source/model.py

from __future__ import annotations

import torch
from torch import nn
from torch.nn import functional as F


def binary_weight(weight: torch.Tensor) -> torch.Tensor:
    scale = weight.abs().mean(dim=1, keepdim=True).clamp_min(1e-6)
    quantized = scale * torch.where(weight >= 0, 1.0, -1.0)
    return weight + (quantized - weight).detach()


def ternary_weight(weight: torch.Tensor) -> torch.Tensor:
    scale = weight.abs().mean(dim=1, keepdim=True).clamp_min(1e-6)
    normalized = weight / scale
    quantized = scale * normalized.round().clamp(-1, 1)
    return weight + (quantized - weight).detach()


class QuantizedLinear(nn.Linear):
    def __init__(self, *args, bits: str = "fp32", **kwargs) -> None:
        super().__init__(*args, **kwargs)
        self.bits = bits

    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
        if self.bits == "binary":
            weight = binary_weight(self.weight)
        elif self.bits == "ternary":
            weight = ternary_weight(self.weight)
        else:
            weight = self.weight
        return F.linear(inputs, weight, self.bias)


class BitMLP(nn.Module):
    def __init__(self, bits: str = "fp32") -> None:
        super().__init__()
        self.bits = bits
        self.hidden = QuantizedLinear(64, 64, bits=bits)
        self.output = QuantizedLinear(64, 10, bits=bits)

    def forward(self, images: torch.Tensor) -> torch.Tensor:
        flattened = images.flatten(1)
        return self.output(F.gelu(self.hidden(flattened)))


def parameter_count(module: nn.Module) -> int:
    return sum(parameter.numel() for parameter in module.parameters())