Compute globally weighted R² score (official Kaggle competition metric). All (image, target) pairs are combined into one long list, and each

Create a Gaussian head as ModuleList: [hidden, mu, log_var].

Create a Gaussian head as ModuleList: [hidden, mu, log_var].

Forward pass through a Gaussian head ModuleList.

Forward pass through a Gaussian head, returning only mu.

Compute pairwise ranking loss for a single target. For all pairs (i, j) where true[i] > true[j]: loss += max(0, margin -

Convert bin logits to expected value (regression output). Args: logits: [B, num_bins] raw logits bin_centers_tensor: [num_bi

Override train to keep backbone in eval mode (frozen).

Convert continuous values to bin indices. Args: values: np.array of values edges: np.array of bin edges [e0, e1, e2, ...]

Convert patch tokens [B, N, C] to 2D spatial grid [B, C, H, W].

Convert patch tokens [B, N, C] to 2D spatial grid [B, C, H, W].

Compute globally weighted R2 score (official Kaggle competition metric). Formula: SS_res = sum(w_i * (y_i - y_pred_i)^2) SS_

Get vision transformer blocks for gradual unfreezing.

Load a model for TTT (with trainable heads).

Main inference function with fold ensemble. Args: config: Configuration dict model_dir: Directory containing model checkpoin

Unfreeze the last n transformer blocks of the vision encoder.

Apply Gaussian blur with specified radius.

Apply Gaussian blur with specified radius.

Load left or right half of an image.

Optional physics-consistent post-processing. mode: - "none": no changes - "total_correction": Total = Green + Dead + Clover

Optional physics-consistent post-processing. mode: - "none": no changes - "total_correction": Total = Green + Dead + Clover

Optionally append label-preserving "quality diversity" augmentation based on config.

Calculate Dry_Dead_g and Dry_Clover_g from predicted values.

Create a debug CSV with predictions for sampled train images.

Create submission file from predictions.

Create submission file from predictions.

Default augmentation (same as child-exp000).

Load YAML configuration file.

Load YAML configuration file.

Run prediction with one fold model.

Return the first existing path from candidates.

Main inference function with fold ensemble. Args: config: Configuration dict model_dir: Directory containing model checkpoin

Set patch dropout probability. Useful for enabling/disabling during training (epoch control) or MC inference. Args:

Set patch dropout probability for MC inference.

Train for one epoch (TTT mode - heads only).

Robust z-score using per-sample median/IQR. residual: [B, 1, H, W] -> z: [B, 1, H, W]

Initialize log_var bias to 0 (variance = 1).

Load depth map and process to patch grid size.

Load depth map and split into left/right halves.

Load a single fold model for inference. Returns model or None if not found. Args: fold: Fold number config: Configuratio

Load a single fold model for inference. Args: fold: Fold number config: Configuration dict model_dir: Directory cont

Normalize height using log1p (handles right-skewed distribution).

Apply CutMix between left and right patches of the same image. Since both patches have the same label (y/2), the mixed label remains y/2.

Build augmentation to apply on the FULL 2000x1000 image BEFORE splitting. This is used to keep left/right halves consistent (same photometric/geo

Depth attention regularization (train-only; inference unchanged). Builds attention from detrended depth residual: attn = relu(z)^heigh

Compute per-image sample weights (length = len(df_wide)). Intended for train-only reweighting / oversampling; weights are applied to both ha

Compute view consistency loss between original and flipped predictions. Enforces: pred(original) ≈ pred(flipped) Args:

Create bin edges and centers from config. Args: bin_edges_config: dict of {target_name: [edge0, edge1, ...]} Returns: b

Create debug output with predictions vs ground truth.

Create a debug CSV with predictions for sampled train images.

Create train/val split for a specific fold. Uses StratifiedGroupKFold by State and Sampling_Date (same as EXP060-000).

Generate Out-of-Fold predictions after training. Uses OOF approach: sum predictions from left and right halves. When enable_tta=True, ou

Generate Out-of-Fold predictions after training.

Create submission file.

Convert depth map to integration weight. Uses rank normalization (robust to outliers) and mean=1 normalization to keep densi

Get vision transformer blocks for gradual unfreezing.

Get vision transformer blocks for gradual unfreezing.

Get vision transformer blocks for gradual unfreezing.

Get vision transformer blocks for gradual unfreezing.

Get vision transformer blocks for gradual unfreezing.

Calculate GRL lambda with warmup schedule. Uses sigmoid schedule: λ(p) = λ_max × (2 / (1 + exp(-10p)) - 1) where p = epoch / total_epoch

Calculate GRL lambda with warmup schedule. Uses sigmoid schedule: λ(p) = λ_max × (2 / (1 + exp(-10p)) - 1) where p = epoch / total_epoch

Enhanced augmentation pipeline for training. All augmentations preserve the amount of grass in the image. Args: aug_version: 'de

Augmentation pipeline for training. All augmentations preserve the amount of grass in the image. Args: config: Optional config d

Augmentation pipeline for 3-crop dataset. IMPORTANT: This is applied to the FULL 2000x1000 image BEFORE splitting. Therefore, we must EX

Create a schedule for gradual block unfreezing for Vision Transformer. Args: total_epochs: Total number of training epochs f

Create a schedule for gradual block unfreezing. Returns: List of tuples (epoch, num_blocks_to_unfreeze, backbone_lr)

Check if model is an EVA-CLIP model.

Load YAML configuration file.

Load YAML configuration file.