"""CNN perception stage used by Sudoku training from scratch.""" from __future__ import annotations import torch import torch.nn as nn import torch.nn.functional as F from softasp_solver.base_model import BaseSoftASPModel from softasp_solver.registry import register_model from softasp_solver.sudoku.softasp_sudoku import CellCNN @register_model("perception_only_sudoku") class PerceptionOnlySudoku(BaseSoftASPModel): """Perception-only baseline: CNN encoder + direct classification. No constraint reasoning. Blank cells are predicted from their (empty) image alone, so accuracy on blank cells should be ~11% (random guessing among 9). Args: d_model: Embedding dimension. """ def __init__(self, d_model: int = 128) -> None: super().__init__() self.d_model = d_model self.cell_cnn = CellCNN(d_model=d_model) self.head = nn.Linear(d_model, 9) def _extract_cells(self, images: torch.Tensor) -> torch.Tensor: B = images.shape[0] x = images.reshape(B, 1, 9, 28, 9, 28) x = x.permute(0, 2, 4, 1, 3, 5) return x.reshape(B * 81, 1, 28, 28) def forward(self, batch: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]: images = batch["images"] labels = batch["labels"] puzzle_mask = batch["puzzle_mask"] # (B, 81) int64, 1 = clue cell B = images.shape[0] if images.dim() == 4 and images.shape[-1] == 1: images = images.permute(0, 3, 1, 2) if images.dtype != torch.float32: images = images.float() / 255.0 cells = self._extract_cells(images) embeds = self.cell_cnn(cells).reshape(B, 81, self.d_model) logits = self.head(embeds) # (B, 81, 9) # CE loss only on clue cells — blank cells are featureless black # patches; the model cannot learn anything from them. clue_flat = puzzle_mask.reshape(-1).bool() # (B*81,) loss = F.cross_entropy( logits.reshape(-1, 9)[clue_flat], labels.reshape(-1)[clue_flat], ) predictions = logits.argmax(dim=-1) # (B, 81) # Flat clue-cell tensors for the clue_accuracy metric in the evaluator clue_logits = logits.reshape(-1, 9)[clue_flat] # (n_clue, 9) clue_labels = labels.reshape(-1)[clue_flat] # (n_clue,) return { "loss": loss, "logits": logits, "predictions": predictions, "clue_logits": clue_logits, "clue_labels": clue_labels, }