File size: 3,663 Bytes
bffb03e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 | """Run multi-scale reconstruction inference."""
import importlib.util
from pathlib import Path
import numpy as np, torch, yaml
import argparse
import random
ROOT = Path(__file__).resolve().parents[1]
def validate_config(cfg):
data, model = cfg["data"], cfg["model"]
if data["image_size"] != model["image_size"] or data["channels"] != model["in_channels"]:
raise ValueError("data and model image shape settings must match")
if data["scales"] != model["scales"]:
raise ValueError("data.scales and model.scales must match")
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--config", type=Path, default=ROOT / "conf/config.yaml")
parser.add_argument("--seed", type=int, default=None)
parser.add_argument("--device", choices=("auto", "cpu", "cuda"), default=None)
parser.add_argument("--data", type=Path, default=None)
parser.add_argument("--checkpoint", type=Path, default=None)
parser.add_argument("--output-dir", type=Path, default=None)
args = parser.parse_args()
cfg = yaml.safe_load(args.config.read_text())
validate_config(cfg)
seed = cfg["seed"] if args.seed is None else args.seed
random.seed(seed); np.random.seed(seed); torch.manual_seed(seed)
spec = importlib.util.spec_from_file_location("satmaepp", ROOT / "model/satmaepp.py"); module = importlib.util.module_from_spec(spec); spec.loader.exec_module(module)
model_cfg = {k: v for k, v in cfg["model"].items() if k not in {"architecture", "runtime_profile"}}
requested = args.device or cfg["runtime"]["device"]
if requested == "cuda" and not torch.cuda.is_available(): raise RuntimeError("CUDA was requested but is unavailable")
device = torch.device("cuda" if requested == "cuda" or (requested == "auto" and torch.cuda.is_available()) else "cpu")
model = module.SatMAEPP(**model_cfg).to(device)
checkpoint = args.checkpoint or ROOT / cfg["paths"]["checkpoint"]
if not checkpoint.exists(): raise FileNotFoundError("Run training before inference")
model.load_state_dict(torch.load(checkpoint, map_location="cpu", weights_only=False)["model"]); model.eval()
archive = np.load(args.data or ROOT / cfg["data"]["root"] / "test.npz")
images = torch.from_numpy(archive["images"])
expected = (cfg["data"]["channels"], cfg["data"]["image_size"], cfg["data"]["image_size"])
if images.ndim != 4 or tuple(images.shape[1:]) != expected:
raise ValueError(f"test images must have shape [N, {expected[0]}, {expected[1]}, {expected[2]}]")
targets = {}
for scale in cfg["model"]["scales"]:
if scale != 1:
field = f"images_{scale}x"
if field not in archive:
raise ValueError(f"test dataset is missing native target {field}")
targets[str(scale)] = torch.from_numpy(archive[field])
images, targets = images.to(device), {key: value.to(device) for key, value in targets.items()}
with torch.no_grad(): output = model(images, high_resolution_targets=targets)
out = args.output_dir or ROOT / cfg["paths"]["inference_dir"]; out.mkdir(parents=True, exist_ok=True)
payload = {"target": images.cpu().numpy(), "prediction": output["reconstruction"].cpu().numpy(), "mask": output["mask"].cpu().numpy(), "labels": archive["labels"]}
for scale, value in output["predictions"].items():
payload[f"prediction_{scale}"] = value.cpu().numpy()
payload[f"target_{scale}"] = output["targets"][scale].cpu().numpy()
np.savez_compressed(out / "reconstruction.npz", **payload)
print("inference=", out / "reconstruction.npz")
if __name__ == "__main__": main()
|