Load YAML configuration file.

Load YAML configuration file.

Load model with FP16 optimization.

Mixup augmentation for regression tasks. Combines two samples with random interpolation: x_new = lambda * x_i + (1-lambda) * x_j

Fast prediction with FP16 and no TTA. Args: use_fp16: Enable FP16 inference use_autocast: If True, use autocast (for img_siz

Monte Carlo prediction with PatchDropout for ensemble diversity. Args: mc_samples: Number of MC samples to average patch_dro

Run prediction with one fold model on split images (supports both 3-head and 5-head). For each image: - Predict left half - Predict

Run prediction with one fold model (with optional TTA). For each image: - Split into left/right halves - Predict each half - Sum

Run 3-crop prediction with one fold model (matches train.py validate_oof_3crop). For each image: - Predict left, center, right crops

Convert preds_np (keys: total,gdm,green[,dead,clover]) into 5 target arrays. For 3-head models, dead/clover are derived with non-negativity c

Convert preds_np dict to 5 target arrays.

Fast inference without TTA. Args: debug_mode: If True, use train data for debugging/batch size estimation debug_n: Number of

Main Test Time Training inference function. 1. Load test data and generate pseudo labels with pre-trained models 2. For each fold, re-tr

Set epoch for seed variation (call at start of each epoch).

Set the GRL lambda value for domain adversarial training.

Set the GRL lambda value for domain adversarial training.

Set random seed for reproducibility.

Set random seed for reproducibility.

Setup EVA-CLIP repository if model requires it. Args: config: Configuration dict Returns: bool: True if setup was succe

Add EVA-CLIP repository to sys.path. Args: custom_path: Optional custom path to EVA-CLIP repo (e.g., '/kaggle/input/my-evaclip-repo/

Add EVA-CLIP repository to sys.path.

Smooth L1 (Huber) loss with explicit beta for compatibility across PyTorch versions. Matches torch.nn.SmoothL1Loss(beta=beta, reduction='mea

Train a single fold with gradual unfreezing.

Train a single fold with gradual unfreezing.

Train for one epoch (supports both 3-head and 5-head models, including binning and Gaussian). Args: mixup_alpha: Beta distribution p

Train for one epoch with component logging and optional auxiliary tasks.

Train for one epoch with View Consistency regularization. Uses BiomassDatasetSplitViewConsistency which returns (orig, flip) pairs. Enfo

Validate using OOF approach - sum predictions from left and right halves. Supports standard 3-head, 5-head, and Gaussian (10-output) models.

Validate using OOF approach (sum left + right predictions).

Validate using OOF approach with 3-crop - sum predictions from left, center, right and multiply by 2/3. Supports standard 3-head, 5-head, and

Compare inference predictions with OOF predictions to verify inference pipeline. This function: 1. Loads OOF predictions 2. Matches

Args: model_name: timm model name pretrained: use pretrained weights dropout: dropout rate for heads