训练sample修改

brep2sdf/networks/sample.py

@@ -3,7 +3,7 @@ import torch
 
 class NormalPerPoint():
 
-    def __init__(self, global_sigma, local_sigma=0.1):
+    def __init__(self, global_sigma, local_sigma=0.5):
         self.global_sigma = global_sigma
         self.local_sigma = local_sigma
 
@@ -31,11 +31,11 @@ class NormalPerPoint():
         """
         sample_size, dim = pc_input.shape
 
-        # 生成随机位移值
+        # 生成随机位移值（确保有正有负）
         if local_sigma is not None:
-            psdf = torch.randn(sample_size, device=pc_input.device) * local_sigma
+            psdf = (torch.rand(sample_size, device=pc_input.device) * 2 - 1) * local_sigma
         else:
-            psdf = torch.randn(sample_size, device=pc_input.device) * self.local_sigma
+            psdf = (torch.rand(sample_size, device=pc_input.device) * 2 - 1) * self.local_sigma
 
         # 沿法线方向偏移
         sample = pc_input + normals * psdf.unsqueeze(-1)