https://github.com/galmetzer/dipole-normal-prop
Raw File
Tip revision: 0887b893b153f4ce6b09e4dd485f9b042f15b706 authored by Gal Metzer on 08 September 2021, 21:53 UTC
Update README.md
Tip revision: 0887b89
orient_large.py
import field_utils
import options
from pathlib import Path
from field_utils import *
from util import orient_center

from inference_utils import load_model_from_file, fix_n_filter, voting_policy
torch.manual_seed(1)

def run(opts):
    export_path: Path = opts.export_dir
    export_path.mkdir(exist_ok=True)

    max_patch_size = 500
    device = torch.device(torch.cuda.current_device() if torch.cuda.is_available() else torch.device('cpu'))
    MyTimer = util.timer_factory()
    with MyTimer('load pc', count=False):
        input_pc = util.xyz2tensor(open(opts.pc, 'r').read(), append_normals=False).to(device)

    input_pc, transform = util.Transform.trans(input_pc)

    if opts.estimate_normals:
        with MyTimer('estimating normals'):
            input_pc = util.estimate_normals(input_pc, max_nn=opts.n)

    softmax = torch.nn.Softmax(dim=-1)


    n_models = len(opts.models)
    models = [load_model_from_file(opts.models[i], device) for i in range(n_models)]

    with MyTimer('divide patches'):
        patch_indices = util.divide_pc(input_pc[:, :3], opts.number_parts,
                                       min_patch=opts.minimum_points_per_patch)
        all_patches_indices = [x.clone() for x in patch_indices]

    with MyTimer('filter patches'):
        patch_indices = fix_n_filter(input_pc, patch_indices, opts.curvature_threshold)

    num_patches = len(patch_indices)
    num_all_patches = len(all_patches_indices)
    print(f'number of patches {num_patches}/{num_all_patches}')

    with MyTimer('orient center'):
        for i, p in patch_indices:
            input_pc[p] = orient_center(input_pc[p])

    with MyTimer('find reps'):
        represent = []
        for p in all_patches_indices:
            permutation = torch.randperm(p.shape[0])
            represent.append((p[permutation[:max_patch_size]], p[permutation[max_patch_size:]]))

    pc_probs = torch.ones_like(input_pc[:, 0])

    with MyTimer('network orientation'):
        for i, _ in patch_indices:
            with torch.no_grad():
                current_reps, non_reps_points = represent[i]
                data = input_pc[current_reps]
                data = data.to(device)
                for _ in range(opts.iters):
                    votes = [model(data.clone()) for model in models]
                    vote_probabilities = [softmax(scores)[:, 1] for scores in votes]
                    flip, probs = voting_policy(vote_probabilities)
                    pc_probs[current_reps] = probs
                    input_pc[current_reps[flip], 3:] *= -1


    [model.to('cpu') for model in models]
    with MyTimer('propagating field'):
        strongest_field_propagation_reps(input_pc, represent, diffuse=True)

    with MyTimer('fix global orientation'):
        if field_utils.measure_mean_potential(input_pc) < 0:
            # if average global potential is negative, flip all normals
            input_pc[:, 3:] *= -1

    with MyTimer('exporting result', count=False):
        util.export_pc(transform.inverse(input_pc).transpose(0, 1), export_path / f'final_result.xyz')

    MyTimer.print_total_time()


if __name__ == '__main__':
    opts = options.get_parser().parse_args()
    opts.export_dir.mkdir(exist_ok=True, parents=True)

    options.export_options(opts)
    run(opts)
back to top