Arabic Sign Language

constants.py

2 min read

constants.py

source core configuration constants

File Path: src/core/constants.py

Purpose: System-wide constants, configuration values, and path definitions.

Overview

Central configuration module that defines:

  • Device configuration (CPU/GPU)
  • Comprehensive directory structure (local vs Kaggle)
  • Dataset parameters and split types
  • Advanced model architecture logic (ModelSize, HeadSize)

Environment & Device

  • DEVICE: Execution device for PyTorch/ONNX. Automatically selects "cuda" if available and USE_CPU is “0”.
  • LOCAL_DEV: (int) Toggles between Kaggle (0) and Local (1) environments.

Directory Structure

Base Directories

  • PROJECT_ROOT_DIR: Root folder containing the src/ directory.
  • STATIC_ASSETS_DIR: {PROJECT_ROOT_DIR}/static.
  • MODELS_DIR: {PROJECT_ROOT_DIR}/models.
  • LOGS_DIR: {PROJECT_ROOT_DIR}/logs.
  • LANDMARKERS_DIR: Folder containing MediaPipe .task files.
  • PROJECT_DATA_DIR: {PROJECT_ROOT_DIR}/data.

Dataset Root Paths

  • DATA_INPUT_DIR: Input path (e.g., /kaggle/input or ./data).
  • DATA_OUTPUT_DIR: Output path (e.g., /kaggle/working or ./data).
  • KARSL_DATA_DIR: {DATA_INPUT_DIR}/karsl-502.
  • TRAIN_CHECKPOINTS_DIR: {DATA_OUTPUT_DIR}/checkpoints.

Specialized Data Paths (KAGGLE/LOCAL)

  • NPZ_KPS_DIR: Extracted keypoints in NPZ format.
  • MMAP_PREPROCESSED_DIR: Local/Input directory for preprocessed memory-mapped files.
  • MMAP_OUTPUT_PREPROCESSED_DIR: Output directory for generated memory-mapped files.

Data Files & Assets

  • LABELS_PATH: Path to the KARSL-502_Labels.xlsx file.
  • LABELS_JSON_PATH: Path to the KARSL-502_Labels.json file.
  • FACE_SYMMETRY_MAP_PATH: Path to the .npy symmetry map used for face mesh processing.
  • SIMPLIFIED_FACE_CONNECTIONS_PATH: Path to the simplified_face_connections.json file.

Model Architecture Definitions

HeadSize (StrEnum)

Defines supported attention head sizes and their corresponding feature dimensions.

  • tiny (‘t’): 16 dimensions
  • small (‘s’): 32 dimensions
  • medium (‘m’): 64 dimensions
  • large (‘l’): 128 dimensions

ModelSize (Class)

Manages the complexity of the transformer model (heads and layers) based on the chosen HeadSize.

HeadSizeNum HeadsNum Layers
Tiny42
Small44
Medium46
Large68

Configuration & Performance

Feature Configuration

  • SEQ_LEN: 50 (Temporal window size).
  • FEAT_NUM: 184 (Total keypoints: POSE + FACE + 2*HANDS).
  • FEAT_DIM: 4 (x, y, z, visibility).

Timing & Performance

  • MAX_WORKERS: 4 (Thread pool size for extraction).
  • MS_30FPS: ~33.33ms (Float interval for 30 FPS).
  • MS_30FPS_INT: 33ms (Integer interval for 30 FPS).

Enumerations & Types

SplitType

  • train, val, test

DatasetType

  • lazy, mmap

KarslDatasetType

  • Union of LazyKArSLDataset and MmapKArSLDataset.

Used By:

Graph View

Backlinks