loader.py

source dashboard data-loading caching

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

Purpose: Data access layer for the dashboard. Handles efficient loading of models and datasets using Streamlit caching.

Overview

Provides cached wrappers around core functions to ensure the dashboard remains responsive while handling large models and datasets.

Functions

get_checkpoints_metadata(checkpoint_path)

Decorator: @st.cache_data Extracts both class count and architecture from the checkpoint file. Returns: (num_signs, ModelSize)

load_cached_checkpoints(checkpoints_dir)

Decorator: @st.cache_data Recursively scans the directory for .pth files to populate the selection sidebar.

load_cached_model(checkpoint_path, _metadata)

Decorator: @st.cache_resource Loads the PyTorch model into memory on the configured device and sets it to evaluation mode.

get_cached_dataloaders(num_signs)

Decorator: @st.cache_resource Initializes and caches lazy dataloaders for all three data splits (train, val, test).

run_inference(_model, _dataloader, ...)

Decorator: @st.cache_data Executes a full inference pass on the provided dataloader. Displays a progress bar in the UI. Returns: (y_true, y_pred, y_probs) as NumPy arrays.

Related Documentation

• dataloader.py - Underlying data provider logic.
• utils.py - extract_metadata_from_checkpoint implementation.
• model.py - load_model implementation.
• app.py - Consumer of these loader functions.