加了全局采样训练

brep2sdf/data/pre_process_by_mesh.py

@@ -433,7 +433,7 @@ def parse_solid(step_path,sample_normal_vector=False,sample_sdf_points=False):
             data['sampled_points_normals_sdf'] = sample_sdf_points_and_normals(
                 trimesh_mesh_ncs=trimesh_mesh_ncs,
                 surf_bbox_ncs=data['surf_bbox_ncs'],
-                num_sdf_samples=4096, # <-- 传递固定数量
+                num_sdf_samples=50000, # <-- 传递固定数量
                 sdf_sampling_std_dev=0.0001
             )
         else:

brep2sdf/train.py

@@ -39,7 +39,7 @@ parser.add_argument(
     help='强制重新进行数据预处理，忽略缓存或已有结果'
 )
 parser.add_argument(
-    '--resume-checkpoint-path', 
+    '--resume-checkpoint-path', '-r',
     type=str,
     default=None,
     help='从指定的checkpoint文件继续训练'
@@ -99,6 +99,7 @@ class Trainer:
             )
             # 合并表面点数据和采样点数据
             self.sdf_data = torch.cat([surface_sdf_data, sampled_sdf_data], dim=0)
+            self.train_surf_ncs = sampled_sdf_data
         else:
             self.sdf_data = surface_sdf_data
         print_data_distribution(self.sdf_data)
@@ -403,7 +404,7 @@ class Trainer:
                 logger.debug(f"Patch [{patch_id:2d}] | train pnt number {points.shape[0]}")
                 
                 mnfld_pnts = points[:, 0:3]
-                gt_sdf = torch.zeros(mnfld_pnts.shape[0], device=self.device)
+                gt_sdf = points[:, -1]
                 
                 normals = points[:, 3:6]