Fixed Overfitting with Dropout Rate, Weight Decay, and on the fly data augmentation.

[sridhs21]

Add Regularization and Data Augmentation

This PR adds regularization techniques and on-the-fly data augmentation to improve model training.

Changes

New Command-Line Arguments

• --dropoutRate (default: 0.3) - Control dropout strength
• --weightDecay (default: 5e-4) - Control L2 regularization
• --seed (default: None) - Set random seed for reproducibility
• --require-gpu - Exit if GPU not available

Dropout Improvements

• UNet now accepts dropout_rate parameter
• Added dropout layers after each decoder block (dec1_dropout, dec2_dropout, dec3_dropout, dec4_dropout)
• Added dropout after bottleneck (bottleneck_dropout)

Weight Decay

• Changed from hardcoded weight_decay=1e-5 to configurable weight_decay=args.weightDecay
• Logged in optimizer initialization output

On-the-Fly Data Augmentation

• Added augment parameter to XPointPatchDataset
• Added seed parameter to XPointPatchDataset for reproducible augmentation
• New _apply_augmentation() method with:
  • Random rotation (90°, 180°, 270°) - 75% probability
  • Random horizontal flip - 50% probability
  • Random vertical flip - 50% probability
  • Gaussian noise injection - 30% probability
  • Brightness/contrast adjustment - 30% probability
  • Random cutout/erasing - 20% probability
• Training uses augment=True, validation uses augment=False
• Removed static augmentation: XPointDataset(..., rotateAndReflect=False)

Reproducibility

• New set_seed() function sets seeds for Python, NumPy, and PyTorch
• Configures PyTorch deterministic mode when seed is provided
• Seeds propagated to patch datasets

GPU Checking

• --require-gpu flag exits with error if CUDA unavailable
• Provides diagnostic messages for troubleshooting

[cwsmith]

There is one small comment below. In addition to that, we discussed offline about changing the default checkpoint rate to every 100 epochs. Each checkpoint was about 100MB which could quickly fill up disk space for someone who doesn't expect it.

[cwsmith]

@sridhs21 would you also please change the default checkpoint interval to 100 epochs?

[cwsmith]

Thank you.