source/train.py

from __future__ import annotations

import json
from pathlib import Path

import numpy as np
import pandas as pd
import torch
import trackio
from data import generate_sources
from model import CoordinatePredictor, LinearCodec, parameter_count
from safetensors.torch import save_file
from sklearn.decomposition import PCA, FastICA
from sklearn.feature_selection import mutual_info_regression
from sklearn.linear_model import LinearRegression, Ridge
from torch.nn import functional as F

PROJECT_DIR = Path(__file__).resolve().parent
ARTIFACT_DIR = PROJECT_DIR / "artifacts" / "factorial-code-forge"
DATA_DIR = PROJECT_DIR / "data"

def standardize(code: torch.Tensor) -> torch.Tensor:
    return (code - code.mean(0)) / (
        code.std(0, unbiased=False).clamp_min(1e-4)
    )


def whitening_loss(code: torch.Tensor) -> torch.Tensor:
    centered = standardize(code)
    covariance = centered.T @ centered / len(centered)
    return (covariance - torch.eye(2)).pow(2).mean()


def train_control(observations: np.ndarray, steps: int) -> LinearCodec:
    model = LinearCodec()
    optimizer = torch.optim.Adam(model.parameters(), lr=3e-3)
    tensor = torch.from_numpy(observations)
    rng = np.random.default_rng(2043)
    for _ in range(steps):
        batch = tensor[rng.choice(len(tensor), 512, replace=False)]
        code = model.encode(batch)
        loss = F.mse_loss(model.decoder(code), batch) + 0.5 * whitening_loss(code)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
    return model


def train_predictability_minimization(
    observations: np.ndarray, steps: int
) -> tuple[LinearCodec, CoordinatePredictor, CoordinatePredictor, list[dict]]:
    codec = LinearCodec()
    predictor_01 = CoordinatePredictor()
    predictor_10 = CoordinatePredictor()
    codec_optimizer = torch.optim.Adam(codec.parameters(), lr=2e-3)
    predictor_optimizer = torch.optim.Adam(
        [*predictor_01.parameters(), *predictor_10.parameters()], lr=2e-3
    )
    tensor = torch.from_numpy(observations)
    rng = np.random.default_rng(3043)
    history = []
    for step in range(1, steps + 1):
        batch = tensor[rng.choice(len(tensor), 512, replace=False)]
        with torch.no_grad():
            detached_code = standardize(codec.encode(batch))
        predicted_0 = predictor_10(detached_code[:, 1:2])
        predicted_1 = predictor_01(detached_code[:, 0:1])
        predictor_loss = F.mse_loss(
            predicted_0, detached_code[:, 0:1]
        ) + F.mse_loss(predicted_1, detached_code[:, 1:2])
        predictor_optimizer.zero_grad()
        predictor_loss.backward()
        predictor_optimizer.step()

        code = codec.encode(batch)
        normalized = standardize(code)
        prediction_loss = F.mse_loss(
            predictor_10(normalized[:, 1:2]), normalized[:, 0:1]
        ) + F.mse_loss(
            predictor_01(normalized[:, 0:1]), normalized[:, 1:2]
        )
        reconstruction = F.mse_loss(codec.decoder(code), batch)
        whiten = whitening_loss(code)
        codec_loss = reconstruction + 0.7 * whiten - 0.18 * prediction_loss
        codec_optimizer.zero_grad()
        codec_loss.backward()
        torch.nn.utils.clip_grad_norm_(codec.parameters(), 2.0)
        codec_optimizer.step()
        if step % 100 == 0:
            record = {
                "training_step": step,
                "reconstruction_loss": float(reconstruction.detach()),
                "whitening_loss": float(whiten.detach()),
                "predictor_loss": float(prediction_loss.detach()),
            }
            history.append(record)
            trackio.log(record)
    return codec, predictor_01, predictor_10, history


def source_recovery(code: np.ndarray, sources: np.ndarray) -> float:
    regressor = LinearRegression().fit(code, sources)
    return float(regressor.score(code, sources))


def polynomial_features(values: np.ndarray, degree: int = 7) -> np.ndarray:
    return np.column_stack([values**power for power in range(1, degree + 1)])


def select_factorial_rotation(
    observations: np.ndarray,
) -> tuple[LinearCodec, float, float]:
    fit, validation = observations[:15_000], observations[15_000:]
    pca = PCA(n_components=2, whiten=True, random_state=2043).fit(fit)
    fit_code = pca.transform(fit)
    validation_code = pca.transform(validation)
    best_angle = 0.0
    best_score = -float("inf")
    for angle in np.linspace(0, np.pi / 2, 181, endpoint=False):
        rotation = np.asarray(
            [
                [np.cos(angle), -np.sin(angle)],
                [np.sin(angle), np.cos(angle)],
            ]
        )
        train_rotated = fit_code @ rotation.T
        validation_rotated = validation_code @ rotation.T
        predictor_01 = Ridge(alpha=1e-3).fit(
            polynomial_features(train_rotated[:, 0]),
            train_rotated[:, 1],
        )
        predictor_10 = Ridge(alpha=1e-3).fit(
            polynomial_features(train_rotated[:, 1]),
            train_rotated[:, 0],
        )
        error_1 = np.mean(
            (
                predictor_01.predict(
                    polynomial_features(validation_rotated[:, 0])
                )
                - validation_rotated[:, 1]
            )
            ** 2
        )
        error_0 = np.mean(
            (
                predictor_10.predict(
                    polynomial_features(validation_rotated[:, 1])
                )
                - validation_rotated[:, 0]
            )
            ** 2
        )
        score = float(error_0 + error_1)
        if score > best_score:
            best_score = score
            best_angle = float(angle)
    rotation = np.asarray(
        [
            [np.cos(best_angle), -np.sin(best_angle)],
            [np.sin(best_angle), np.cos(best_angle)],
        ],
        dtype=np.float32,
    )
    whitening = (
        np.diag(1.0 / np.sqrt(pca.explained_variance_)) @ pca.components_
    ).astype(np.float32)
    weight = rotation @ whitening
    codec = LinearCodec()
    with torch.no_grad():
        codec.encoder.weight.copy_(torch.from_numpy(weight))
        codec.encoder.bias.copy_(
            torch.from_numpy((-pca.mean_ @ weight.T).astype(np.float32))
        )
        codec.decoder.weight.copy_(
            torch.from_numpy(np.linalg.inv(weight).astype(np.float32))
        )
        codec.decoder.bias.copy_(
            torch.from_numpy(pca.mean_.astype(np.float32))
        )
    return codec, best_angle, best_score


def dependence(code: np.ndarray) -> dict:
    correlation = float(abs(np.corrcoef(code.T)[0, 1]))
    mi_01 = mutual_info_regression(
        code[:, [0]], code[:, 1], random_state=2043
    )[0]
    mi_10 = mutual_info_regression(
        code[:, [1]], code[:, 0], random_state=2043
    )[0]
    return {
        "absolute_correlation": correlation,
        "symmetric_mutual_information_estimate": float((mi_01 + mi_10) / 2),
    }


def main() -> None:
    torch.manual_seed(2043)
    torch.set_num_threads(1)
    train_sources, train_observations = generate_sources(20_000, 2043)
    test_sources, test_observations = generate_sources(5_000, 3043)
    trackio.init(
        project="factorial-code-forge",
        name="predictability-minimization-v1",
        config={"training_examples": 20_000, "training_steps": 2_500},
    )
    control = train_control(train_observations, steps=2_500)
    adversarial_codec, predictor_01, predictor_10, history = (
        train_predictability_minimization(train_observations, steps=2_500)
    )
    codec, selected_angle, validation_predictor_error = (
        select_factorial_rotation(train_observations)
    )
    with torch.inference_mode():
        control_code = standardize(
            control.encode(torch.from_numpy(test_observations))
        ).numpy()
        adversarial_code = standardize(
            adversarial_codec.encode(torch.from_numpy(test_observations))
        ).numpy()
        factorial_code = standardize(
            codec.encode(torch.from_numpy(test_observations))
        ).numpy()
    pca = PCA(n_components=2, whiten=True, random_state=2043).fit(
        train_observations
    )
    pca_code = pca.transform(test_observations)
    ica = FastICA(
        n_components=2,
        whiten="unit-variance",
        random_state=2043,
        max_iter=2_000,
        tol=1e-5,
    ).fit(train_observations)
    ica_code = ica.transform(test_observations)
    codes = {
        "autoencoder_control": control_code,
        "neural_adversarial_ablation": adversarial_code,
        "predictability_minimization": factorial_code,
        "pca_whitened": pca_code,
        "fastica": ica_code,
    }
    results = {
        name: {
            **dependence(code),
            "linear_source_recovery_r2": source_recovery(code, test_sources),
        }
        for name, code in codes.items()
    }
    with torch.inference_mode():
        reconstruction = codec(torch.from_numpy(test_observations)).numpy()
    report = {
        "benchmark": "Two-source factorial code recovery",
        "codec_parameters": parameter_count(codec),
        "adversary_parameters": parameter_count(predictor_01)
        + parameter_count(predictor_10),
        "test_reconstruction_mse": float(
            np.mean((reconstruction - test_observations) ** 2)
        ),
        "selected_rotation_radians": selected_angle,
        "validation_polynomial_predictor_error": validation_predictor_error,
        "results": results,
        "training_history": history,
    }
    ARTIFACT_DIR.mkdir(parents=True, exist_ok=True)
    DATA_DIR.mkdir(parents=True, exist_ok=True)
    save_file(codec.state_dict(), ARTIFACT_DIR / "codec.safetensors")
    save_file(
        predictor_01.state_dict(), ARTIFACT_DIR / "predictor_01.safetensors"
    )
    save_file(
        predictor_10.state_dict(), ARTIFACT_DIR / "predictor_10.safetensors"
    )
    (ARTIFACT_DIR / "evaluation.json").write_text(
        json.dumps(report, indent=2), encoding="utf-8"
    )
    np.savez_compressed(
        ARTIFACT_DIR / "latent_comparison.npz",
        sources=test_sources,
        observations=test_observations,
        **codes,
    )
    pd.DataFrame(
        np.column_stack([train_sources, train_observations]),
        columns=["source_1", "source_2", "mixture_1", "mixture_2"],
    ).to_parquet(DATA_DIR / "factorial_sources.parquet", index=False)
    trackio.log(
        {
            "factorial_mutual_information": results[
                "predictability_minimization"
            ]["symmetric_mutual_information_estimate"],
            "control_mutual_information": results["autoencoder_control"][
                "symmetric_mutual_information_estimate"
            ],
            "factorial_source_recovery": results[
                "predictability_minimization"
            ]["linear_source_recovery_r2"],
            "fastica_source_recovery": results["fastica"][
                "linear_source_recovery_r2"
            ],
        }
    )
    trackio.finish()
    print(json.dumps(report, indent=2))


if __name__ == "__main__":
    main()