a big cleanup: data fix seed; ngp cubic box

examples/datasets/nerf_360_v2.py

@@ -258,6 +258,8 @@ class SubjectLoader(torch.utils.data.Dataset):
         )
         self.K = torch.tensor(self.K).to(torch.float32).to(device)
         self.height, self.width = self.images.shape[1:3]
+        self.g = torch.Generator(device=device)
+        self.g.manual_seed(42)
 
     def __len__(self):
         return len(self.images)
@@ -274,7 +276,7 @@ class SubjectLoader(torch.utils.data.Dataset):
 
         if self.training:
             if self.color_bkgd_aug == "random":
-                color_bkgd = torch.rand(3, device=self.images.device)
+                color_bkgd = torch.rand(3, device=self.images.device, generator=self.g)
             elif self.color_bkgd_aug == "white":
                 color_bkgd = torch.ones(3, device=self.images.device)
             elif self.color_bkgd_aug == "black":
@@ -304,14 +306,15 @@ class SubjectLoader(torch.utils.data.Dataset):
                     len(self.images),
                     size=(num_rays,),
                     device=self.images.device,
+                    generator=self.g,
                 )
             else:
                 image_id = [index] * num_rays
             x = torch.randint(
-                0, self.width, size=(num_rays,), device=self.images.device
+                0, self.width, size=(num_rays,), device=self.images.device, generator=self.g
             )
             y = torch.randint(
-                0, self.height, size=(num_rays,), device=self.images.device
+                0, self.height, size=(num_rays,), device=self.images.device, generator=self.g
             )
         else:
             image_id = [index]

examples/datasets/nerf_synthetic.py

@@ -124,6 +124,8 @@ class SubjectLoader(torch.utils.data.Dataset):
         self.camtoworlds = self.camtoworlds.to(device)
         self.K = self.K.to(device)
         assert self.images.shape[1:3] == (self.HEIGHT, self.WIDTH)
+        self.g = torch.Generator(device=device)
+        self.g.manual_seed(42)
 
     def __len__(self):
         return len(self.images)
@@ -141,7 +143,7 @@ class SubjectLoader(torch.utils.data.Dataset):
 
         if self.training:
             if self.color_bkgd_aug == "random":
-                color_bkgd = torch.rand(3, device=self.images.device)
+                color_bkgd = torch.rand(3, device=self.images.device, generator=self.g)
             elif self.color_bkgd_aug == "white":
                 color_bkgd = torch.ones(3, device=self.images.device)
             elif self.color_bkgd_aug == "black":
@@ -172,14 +174,15 @@ class SubjectLoader(torch.utils.data.Dataset):
                     len(self.images),
                     size=(num_rays,),
                     device=self.images.device,
+                    generator=self.g,
                 )
             else:
                 image_id = [index] * num_rays
             x = torch.randint(
-                0, self.WIDTH, size=(num_rays,), device=self.images.device
+                0, self.WIDTH, size=(num_rays,), device=self.images.device, generator=self.g
             )
             y = torch.randint(
-                0, self.HEIGHT, size=(num_rays,), device=self.images.device
+                0, self.HEIGHT, size=(num_rays,), device=self.images.device, generator=self.g
             )
         else:
             image_id = [index]

examples/radiance_fields/ngp.py

@@ -85,6 +85,13 @@ class NGPRadianceField(torch.nn.Module):
         super().__init__()
         if not isinstance(aabb, torch.Tensor):
             aabb = torch.tensor(aabb, dtype=torch.float32)
+
+        # Turns out rectangle aabb will leads to uneven collision so bad performance.
+        # We enforce a cube aabb here.
+        center = (aabb[..., :num_dim] + aabb[..., num_dim:]) / 2.0
+        size = (aabb[..., num_dim:] - aabb[..., :num_dim]).max()
+        aabb = torch.cat([center - size / 2.0, center + size / 2.0], dim=-1)
+
         self.register_buffer("aabb", aabb)
         self.num_dim = num_dim
         self.use_viewdirs = use_viewdirs