params

nerfstudio/configs/method_configs.py

@@ -602,7 +602,7 @@ method_configs["splatfacto"] = TrainerConfig(
     pipeline=VanillaPipelineConfig(
         datamanager=FullImageDatamanagerConfig(
             _target=FullImageDatamanager[DepthDataset],
-            dataparser=NerfstudioDataParserConfig(load_3D_points=True,depth_unit_scale_factor=1.0),
+            dataparser=NerfstudioDataParserConfig(load_3D_points=True,depth_unit_scale_factor=1.0,train_split_fraction=.5),
             cache_images_type="uint8",
         ),
         model=SplatfactoModelConfig(
@@ -630,14 +630,14 @@ method_configs["splatfacto"] = TrainerConfig(
             "scheduler": None,
         },
         "scales": {
-            "optimizer": AdamOptimizerConfig(lr=0.005, eps=1e-15),
+            "optimizer": AdamOptimizerConfig(lr=0.003, eps=1e-15),
             "scheduler": None,
         },
         "quats": {"optimizer": AdamOptimizerConfig(lr=0.001, eps=1e-15), "scheduler": None},
         "camera_opt": {
-            "optimizer": AdamOptimizerConfig(lr=2e-5, eps=1e-15),
+            "optimizer": AdamOptimizerConfig(lr=1e-5, eps=1e-15),
             "scheduler": ExponentialDecaySchedulerConfig(
-                lr_final=5e-7, max_steps=30000, warmup_steps=3000, lr_pre_warmup=0
+                lr_final=5e-7, max_steps=30000, warmup_steps=6000, lr_pre_warmup=0
             ),
         },
     },

nerfstudio/models/splatfacto.py

@@ -145,7 +145,7 @@ class SplatfactoModelConfig(ModelConfig):
     """below this size, gaussians are *duplicated*, otherwise split"""
     n_split_samples: int = 2
     """number of samples to split gaussians into"""
-    sh_degree_interval: int = 3000
+    sh_degree_interval: int = 2000
     """every n intervals turn on another sh degree"""
     split_screen_size: float = 0.05
     """if a gaussian is more than this percent of screen space, split it"""
@@ -820,7 +820,7 @@ class SplatfactoModel(Model):
             #convert to float32
             mask = self._downscale_if_required(batch["mask"].float())
             mask = mask.to(self.device)
-            mask[mask<1]=0.0
+            mask[mask<1] = 0.0
             assert mask.shape[:2] == gt_img.shape[:2] == pred_img.shape[:2]
             gt_img = gt_img * mask
             pred_img = pred_img * mask
@@ -842,8 +842,8 @@ class SplatfactoModel(Model):
         if 'depth_image' in batch:
             gt_depth = self.get_gt_img(batch['depth_image'])
             ignores = gt_depth <= .001
-            # if "mask" in batch:
-            #     ignores = ignores | (mask == 0)
+            if "mask" in batch:
+                ignores = ignores | (mask == 0)
             depth_loss = (gt_depth - outputs['depth'])[~ignores].abs().mean()
         else:
             depth_loss = torch.tensor(0.0).to(self.device)