wabdalm's picture
Release AS2 framework with Hydra-selected instantiations
69928d6 verified
Raw History Blame Contribute Delete
2.56 kB
"""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,
}