LC-specific metrics on Detectron datasets, and bounding_box_metrics_collector for bounding box predictions.register_coco_instances serves as a drop-in replacement for detectron2.data.datasets.register_coco_instances.umap_reduce_embeddings_hook allows for UMAP reduction of embeddings generated during a run.The tlc.metrics.semantic_segmentation module provides offline Python-side metrics for semantic segmentation based on a $C \times C$ pixel confusion matrix (where $C$ is the number of real classes). It supports standard metrics including pixel accuracy, per-class recall/accuracy, per-class IoU, mean IoU, frequency-weighted IoU, and Dice/F1.
Aggregation Duality:
To achieve benchmark-correct cumulative mean IoU using per-image data, sum the per-image confusion matrices and then derive the metrics from the resulting total matrix.
tlc.metrics package is used to collect per-sample metrics from a model operating on a tlc.Table and write those results to a tlc.Run. The primary workflow involves using tlc.collect_metrics() during inference passes, utilizing various metrics_collectors to gather data.tlc Python package provides the core functionality for Three Lines of Code (3LC). The public API is designed for end-users; internal symbols may be visible via introspection but are not part of the public API and are subject to change in future releases. For high-level concepts, refer to the Getting Started and User Guide documentation.tlc.Table. Tables are interoperable with common ML/Python tabular types such as PyTorch Datasets, Pandas DataFrames, and Hugging Face Datasets. Edits made to Tables are non-destructive and tracked via a revision model, ensuring original data remains unmodified.