experiments

opt/opt.py

@@ -183,6 +183,7 @@
 group.add_argument('--tv_lumisphere_dir_factor', type=float, default=0.0)
 
 group.add_argument('--lambda_l2_sh', type=float, default=0.0)#1e-4)
+group.add_argument('--tv_early_only', type=int, default=1) # Turn off TV regularization after the first split/prune
 # End Foreground TV
 
 group.add_argument('--lambda_sparsity', type=float, default=
@@ -211,6 +212,7 @@
 
 group.add_argument('--lr_decay', action='store_true', default=True)
 
+group.add_argument('--n_train', type=int, default=None, help='Number of training images. Defaults to use all avaiable.')
 args = parser.parse_args()
 config_util.maybe_merge_config_file(args)
 
@@ -239,6 +241,7 @@
                split="train",
                device=device,
                factor=factor,
+               n_images=args.n_train,
                **config_util.build_data_options(args))
 
 if args.background_nlayers > 0 and not dset.should_use_background:
@@ -554,8 +557,10 @@ def train_step():
         last_upsamp_step = gstep_id_base
         if reso_id < len(reso_list) - 1:
             print('* Upsampling from', reso_list[reso_id], 'to', reso_list[reso_id + 1])
-            #  args.lambda_tv = 0
-            #  args.lambda_tv_sh = 0
+            if args.tv_early_only > 0:
+                print('turning off TV regularization')
+                args.lambda_tv = 0.0
+                args.lambda_tv_sh = 0.0
             reso_id += 1
             use_sparsify = True
             z_reso = reso_list[reso_id] if isinstance(reso_list[reso_id], int) else reso_list[reso_id][2]

opt/tasks/interpablate.json

@@ -0,0 +1,18 @@
+{
+    "eval": true,
+    "data_root": "/home/sxyu/data/nerf_synthetic",
+    "train_root": "/home/sfk/svox2/opt/ckpt_auto/nearestneighbor",
+    "variables": {
+        "scene": ["lego", "mic", "ship", "chair", "ficus", "materials", "drums", "hotdog"], 
+        "reso": ["[[256,256,256]]", "[[128,128,128]]"]
+    },
+    "tasks": [{
+        "train_dir": "{reso}_3e0_{scene}",
+        "data_dir": "{scene}",
+        "flags": [
+            "-B", "svox1",
+            "--lr_sigma", "3e0",
+            "--reso", "{reso}"
+        ]
+    }]
+}

opt/tasks/ntrainablate.json

@@ -0,0 +1,16 @@
+{
+    "eval": true,
+    "data_root": "/home/sxyu/data/nerf_synthetic",
+    "train_root": "/home/sfk/svox2/opt/ckpt_auto/256_to_512_tvearlyonly_ntrain",
+    "variables": {
+        "scene": ["lego", "mic", "ship", "chair", "ficus", "materials", "drums", "hotdog"],
+        "n_train": [25]
+    },
+    "tasks": [{
+        "train_dir": "train_{n_train}_{scene}",
+        "data_dir": "{scene}",
+        "flags": [
+            "--n_train", "{n_train}"
+        ]
+    }]
+}

opt/tasks/tvearlyonly.json

@@ -0,0 +1,16 @@
+{
+    "eval": true,
+    "data_root": "/home/sxyu/data/nerf_synthetic",
+    "train_root": "/home/sfk/svox2/opt/ckpt_auto/256_to_512",
+    "variables": {
+        "scene": ["lego", "mic", "ship", "chair", "ficus", "materials", "drums", "hotdog"],
+        "tv_early_only": [0, 1]
+    },
+    "tasks": [{
+        "train_dir": "tv_early_only_{tv_early_only}_{scene}",
+        "data_dir": "{scene}",
+        "flags": [
+            "--tv_early_only", "{tv_early_only}"
+        ]
+    }]
+}

opt/util/obj_dataset.py

@@ -35,6 +35,7 @@ def __init__(
         scale : Optional[float] = None,
         permutation: bool = True,
         white_bkgd: bool = True,
+        n_images = None,
         **kwargs
     ):
         assert path.isdir(root), f"'{root}' is not a directory"
@@ -88,6 +89,15 @@ def __init__(
                 self.gt = self.gt[..., :3]
 
         self.n_images, self.h_full, self.w_full, _ = self.gt.shape
+        # Choose a subset of training images
+        if n_images is not None:
+            if n_images > self.n_images:
+                print(f'using {self.n_images} available training views instead of the requested {n_images}.')
+                n_images = self.n_images
+            self.n_images = n_images
+            self.gt = self.gt[0:n_images,...]
+            self.c2w = self.c2w[0:n_images,...]
+
         self.intrins_full : Intrin = Intrin(focal, focal,
                                             self.w_full * 0.5,
                                             self.h_full * 0.5)