Arabic Sign Language

app.py

2 min read

app.py

source dashboard streamlit analytics

File Path: src/modelling/dashboard/app.py

Purpose: Entry point for the Streamlit-based KArSL Analytics Dashboard.

Overview

The dashboard provides a visual interface to explore model performance, inspect specific data samples, and experiment with data augmentations. It uses a sidebar for global configuration and a tabbed interface for focused analysis.

Dashboard Structure

  • Checkpoint Selector: Lists all .pth files found in TRAIN_CHECKPOINTS_DIR.
  • Split Selector: Choice between train, val, and test splits.
  • Run Evaluation: Triggers the inference pipeline for the selected model and split.

Analysis Tabs

  1. Global Metrics: (Conditional) Displays confusion matrices and per-class performance plots.
  2. Error Analysis: (Conditional) Detailed view of prediction probabilities and specific misclassification instances.
  3. Sample Inspector: Explores the raw keypoints and labels of the dataset.
  4. Augmentation Lab: Interactive playground for testing DataAugmentor transformations.

Workflow

  1. State Management: Uses st.session_state to persist evaluation results and random visualization keys.
  2. Caching: Leverages st.cache_data and st.cache_resource (via loader.py) for efficient model/data loading across reruns.
  3. Dynamic Tabs: The “Global Metrics” and “Error Analysis” tabs only appear after an evaluation has been successfully run.

Key Functions

main()

Orchestrates the sidebar logic, state initialization, and tab rendering.

  • loader.py - Backend logic for data and model caching.
  • views.py - UI components for each dashboard tab.
  • train.py - Source of the training checkpoints.

Graph View

Backlinks