| """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) |
| |
| 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() |
|
|
