Forward pass. Args: x: [B, C, H, W] preprocessed images Returns: pred_total, pred_gdm, pred_green:

Forward pass. Args: x: [B, C, H, W] preprocessed images Returns: pred_total, pred_gdm, pred_green,

Forward pass with Height FiLM conditioning. Args: x: Image tensor [B, C, H, W] height_gt: Ground truth heigh

Forward pass. Args: x: [B, C, H, W] preprocessed images return_variance: If True, return (mu, var) tuples. I

Forward pass. Args: x: [B, C, H, W] preprocessed images Returns: pred_total, pred_gdm, pred_green,

Forward pass. Args: x: [B, C, H, W] preprocessed images (float16) Returns: pred_total, pred_gdm, pr

Forward pass. Args: x: [B, C, H, W] preprocessed images return_logits: if True, return raw logits; if False,

Calculate weighted CrossEntropy loss. Args: logits_*: [B, num_bins] raw logits target_*: [B, 1] continuous t

Calculate weighted prediction loss + consistency loss. Returns: total_loss, pred_loss, consistency_loss, individual loss

Calculate Gaussian NLL loss + consistency loss. Args: mu_*: [B, 1] predicted means for each target var_*: [B

Compute weighted ranking loss across all 5 targets. Args: pred_*: (B, 1) predictions for each target true_*:

Forward pass. Args: x: [B, C, H, W] preprocessed images Returns: pred_total, pred_gdm, pred_green:

Forward pass. Args: x: [B, C, H, W] preprocessed images Returns: pred_total, pred_gdm, pred_green,

Forward pass - returns only mu (mean) predictions for inference. Returns: (mu_total, mu_gdm, mu_green, mu_dead, mu_clove

Forward pass with Height FiLM conditioning. For inference, always uses predicted height. Returns: Tuple of (pred

Forward pass. Args: x: Input tensor [B, C, H, W] return_components: If True, also return global preds and de

Args: x_2d: [B, D, H, W] - patch features in 2D spatial format Returns: density: [B, 5, H, W]

Forward pass. Args: x: Input tensor [B, C, H, W] depth: Optional depth tensor [B, 1, H, W] for weighted inte

Forward pass. Args: x: [B, C, H, W] preprocessed images Returns: pred_total, pred_gdm, pred_green,

Forward pass.

Transform for InternViT using CLIPImageProcessor.

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

Unfreeze the last n transformer blocks of the vision encoder. Args: n: Number of blocks to unfreeze from the end

Unfreeze the last n transformer blocks of the vision encoder. Args: n: Number of blocks to unfreeze from the end

Unfreeze the last n transformer blocks of the vision encoder.

Unfreeze the last n transformer blocks of the vision encoder.

Unfreeze the last n transformer blocks.