Clean up

Author: d2nerf

configs/decompose/train.gin

@@ -32,8 +32,6 @@ TrainConfig.log_every = 100
 TrainConfig.histogram_every = 100
 TrainConfig.save_every = 2000
 TrainConfig.use_background_loss = False
-TrainConfig.use_bg_decompose_loss = False
-TrainConfig.bg_decompose_loss_weight = 1.0
 # TrainConfig.blendw_loss_weight = 0.001
 # TrainConfig.blendw_loss_weight_schedule = ('constant', 0.001)
 TrainConfig.blendw_loss_weight_schedule = {
@@ -43,20 +41,9 @@ TrainConfig.blendw_loss_weight_schedule = {
   'num_steps': 200000,
 }
 TrainConfig.blendw_loss_skewness = 1.25
-TrainConfig.blendw_pixel_loss_weight_schedule = ('constant', 0.0)
-TrainConfig.blendw_ray_loss_weight = 0.001
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = True
 TrainConfig.blendw_area_loss_weight = 0.0001
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 
-# initialize static model for some iterations first
-TrainConfig.fix_blendw_steps = 0
-TrainConfig.force_blendw_steps = 0
-TrainConfig.force_blendw_loss_weight = .1
-TrainConfig.fix_blendw_value = 0.15
 
 DecomposeNerfModel.blend_mode = 'nsff'
 DecomposeNerfModel.extra_renders = ('static_full', 'dynamic', 'blendw')

configs/decompose/train_kubric_template.gin

@@ -23,13 +23,8 @@ TrainConfig.blendw_loss_weight_schedule = {
   'num_steps': 75000 #%max_steps,
 }
 TrainConfig.blendw_loss_skewness = 2.5
-TrainConfig.blendw_ray_loss_weight = 0.
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = False
 TrainConfig.blendw_area_loss_weight = 0.00001
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 
 DecomposeNerfModel.separate_shadow_model = True
 # TrainConfig.shadow_r_loss_weight = ('constant', 0.01)
@@ -39,7 +34,6 @@ TrainConfig.shadow_r_loss_weight = {
   'final_value': 0.1,
   'num_steps': 100000 #%max_steps,
 }
-TrainConfig.shadow_r_l2_loss_weight = 0.01
 DecomposeNerfModel.sigma_threshold = 0.1
 DecomposeNerfModel.shadow_r_shift = 0.
 

configs/decompose/train_lf_template.gin

@@ -22,21 +22,8 @@ TrainConfig.blendw_loss_weight_schedule = {
   }
 TrainConfig.blendw_loss_skewness = 3.0
 
-# TrainConfig.blendw_pixel_loss_weight_schedule = {
-#   'type': 'exp_increase',
-#   'initial_value': 1e-06,
-#   'final_value': 0.01,
-#   'num_steps': 100000,
-#   }
-# TrainConfig.blendw_pixel_loss_skewness = 1.0
-
-TrainConfig.blendw_ray_loss_weight = 0.
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = False
 TrainConfig.blendw_area_loss_weight = 0.0001
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 TrainConfig.sigma_s_ray_loss_weight = 0.0
 
 DecomposeNerfModel.separate_shadow_model = True
@@ -48,7 +35,6 @@ TrainConfig.shadow_r_loss_weight = {
     (0, ('linear', 0.0005, 1.0, 90000))
   ],
 }
-TrainConfig.shadow_r_l2_loss_weight = 0.
 DecomposeNerfModel.sigma_threshold = 0.1
 DecomposeNerfModel.shadow_r_shift = 0.
 

configs/rl/001.gin

@@ -22,13 +22,8 @@ TrainConfig.blendw_loss_weight_schedule = {
 }
 TrainConfig.blendw_loss_skewness = 1.75
 
-TrainConfig.blendw_ray_loss_weight = 0.
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = False
 TrainConfig.blendw_area_loss_weight = 0.001
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 
 DecomposeNerfModel.separate_shadow_model = True
 TrainConfig.shadow_r_loss_weight = {
@@ -37,7 +32,6 @@ TrainConfig.shadow_r_loss_weight = {
   'final_value': 0.1,
   'num_steps': 100000 #%max_steps,
 }
-TrainConfig.shadow_r_l2_loss_weight = 0.01
 DecomposeNerfModel.sigma_threshold = 0.1
 DecomposeNerfModel.shadow_r_shift = 0.
 

configs/rl/002.gin

@@ -21,13 +21,8 @@ TrainConfig.blendw_loss_weight_schedule = {
   }
 TrainConfig.blendw_loss_skewness = 3.0
 
-TrainConfig.blendw_ray_loss_weight = 0.
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = False
 TrainConfig.blendw_area_loss_weight = 0.0001
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 TrainConfig.sigma_s_ray_loss_weight = 0.001
 
 DecomposeNerfModel.separate_shadow_model = True
@@ -37,7 +32,6 @@ TrainConfig.shadow_r_loss_weight = {
   'final_value': 0.1,
   'num_steps': 100000 #%max_steps,
 }
-TrainConfig.shadow_r_l2_loss_weight = 0.
 DecomposeNerfModel.sigma_threshold = 0.1
 DecomposeNerfModel.shadow_r_shift = 0.
 

configs/rl/005.gin

@@ -23,28 +23,16 @@ TrainConfig.blendw_loss_weight_schedule = {
   'num_steps': 100000,
 }
 TrainConfig.blendw_loss_skewness = 1.5
-TrainConfig.blendw_ray_loss_weight = 0
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = False
 TrainConfig.blendw_area_loss_weight = 0.1
 TrainConfig.sigma_s_ray_loss_weight = 0.0
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 TrainConfig.sigma_s_ray_loss_weight = 0.01
 
 DecomposeNerfModel.separate_shadow_model = True
 # TrainConfig.shadow_r_loss_weight = ('constant', 0.01)
 TrainConfig.shadow_r_loss_weight = ('constant', 0.01)
 
-# TrainConfig.shadow_r_consistency_loss_weight_schedule = {
-#   'type': 'linear',
-#   'initial_value': 0.00001,
-#   'final_value': 0.1,
-#   'num_steps': 50000 #%max_steps,
-# }
 
-TrainConfig.shadow_r_l2_loss_weight = 0.0
 DecomposeNerfModel.sigma_threshold = 0.1
 DecomposeNerfModel.shadow_r_shift = 0.
 

configs/synthetic/002.gin

@@ -22,13 +22,8 @@ TrainConfig.blendw_loss_weight_schedule = {
   'num_steps': 75000 #%max_steps,
 }
 TrainConfig.blendw_loss_skewness = 1.25
-TrainConfig.blendw_ray_loss_weight = 0.
-TrainConfig.blendw_ray_loss_threshold = .5
 TrainConfig.use_ex_ray_entropy_loss = False
 TrainConfig.blendw_area_loss_weight = 0.0001
-TrainConfig.shadow_loss_threshold = 0.2
-TrainConfig.shadow_loss_weight = 0.
-TrainConfig.blendw_sample_loss_weight = 0.
 
 
 EvalConfig.niter_runtime_eval = 10000

hypernerf/configs.py

@@ -102,43 +102,17 @@ class TrainConfig:
   elastic_loss_type: str = 'log_svals'
   # Whether to use background regularization.
   use_background_loss: bool = False
-  # Whether to use background decompose loss. This loss is only defined for DecomposeNerfModel
-  use_bg_decompose_loss: bool = False
-  # The weight for the background loss.
   background_loss_weight: float = 0.0
-  bg_decompose_loss_weight: float = 0.0
   # Adaptive blendw loss weight setting
   blendw_loss_weight_schedule: Optional[ScheduleDef] = None
-  # Adaptive blendw loss weight setting
-  blendw_pixel_loss_weight_schedule: Optional[ScheduleDef] = immutabledict.immutabledict({
-      'type': 'constant',
-      'value': 0.0,
-  })
   blendw_loss_skewness: float = 1.0
-  blendw_pixel_loss_skewness: float = 1.0
-  force_blendw_loss_weight: float = 1.0
-  blendw_ray_loss_weight: float = 0.0
   sigma_s_ray_loss_weight: float = 0.0
-  sigma_d_ray_loss_weight: float = 0.0
   blendw_area_loss_weight: float = 0.0
-  blendw_ray_loss_threshold: float = 1.0
-  shadow_loss_threshold: float = 0.2
-  shadow_loss_weight: float = 0.0
-  blendw_sample_loss_weight: float = 0.0
   shadow_r_loss_weight: Optional[ScheduleDef] = immutabledict.immutabledict({
       'type': 'constant',
       'value': 0.0,
   })
-  cubic_shadow_r_loss_weight_schedule: Optional[ScheduleDef] = immutabledict.immutabledict({
-      'type': 'constant',
-      'value': 0.0,
-  })
-  shadow_r_consistency_loss_weight_schedule: Optional[ScheduleDef] = immutabledict.immutabledict({
-      'type': 'constant',
-      'value': 0.0,
-  })
-  shadow_r_l2_loss_weight: float = 0.0
-  blendw_spatial_loss_weight: float = 0.0
+
   # The batch size for background regularization loss.
   background_points_batch_size: int = 16384
   # Whether to use the warp reg loss.
@@ -177,21 +151,9 @@ class TrainConfig:
   # Use decompose NeRF or not
   use_decompose_nerf: bool = False
 
-  # Initialize the static model for several iterations first
-  # During this, dynamic model would be frozen
-  freeze_dynamic_steps: int = 0
-  # Fix blendw to certain values for a fixed number of iters
-  fix_blendw_steps: int = 0
-  # Encourage the predicted blendw to be close to a certain value for some iters
-  # This is different to fixing the value
-  force_blendw_steps: int = 0
-  fix_blendw_value: float = 0.25
-
-  # Use provided dynmaic object mask to separately train the dynamic and static component
-  use_mask_sep_train: bool = False
-
   # Regularize the decompose model using rays combined with different time frame
   use_ex_ray_entropy_loss: bool = False
+  use_lap_blendw_loss: bool = False
 
 @gin.configurable()
 @dataclasses.dataclass
@@ -234,8 +196,6 @@ class EvalConfig:
   # Evalution setting for fixed time circulating camera experiments
   novel_view_eval: bool = False
 
-  # test_time_id: int = 0
-
   # Evalution setting for fixed view varying time experiments
   fix_view_eval: bool = False
   # Number of views rendered for each fixed time frame

hypernerf/datasets/core.py

@@ -215,7 +215,6 @@ def __init__(self,
                use_camera_id=False,
                use_warp_id=False,
                use_depth=False,
-               use_mask=False,
                mask_interest_region=False,
                use_relative_depth=False,
                use_time=False,
@@ -233,7 +232,6 @@ def __init__(self,
     self.use_warp_id = use_warp_id
     self.use_depth = use_depth
     self.use_time = use_time
-    self.use_mask = use_mask
     self.mask_interest_region = mask_interest_region
     self.use_relative_depth = use_relative_depth
     self.rng = np.random.RandomState(random_seed)
@@ -489,25 +487,6 @@ def _prepare_array(x):
 
     dataset = tf.data.Dataset.from_tensor_slices(out_dict)
 
-    if self.use_mask:
-      # separate two datasets, one with all rays that are on static background and the other for dynamic object
-      static_ray_ids = (out_dict['mask'][:,0]==0).nonzero()[0]
-      dynamic_ray_ids = np.setdiff1d(np.arange(num_rays), static_ray_ids)
-      static_dict = {}
-      dynamic_dict = {}
-      for key, value in out_dict.items():
-        static_dict[key] = jax.tree_map(lambda x: x[static_ray_ids], value)
-        dynamic_dict[key] = jax.tree_map(lambda x: x[dynamic_ray_ids], value)
-
-      static_data = tf.data.Dataset.from_tensor_slices(static_dict).repeat().batch(batch_size)
-      dynamic_data = tf.data.Dataset.from_tensor_slices(dynamic_dict).repeat().batch(batch_size)
-
-      zipped = tf.data.Dataset.zip((static_data, dynamic_data))
-      dataset = zipped.flat_map(lambda x0, x1: tf.data.Dataset.from_tensors(x0).concatenate(tf.data.Dataset.from_tensors(x1)))
-      dataset = dataset.unbatch()
-
-      # dataset = tf.data.Dataset.from_tensor_slices(static_dict)
-
     return dataset
 
   def _create_lazy_dataset(self,

hypernerf/model_utils.py

@@ -33,13 +33,6 @@ class TrainState:
   warp_alpha: Optional[jnp.ndarray] = None
   hyper_alpha: Optional[jnp.ndarray] = None
   hyper_sheet_alpha: Optional[jnp.ndarray] = None
-  # not implemented modification
-  # render_mode: jnp.ndarray = jnp.array([RENDER_MODE['regular']])
-  freeze_static: jnp.ndarray = jnp.array([False])
-  freeze_dynamic: jnp.ndarray = jnp.array([False])
-  freeze_blendw: jnp.ndarray = jnp.array([False])
-  force_blendw: jnp.ndarray = jnp.array([False])
-  freeze_blendw_value: jnp.ndarray = jnp.array([0.5])
 
   @property
   def extra_params(self):
@@ -48,12 +41,6 @@ def extra_params(self):
         'warp_alpha': self.warp_alpha,
         'hyper_alpha': self.hyper_alpha,
         'hyper_sheet_alpha': self.hyper_sheet_alpha,
-        # 'render_mode': self.render_mode,
-        'freeze_static': self.freeze_static,
-        'freeze_dynamic': self.freeze_dynamic,
-        'freeze_blendw': self.freeze_blendw,
-        'force_blendw': self.force_blendw,
-        'freeze_blendw_value': self.freeze_blendw_value
     }