source/packing.py

from __future__ import annotations

from pathlib import Path

import numpy as np
import torch
from model import BitMLP


def pack_binary_model(model: BitMLP, destination: Path) -> dict:
    payload = {}
    payload_bytes = 0
    for name in ["hidden", "output"]:
        layer = getattr(model, name)
        weight = layer.weight.detach().cpu().numpy()
        scales = np.mean(np.abs(weight), axis=1).astype(np.float32)
        signs = np.packbits((weight >= 0).reshape(-1), bitorder="little")
        shape = np.asarray(weight.shape, dtype=np.int32)
        bias = layer.bias.detach().cpu().numpy().astype(np.float32)
        payload[f"{name}_signs"] = signs
        payload[f"{name}_scales"] = scales
        payload[f"{name}_shape"] = shape
        payload[f"{name}_bias"] = bias
        payload_bytes += signs.nbytes + scales.nbytes + bias.nbytes
    np.savez(destination, **payload)
    return {
        "packed_payload_bytes": int(payload_bytes),
        "container_bytes": int(destination.stat().st_size),
    }


def load_binary_model(source: Path) -> BitMLP:
    packed = np.load(source)
    model = BitMLP(bits="binary")
    with torch.no_grad():
        for name in ["hidden", "output"]:
            layer = getattr(model, name)
            shape = tuple(packed[f"{name}_shape"].astype(int))
            total = int(np.prod(shape))
            bits = np.unpackbits(
                packed[f"{name}_signs"], bitorder="little"
            )[:total]
            signs = np.where(bits.reshape(shape) == 1, 1.0, -1.0)
            scales = packed[f"{name}_scales"][:, None]
            layer.weight.copy_(
                torch.from_numpy((signs * scales).astype(np.float32))
            )
            layer.bias.copy_(torch.from_numpy(packed[f"{name}_bias"]))
    return model