gradually increase density threshold to remove floater

train.py

@@ -5,13 +5,14 @@
 import numpy as np
 import cv2
 
+# data
 from torch.utils.data import DataLoader
 from datasets import dataset_dict
 
 # models
 from kornia.utils.grid import create_meshgrid3d
 from models.networks import NGP
-from models.rendering import render
+from models.rendering import render, MAX_SAMPLES
 
 # optimizer, losses
 from apex.optimizers import FusedAdam
@@ -26,6 +27,8 @@
 from pytorch_lightning.callbacks import TQDMProgressBar, ModelCheckpoint
 from pytorch_lightning.loggers import TensorBoardLogger
 
+from utils import slim_ckpt
+
 import warnings; warnings.filterwarnings("ignore")
 
 
@@ -103,7 +106,9 @@ def val_dataloader(self):
 
     def training_step(self, batch, batch_nb):
         if self.global_step%self.S == 0:
-            self.model.update_density_grid(warmup=self.global_step<256)
+            a_thr = min(self.current_epoch+1, 25)/50 # alpha threshold
+            self.model.update_density_grid(a_thr*MAX_SAMPLES/(2*3**0.5),
+                                           warmup=self.global_step<256)
 
         rays, rgb = batch['rays'], batch['rgb']
         results = self(rays, split='train')
@@ -179,3 +184,7 @@ def validation_epoch_end(self, outputs):
                       precision=16)
 
     trainer.fit(system, ckpt_path=hparams.ckpt_path)
+
+    # save slimmed ckpt for the last epoch
+    ckpt_ = slim_ckpt(f'ckpts/{hparams.exp_name}/epoch={hparams.num_epochs-1}.ckpt')
+    torch.save(f'ckpts/{hparams.exp_name}/epoch={hparams.num_epochs-1}_slim.ckpt', ckpt_)

utils.py

@@ -22,4 +22,13 @@ def load_ckpt(model, ckpt_path, model_name='model', prefixes_to_ignore=[]):
     model_dict = model.state_dict()
     checkpoint_ = extract_model_state_dict(ckpt_path, model_name, prefixes_to_ignore)
     model_dict.update(checkpoint_)
-    model.load_state_dict(model_dict)
+    model.load_state_dict(model_dict)
+
+
+def slim_ckpt(ckpt_path):
+    ckpt = torch.load(ckpt_path)
+    # pop unused parameters
+    ckpt['state_dict'].pop('weights', None)
+    ckpt['state_dict'].pop('model.density_grid', None)
+    ckpt['state_dict'].pop('model.grid_coords', None)
+    return ckpt['state_dict']