Implement PyTorch DataLoader for exoskeleton dataset

[DylanG5]

Summary

• Implements PyTorch Dataset and DataLoader for exoskeleton Phase1 data from HuggingFace
• Adds comprehensive test suite with 44 unit tests
• Provides configuration and helper utilities

Changes

Core Components

• ExoskeletonDataset (datasets.py): PyTorch Dataset with auto-download

  • Two-layer caching (HF cache + preprocessed tensor cache)
  • Participant-based filtering for LOSO cross-validation
  • Optional feature normalization with computed statistics
  • Variable-length sequence handling with padding

• download.py: Dataset download and cache management

  • Automatic download from HuggingFace Hub
  • Data integrity verification (fields, dimensions, sequence lengths)
  • Cache info and clear cache functions

• create_dataloaders(): Helper for train/val/test DataLoaders

  • Consistent normalization across splits
  • Configurable batch size, workers, shuffling

• collate_fn: Custom collation for variable-length sequences

  • Pads to max sequence length in batch
  • Returns mask for valid timesteps
  • Preserves metadata (participant, trial name, mass)

Configuration

• configs/data/phase1.yaml: YAML config with participant splits, features, and settings

Testing

• 44 unit tests covering:
  • Dataset initialization and caching
  • Download and integrity verification
  • DataLoader iteration and batching
  • Normalization and padding
  • Edge cases (empty filters, single samples)

Dependencies

• Added datasets>=2.14.0 and huggingface-hub>=0.17.0

Test plan

• All 44 unit tests pass: pytest tests/test_data/ -v
• Linting passes: ruff check src/ tests/
• Type checking passes: mypy src/
• Formatting passes: black --check src/ tests/

Usage Example

from exoskeleton_ml.data import ExoskeletonDataset, create_dataloaders

# Create train/val/test DataLoaders
train_loader, val_loader, test_loader = create_dataloaders(
    hf_repo="macexo/exoskeleton-phase1",
    train_participants=["BT01", "BT02", "BT03"],
    val_participants=["BT14"],
    test_participants=["BT15"],
    batch_size=32,
    normalize=True,
)

for batch in train_loader:
    inputs = batch["inputs"]    # (batch, seq_len, 28)
    targets = batch["targets"]  # (batch, seq_len, 4)
    mask = batch["mask"]        # (batch, seq_len)

🤖 Generated with Claude Code