| """Configuration-specific target eligibility; never clamp or rewrite labels.""" |
| from __future__ import annotations |
|
|
| import math |
|
|
| import torch |
|
|
| from .objective import supervision_masks |
|
|
|
|
| def size_target_conflicts(batch, model_config, loss_config): |
| """Report learned, supervised axes outside reference * exp(+-limit). |
| |
| Complete-field masks follow the criterion. Fixed request dimensions bypass |
| the learned size head. Missing labels are selected out before log arithmetic; |
| disabled size/box objectives impose no size-target range requirement. |
| A 1e-6 log tolerance only handles float32 rounding at the exact endpoints. |
| """ |
| masks = supervision_masks(batch) |
| slots = torch.zeros_like(batch["slot_mask"]) |
| if loss_config.size > 0: |
| slots = slots | masks["size"] |
| if loss_config.box > 0: |
| slots = slots | masks["box"] |
| fixed = batch.get("fixed_size_mask", torch.zeros_like(batch["validity"]["size"])) |
| coordinates = slots.unsqueeze(-1) & ~fixed |
| conflicts = [] |
| for b, i, q in coordinates.nonzero().tolist(): |
| target = float(batch["targets"]["size"][b, i, q]) |
| reference = model_config.size_reference[q] |
| log_ratio = math.log(target) - math.log(reference) |
| limit = model_config.size_log_limit |
| if abs(log_ratio) > limit + 1e-6: |
| conflicts = conflicts + [{"batch_index": b, "object_id": batch["objects"][b][i]["id"], |
| "dimension": ("w", "d", "h")[q], "target_m": target, |
| "minimum_m": reference * math.exp(-limit), |
| "maximum_m": reference * math.exp(limit)}] |
| return conflicts |
|
|
