inference.py

import os
import torch
import trimesh
import numpy as np
from pathlib import Path
import time

# Import from MagicArticulate
from skeleton_models.skeletongen import SkeletonGPT
from data_utils.save_npz import normalize_to_unit_cube
from utils.mesh_to_pc import MeshProcessor
from utils.save_utils import (
    pred_joints_and_bones,
    save_skeleton_to_txt,
    merge_duplicate_joints_and_fix_bones,
    save_skeleton_obj,
    save_mesh
)

class SkeletonInferencer:
    """Wrapper class for skeleton generation inference"""
    
    def __init__(self, pretrained_weights, device="cuda", precision="fp16"):
        self.device = device
        self.precision = precision
        
        # Create args object
        class Args:
            def __init__(self):
                self.llm = "facebook/opt-350m"
                self.pad_id = -1
                self.n_discrete_size = 128
                self.n_max_bones = 100
                self.num_beams = 1
                self.seed = 0
        
        self.args = Args()
        
        # Load model
        print(f"Loading model from {pretrained_weights}...")
        self.model = SkeletonGPT(self.args).to(device)
        
        pkg = torch.load(pretrained_weights, map_location=torch.device("cpu"))
        self.model.load_state_dict(pkg["model"])
        self.model.eval()
        
        # Set precision
        if precision == "fp16" and device == "cuda":
            self.model = self.model.half()
        
        print("Model loaded successfully!")
    
    @torch.no_grad()
    def infer(
        self,
        input_path,
        output_dir,
        input_pc_num=8192,
        apply_marching_cubes=False,
        octree_depth=7,
        sequence_type="spatial"
    ):
        """
        Run inference on a single mesh file
        
        Returns:
            dict: Results including paths and statistics
        """
        start_time = time.time()
        
        output_dir = Path(output_dir)
        output_dir.mkdir(parents=True, exist_ok=True)
        
        # Load mesh
        mesh = trimesh.load(input_path, force='mesh')
        
        # Convert to point cloud
        if apply_marching_cubes:
            pc_list = MeshProcessor.convert_meshes_to_point_clouds(
                [mesh], input_pc_num, 
                apply_marching_cubes=True,
                octree_depth=octree_depth
            )
            pc_normal = pc_list[0]
        else:
            # Simple sampling
            points, face_indices = trimesh.sample.sample_surface(mesh, input_pc_num)
            normals = mesh.face_normals[face_indices]
            pc_normal = np.concatenate([points, normals], axis=-1)
        
        # Normalize point cloud
        pc_coor = pc_normal[:, :3]
        normals = pc_normal[:, 3:]
        pc_coor, center, scale = normalize_to_unit_cube(pc_coor, scale_factor=0.9995)
        
        # Prepare transform parameters
        bounds = np.array([pc_coor.min(axis=0), pc_coor.max(axis=0)])
        pc_center = (bounds[0] + bounds[1]) / 2
        pc_scale = (bounds[1] - bounds[0]).max() + 1e-5
        
        transform_params = torch.tensor([
            center[0], center[1], center[2], scale,
            pc_center[0], pc_center[1], pc_center[2], pc_scale
        ], dtype=torch.float32)
        
        # Prepare batch data
        pc_normal_normalized = np.concatenate([pc_coor, normals], axis=-1)
        batch_data = {
            'pc_normal': torch.from_numpy(pc_normal_normalized).half().unsqueeze(0).to(self.device),
            'transform_params': transform_params.unsqueeze(0),
            'vertices': torch.from_numpy(mesh.vertices).unsqueeze(0),
            'faces': torch.from_numpy(mesh.faces).unsqueeze(0),
            'file_name': [Path(input_path).stem]
        }
        
        # Generate skeleton
        pred_bone_coords = self.model.generate(batch_data)
        
        # Process results
        file_name = Path(input_path).stem
        skeleton = pred_bone_coords[0].cpu().numpy()
        pred_joints, pred_bones = pred_joints_and_bones(skeleton.squeeze())
        
        # Post-process
        hier_order = (sequence_type == "hierarchical")
        if hier_order and len(pred_bones) > 0:
            pred_root_index = pred_bones[0][0]
            pred_joints, pred_bones, pred_root_index = merge_duplicate_joints_and_fix_bones(
                pred_joints, pred_bones, root_index=pred_root_index
            )
        else:
            pred_joints, pred_bones = merge_duplicate_joints_and_fix_bones(
                pred_joints, pred_bones
            )
            pred_root_index = None
        
        # Denormalize for saving
        trans = transform_params[:3].numpy()
        scale_val = transform_params[3].item()
        pc_trans = transform_params[4:7].numpy()
        pc_scale_val = transform_params[7].item()
        
        pred_joints_denorm = pred_joints * pc_scale_val + pc_trans
        pred_joints_denorm = pred_joints_denorm / scale_val + trans
        
        # Save files
        pred_rig_filename = output_dir / f"{file_name}_pred.txt"
        pred_skel_filename = output_dir / f"{file_name}_skel.obj"
        mesh_filename = output_dir / f"{file_name}_mesh.obj"
        
        save_skeleton_to_txt(
            pred_joints_denorm, pred_bones, pred_root_index,
            hier_order, mesh.vertices, str(pred_rig_filename)
        )
        
        save_skeleton_obj(
            pred_joints, pred_bones, str(pred_skel_filename),
            pred_root_index if hier_order else None,
            use_cone=hier_order
        )
        
        # Save normalized mesh
        vertices_norm = (mesh.vertices - trans) * scale_val
        vertices_norm = (vertices_norm - pc_trans) / pc_scale_val
        save_mesh(vertices_norm, mesh.faces, str(mesh_filename))
        
        elapsed_time = time.time() - start_time
        
        return {
            'skeleton_file': str(pred_skel_filename),
            'rig_file': str(pred_rig_filename),
            'mesh_file': str(mesh_filename),
            'num_joints': len(pred_joints),
            'num_bones': len(pred_bones),
            'time': elapsed_time
        }