model.py

import torch
import torch.nn as nn
import torch.nn.init as init

# --- CONFIGURATION ---
INPUT_CELLS = 81
NUM_CLASSES = 10
HIDDEN_DIM = 128
ATTN_HEADS = 4   # MUST match training script

class StandardAttention2D(nn.Module):
    """
    Standard O(N^2) Multi-Head Attention for 2D grids.
    Zero-initialized output projection to start as identity.
    """
    def __init__(self, dim, heads=ATTN_HEADS):
        super().__init__()
        self.scale = dim ** -0.5
        self.heads = heads
        self.head_dim = dim // heads

        self.to_qkv = nn.Conv2d(dim, dim * 3, kernel_size=1, bias=False)
        self.to_out = nn.Sequential(
            nn.Conv2d(dim, dim, kernel_size=1),
            nn.GroupNorm(8, dim)
        )

        # Zero-init so attention starts as a no-op
        init.zeros_(self.to_out[0].weight)
        init.zeros_(self.to_out[0].bias)

    def forward(self, x):
        b, c, h, w = x.shape
        n = h * w

        qkv = self.to_qkv(x).view(b, 3 * c, n)
        q, k, v = qkv.chunk(3, dim=1)

        q = q.view(b, self.heads, self.head_dim, n).permute(0, 1, 3, 2)
        k = k.view(b, self.heads, self.head_dim, n).permute(0, 1, 3, 2)
        v = v.view(b, self.heads, self.head_dim, n).permute(0, 1, 3, 2)

        dots = (q @ k.transpose(-2, -1)) * self.scale
        attn = dots.softmax(dim=-1)

        out = (attn @ v).transpose(1, 2).reshape(b, c, h, w)
        return self.to_out(out) + x


class UniversalPotato(nn.Module):
    """
    EXACT match to the Colab-trained HybridPotato architecture.
    No positional embeddings. Blindfold-compatible.
    """
    def __init__(self):
        super().__init__()

        self.embed_clues = nn.Embedding(NUM_CLASSES, HIDDEN_DIM)
        self.embed_board = nn.Embedding(NUM_CLASSES, HIDDEN_DIM)

        self.input_proj = nn.Sequential(
            nn.Conv2d(HIDDEN_DIM * 3, HIDDEN_DIM, kernel_size=1),
            nn.GroupNorm(8, HIDDEN_DIM),
            nn.SiLU()
        )

        self.core = nn.Sequential(
            # Local
            nn.Conv2d(HIDDEN_DIM, HIDDEN_DIM, 3, padding=1),
            nn.GroupNorm(8, HIDDEN_DIM),
            nn.SiLU(),

            # Global
            StandardAttention2D(HIDDEN_DIM),
            nn.SiLU(),

            # Mid-range
            nn.Conv2d(HIDDEN_DIM, HIDDEN_DIM, 3, padding=2, dilation=2),
            nn.GroupNorm(8, HIDDEN_DIM),
            nn.SiLU(),

            # Processing
            nn.Conv2d(HIDDEN_DIM, HIDDEN_DIM, 3, padding=4, dilation=4),
            nn.GroupNorm(8, HIDDEN_DIM),
            nn.SiLU()
        )

        self.head = nn.Conv2d(HIDDEN_DIM, NUM_CLASSES, kernel_size=1)
        self.memory_norm = nn.GroupNorm(8, HIDDEN_DIM)

    def run_core(self, x):
        return self.core(x)

    def forward(self, clues, current_board, memory, blindfold=False):
        b, n = clues.shape

        clues_emb = (
            self.embed_clues(clues)
            .transpose(1, 2)
            .view(b, HIDDEN_DIM, 9, 9)
        )

        board_emb = (
            self.embed_board(current_board)
            .transpose(1, 2)
            .view(b, HIDDEN_DIM, 9, 9)
        )

        if blindfold:
            board_emb = torch.zeros_like(board_emb)

        raw = torch.cat([clues_emb, board_emb, memory], dim=1)
        z = self.input_proj(raw)
        z = self.core(z)

        new_memory = self.memory_norm(memory + z)

        logits = (
            self.head(z)
            .view(b, NUM_CLASSES, 81)
            .transpose(1, 2)
        )

        return logits, new_memory