export_checkpoint.py

"""Export a region chkpnt_fine_N.pth into a snapshot model dir the renderer can load.

Usage: python export_checkpoint.py <model_dir> <iteration>
Writes <model_dir>_<iteration>/point_cloud/iteration_<iteration>/ plus resolved_config.json.
"""
import json
import shutil
import sys
from argparse import Namespace
from pathlib import Path

import torch

sys.path.insert(0, "/home/anujaya/venumi/4DGaussians")
import numpy as np  # noqa: E402
from gaussian_renderer import render  # noqa: E402
from scene import GaussianModel, Scene  # noqa: E402

model_dir, iteration = Path(sys.argv[1]).resolve(), int(sys.argv[2])
out_dir = model_dir.with_name(f"{model_dir.name}_{iteration}")
settings = json.loads((model_dir / "resolved_config.json").read_text())
out_dir.mkdir(exist_ok=True)
settings["model"]["model_path"] = str(out_dir)
gaussians = GaussianModel(settings["model"]["sh_degree"], Namespace(**settings["hidden"]))
# Scene supplies the cameras; its fresh point-cloud init is overwritten by the checkpoint below.
scene = Scene(Namespace(**settings["model"]), gaussians, shuffle=False)

state = torch.load(model_dir / f"chkpnt_fine_{iteration}.pth", weights_only=False, map_location="cuda")
assert state["stage"] == "fine" and state["iteration"] == iteration, (state["stage"], state["iteration"])
(gaussians.active_sh_degree, gaussians._xyz, deform_state, gaussians._deformation_table,
 gaussians._features_dc, gaussians._features_rest, gaussians._scaling, gaussians._rotation,
 gaussians._opacity, *_rest) = state["model"]
gaussians._deformation.load_state_dict(deform_state)
gaussians._deformation_accum = state["deformation_accum"]

point_cloud = out_dir / "point_cloud" / f"iteration_{iteration}"
point_cloud.mkdir(parents=True, exist_ok=True)
gaussians.save_ply(str(point_cloud / "point_cloud.ply"))
gaussians.save_deformation(str(point_cloud))
# The renderer verifies its reload against this probe, as it does for a finished training run.
with torch.no_grad():
    probe = render(scene.getVideoCameras()[0], gaussians, Namespace(**settings["pipeline"]),
                   torch.zeros(3, device="cuda"), stage="fine", cam_type=scene.dataset_type)["render"]
np.save(out_dir / "final_render_probe.npy", probe.cpu().numpy())
(out_dir / "resolved_config.json").write_text(json.dumps(settings, indent=2))
shutil.copy(model_dir / "cfg_args", out_dir / "cfg_args")
print(f"exported {len(gaussians._xyz)} Gaussians -> {point_cloud}")