Training Improvements, LR Scheduler, GPU Fixes, and PyPI Package

[JinruG]

Training Improvements

• cuda() called before train() : The model previously was moved to GPU only after training, causing the entire training loop to run on CPU.
• Best-val-loss checkpointing : Each voice model now saves only the checkpoint that achieved the lowest validation loss, rather than always overwriting with the final epoch.
• ReduceLROnPlateau scheduler : Learning rate is automatically reduced when validation loss plateaus, controlled by the new --lr_patience and --lr_factor options above.

Dependency Compatibility

• Replaced torch.load(...) bare calls with explicit weights_only=True / weights_only=False keyword to silence the FutureWarning introduced in PyTorch 2.0 and enforced as an error in PyTorch 2.4+.
• standalone_mode in the __main__ block is now set to False so that Click returns the result value rather than calling sys.exit(0) on success — required for embedding the entry point in notebooks or test harnesses without process termination.

deepbach_pytorch PyPI Package

https://pypi.org/manage/project/deepbach-pytorch/release/0.3.6/

A test script (test.py) is included that exercises the public API entry point:

test.py