"""Run validated, batched autoregressive solar forecasting inference."""

import importlib.util
from pathlib import Path

import numpy as np
import torch
import yaml

ROOT = Path(__file__).resolve().parents[1]


def main():
    import argparse
    parser = argparse.ArgumentParser(); parser.add_argument("--config", type=Path, default=ROOT / "conf/config.yaml"); parser.add_argument("--data", type=Path); parser.add_argument("--checkpoint", type=Path); parser.add_argument("--output-dir", type=Path); parser.add_argument("--batch-size", type=int, default=4); parser.add_argument("--device", choices=("auto", "cpu", "cuda"), default=None); args = parser.parse_args()
    cfg = yaml.safe_load(args.config.read_text())
    spec = importlib.util.spec_from_file_location("surya_model", ROOT / "model/surya.py")
    module = importlib.util.module_from_spec(spec); spec.loader.exec_module(module)
    model = module.Surya(**cfg["model"])
    checkpoint = args.checkpoint or ROOT / cfg["paths"]["checkpoint"]
    if not checkpoint.exists(): raise FileNotFoundError("Run training before inference")
    device_name = args.device or cfg["runtime"]["device"]
    if device_name == "cuda" and not torch.cuda.is_available(): raise RuntimeError("CUDA requested but unavailable")
    device = torch.device("cuda" if torch.cuda.is_available() and device_name != "cpu" else "cpu")
    model.load_state_dict(torch.load(checkpoint, map_location="cpu", weights_only=False)["model"]); model.to(device).eval()
    data = np.load(args.data or ROOT / cfg["data"]["root"] / "test.npz"); raw_inputs, raw_targets = data["inputs"], data["targets"]
    if raw_inputs.ndim != 5 or raw_inputs.shape[1:] != (2, 13, cfg["data"]["image_size"], cfg["data"]["image_size"]): raise ValueError("test.npz violates the BTCHW protocol")
    mean = np.asarray(cfg["data"]["channel_mean"], dtype=np.float32)[None, None, :, None, None]; std = np.asarray(cfg["data"]["channel_std"], dtype=np.float32)[None, None, :, None, None]
    inputs = (np.sign(raw_inputs) * np.log1p(np.abs(raw_inputs)) - mean) / std
    predictions = []
    with torch.no_grad():
        for start in range(0, len(inputs), args.batch_size): predictions.append(model(torch.from_numpy(inputs[start:start+args.batch_size]).to(device), steps=cfg["data"]["forecast_steps"]).cpu().numpy())
    predictions = np.concatenate(predictions, axis=0)
    # Evaluation consumes the original physical-value representation.
    predictions = np.sign(predictions * std[:, :1] + mean[:, :1]) * (np.expm1(np.abs(predictions * std[:, :1] + mean[:, :1])))
    output = args.output_dir or ROOT / cfg["paths"]["inference_dir"]; output.mkdir(parents=True, exist_ok=True)
    np.savez_compressed(output / "forecast.npz", inputs=raw_inputs, targets=raw_targets,
                        predictions=predictions, activity=data["activity"])
    print("inference=", output / "forecast.npz")


if __name__ == "__main__": main()