Building/src_new/evaluate.py

from __future__ import annotations

import argparse
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch

try:
    from .config import ExperimentConfig, make_rms_forward_config
    from .dataset import get_dataloaders, report_to_text
    from .model import build_model
except ImportError:
    from config import ExperimentConfig, make_rms_forward_config
    from dataset import get_dataloaders, report_to_text
    from model import build_model


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Evaluate direct RMS regression model on harmonic data.")
    parser.add_argument("--split", choices=("train", "val", "test"), default="val")
    parser.add_argument("--sample-index", type=int, default=0)
    parser.add_argument("--device", type=str, default=None)
    parser.add_argument("--checkpoint", type=str, default=None)
    parser.add_argument("--all-samples", action="store_true")
    return parser.parse_args()


def resolve_device(device_name: str | None, config: ExperimentConfig) -> torch.device:
    requested = device_name or config.train.device
    if requested.startswith("cuda") and not torch.cuda.is_available():
        return torch.device("cpu")
    return torch.device(requested)


def resolve_checkpoint_path(config: ExperimentConfig, checkpoint_arg: str | None) -> Path:
    if checkpoint_arg:
        return Path(checkpoint_arg).resolve()
    return config.data.project_root / config.train.checkpoint_dir / "forward_rms" / config.train.best_model_name


def relative_percent_error(true_value: float, pred_value: float) -> float:
    denominator = max(abs(true_value), 1e-12)
    return abs(pred_value - true_value) / denominator * 100.0


def inverse_waveform(array: torch.Tensor, mean: torch.Tensor, std: torch.Tensor) -> torch.Tensor:
    mean = mean.to(array.device, dtype=array.dtype)
    std = std.to(array.device, dtype=array.dtype)
    return array * std + mean


def predict_sample(
    model: torch.nn.Module,
    sample: dict[str, object],
    device: torch.device,
    y_scaler,
    x_scaler,
    y_wave_mean: torch.Tensor,
    y_wave_std: torch.Tensor,
) -> dict[str, float | str | np.ndarray]:
    x = sample["x"].unsqueeze(0).to(device)
    aux = sample["aux"].unsqueeze(0).to(device)
    mask = sample["mask"].unsqueeze(0).to(device)
    x_rms_raw = sample["x_rms_raw"].unsqueeze(0).to(device)
    y_true = float(sample["y_raw"].detach().cpu().numpy().reshape(-1)[0])
    frequency_hz = float(sample["frequency_hz"].item())
    file_name = str(sample["file_name"])
    valid_mask = sample["mask"].detach().cpu().numpy() > 0.5

    with torch.no_grad():
        outputs = model(x, mask, aux)

    pred_target = y_scaler.inverse_transform(outputs["rms"])
    y_pred = float((torch.exp(pred_target) * x_rms_raw).detach().cpu().numpy().reshape(-1)[0])
    pred_wave = inverse_waveform(outputs["waveform"], y_wave_mean, y_wave_std).detach().cpu().numpy().reshape(-1)[valid_mask]
    true_wave = inverse_waveform(sample["y_wave"], y_wave_mean.cpu(), y_wave_std.cpu()).detach().cpu().numpy().reshape(-1)[valid_mask]
    x_wave = x_scaler.inverse_transform(sample["x"]).reshape(-1)[valid_mask]
    return {
        "file_name": file_name,
        "frequency_hz": frequency_hz,
        "true_rms": y_true,
        "pred_rms": y_pred,
        "relative_error_percent": float(relative_percent_error(y_true, y_pred)),
        "x_wave": x_wave.numpy().reshape(-1),
        "true_wave": true_wave,
        "pred_wave": pred_wave,
    }


def evaluate_sample(
    model: torch.nn.Module,
    sample: dict[str, object],
    device: torch.device,
    y_scaler,
    x_scaler,
    y_wave_mean: torch.Tensor,
    y_wave_std: torch.Tensor,
) -> dict[str, float | str]:
    result = predict_sample(model, sample, device, y_scaler, x_scaler, y_wave_mean, y_wave_std)
    return {
        "file_name": str(result["file_name"]),
        "frequency_hz": float(result["frequency_hz"]),
        "true_rms": float(result["true_rms"]),
        "pred_rms": float(result["pred_rms"]),
        "relative_error_percent": float(result["relative_error_percent"]),
    }


def save_single_sample_figure(result: dict[str, float | str | np.ndarray], split: str, save_dir: Path, sample_index: int) -> Path:
    true_rms = float(result["true_rms"])
    pred_rms = float(result["pred_rms"])
    error_percent = relative_percent_error(true_rms, pred_rms)
    x_signal = np.asarray(result["x_wave"], dtype=np.float64).reshape(-1)
    true_wave = np.asarray(result["true_wave"], dtype=np.float64).reshape(-1)
    pred_wave = np.asarray(result["pred_wave"], dtype=np.float64).reshape(-1)
    time_steps = np.arange(x_signal.shape[0], dtype=np.float64)

    fig, axes = plt.subplots(3, 1, figsize=(12, 10))
    fig.suptitle(f"Forward RMS regression | {split} | {result['file_name']}")
    axes[0].plot(time_steps, x_signal, linewidth=1.2)
    axes[0].set_title("Input Base Excitation (Steady-State Segment)")
    axes[0].set_xlabel("Time Step")
    axes[0].set_ylabel("Acceleration")
    axes[0].grid(True, alpha=0.3)
    axes[1].plot(time_steps, true_wave, label="True response", linewidth=1.2, color="tab:blue")
    axes[1].plot(time_steps, pred_wave, label="Pred response", linewidth=1.2, color="tab:orange", alpha=0.85)
    axes[1].set_title("Auxiliary Waveform Head: True vs Predicted Top Response")
    axes[1].set_xlabel("Time Step")
    axes[1].set_ylabel("Acceleration")
    axes[1].grid(True, alpha=0.3)
    axes[1].legend()
    axes[2].bar(["True RMS", "Pred RMS"], [true_rms, pred_rms], color=["tab:blue", "tab:orange"])
    axes[2].set_title(f"True RMS: {true_rms:.4f}, Pred RMS: {pred_rms:.4f}, Error: {error_percent:.2f}%")
    axes[2].set_ylabel("RMS")
    axes[2].grid(True, axis="y", alpha=0.3)
    plt.tight_layout()
    save_dir.mkdir(parents=True, exist_ok=True)
    figure_path = save_dir / f"evaluation_{split}_s{sample_index}_waveform.png"
    plt.savefig(figure_path, dpi=180, bbox_inches="tight")
    plt.show()
    plt.close(fig)
    return figure_path


def main() -> None:
    args = parse_args()
    config = make_rms_forward_config()
    device = resolve_device(args.device, config)
    checkpoint_path = resolve_checkpoint_path(config, args.checkpoint)
    loaders, datasets, reports, x_scaler, aux_scaler, y_scaler = get_dataloaders(config)
    print(report_to_text(reports))

    if not checkpoint_path.exists():
        raise FileNotFoundError(f"Checkpoint not found: {checkpoint_path}")

    dataset = datasets[args.split]
    sample = dataset[args.sample_index]
    normalization = datasets["train"].normalization
    model = build_model(config).to(device)
    checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
    model.load_state_dict(checkpoint["model_state_dict"])
    model.eval()
    save_dir = config.data.project_root / "evaluation_outputs" / "forward_rms"
    if args.all_samples:
        rows = [
            evaluate_sample(
                model,
                dataset[index],
                device,
                y_scaler,
                x_scaler,
                normalization.y_wave_mean,
                normalization.y_wave_std,
            )
            for index in range(len(dataset))
        ]
        result_df = pd.DataFrame(rows)
        result_df = result_df.sort_values(["relative_error_percent", "file_name"]).reset_index(drop=True)
        csv_path = save_dir / f"evaluation_{args.split}_all_samples.csv"
        save_dir.mkdir(parents=True, exist_ok=True)
        result_df.to_csv(csv_path, index=False)
        summary = {
            "count": len(result_df),
            "mean_error_percent": float(result_df["relative_error_percent"].mean()),
            "median_error_percent": float(result_df["relative_error_percent"].median()),
            "max_error_percent": float(result_df["relative_error_percent"].max()),
            "min_error_percent": float(result_df["relative_error_percent"].min()),
        }
        print(f"Checkpoint: {checkpoint_path}")
        print(f"Summary: {summary}")
        print(f"CSV saved to: {csv_path}")
        print(result_df.to_string(index=False))
        return

    result = predict_sample(
        model,
        sample,
        device,
        y_scaler,
        x_scaler,
        normalization.y_wave_mean,
        normalization.y_wave_std,
    )
    figure_path = save_single_sample_figure(
        result=result,
        split=args.split,
        save_dir=save_dir,
        sample_index=args.sample_index,
    )
    print(f"Checkpoint: {checkpoint_path}")
    print(f"Sample file: {result['file_name']}")
    print(f"True RMS: {result['true_rms']:.6f}")
    print(f"Pred RMS: {result['pred_rms']:.6f}")
    print(f"Relative RMS Error (%): {result['relative_error_percent']:.4f}")
    print(f"Figure saved to: {figure_path}")


if __name__ == "__main__":
    main()