model.py

"""Vanilla 3-layer feedforward ANN for modular multiplication.

Input:  digit-encoded (a_red, b_red, p) — each zero-padded to MAX_DIGITS
Hidden: one fully-connected layer with ReLU
Output: MAX_OUT_DIGITS * 10 logits — one 10-way class per output digit position

At inference, a and b are reduced mod p inside predict_digits (allowed by rules)
before being encoded and fed to the network.
"""

from __future__ import annotations

from pathlib import Path

import torch
import torch.nn as nn

from modchallenge.interface.base_model import ModularMultiplicationModel

MAX_DIGITS = 10      # decimal digits per input slot (covers p up to ~4 billion = tier 4)
MAX_OUT_DIGITS = 10  # decimal digits in the answer (answer < p <= tier-4 max)
INPUT_SIZE = 3 * MAX_DIGITS          # 30
HIDDEN_SIZE = 1024
OUTPUT_SIZE = MAX_OUT_DIGITS * 10    # 100


# ---------------------------------------------------------------------------
# Architecture
# ---------------------------------------------------------------------------

class VanillaMLP(nn.Module):
    def __init__(
        self,
        input_size: int = INPUT_SIZE,
        hidden_size: int = HIDDEN_SIZE,
        output_size: int = OUTPUT_SIZE,
    ):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, output_size),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """x: (B, INPUT_SIZE) floats in [0, 1]. returns (B, MAX_OUT_DIGITS, 10) logits."""
        out = self.net(x)                          # (B, OUTPUT_SIZE)
        return out.view(-1, MAX_OUT_DIGITS, 10)    # (B, D, 10)


# ---------------------------------------------------------------------------
# Encoding helpers (shared by train.py and predict_digits)
# ---------------------------------------------------------------------------

def encode_int(n: int, length: int = MAX_DIGITS) -> list[int]:
    """Integer -> zero-padded decimal digit list of fixed length, MSB first."""
    s = str(int(n)).zfill(length)
    if len(s) > length:
        s = s[-length:]   # truncate if somehow too long
    return [int(c) for c in s]


def digits_to_tensor(n: int, length: int = MAX_DIGITS) -> list[float]:
    """Encode integer n as normalized floats in [0, 1]."""
    return [d / 9.0 for d in encode_int(n, length)]


# ---------------------------------------------------------------------------
# Submission entry point
# ---------------------------------------------------------------------------

class VanillaModel(ModularMultiplicationModel):
    def __init__(self):
        self.model: VanillaMLP | None = None
        self.device: torch.device | None = None

    def load(self, model_dir: str) -> None:
        if torch.cuda.is_available():
            self.device = torch.device("cuda")
        else:
            self.device = torch.device("cpu")

        ckpt = torch.load(
            Path(model_dir) / "weights.pt",
            map_location=self.device,
            weights_only=True,
        )
        cfg = ckpt.get("config", {})
        self.model = VanillaMLP(
            input_size=cfg.get("input_size", INPUT_SIZE),
            hidden_size=cfg.get("hidden_size", HIDDEN_SIZE),
            output_size=cfg.get("output_size", OUTPUT_SIZE),
        )
        self.model.load_state_dict(ckpt["state_dict"])
        self.model.to(self.device)
        self.model.eval()

    def preprocess_a(self, a: str):
        return a

    def preprocess_b(self, b: str):
        return b

    def preprocess_p(self, p: str):
        return p

    @torch.no_grad()
    def predict_digits(self, a_enc, b_enc, p_enc) -> list[int]:
        assert self.model is not None

        p = int(p_enc)
        a_red = int(a_enc) % p
        b_red = int(b_enc) % p

        x = digits_to_tensor(a_red) + digits_to_tensor(b_red) + digits_to_tensor(p)
        inp = torch.tensor([x], dtype=torch.float32, device=self.device)  # (1, 30)

        logits = self.model(inp)          # (1, MAX_OUT_DIGITS, 10)
        preds = logits[0].argmax(-1).tolist()   # [d0, d1, ..., d9]

        # Strip leading zeros, return at least [0]
        result = preds
        while len(result) > 1 and result[0] == 0:
            result = result[1:]
        return result