radii pruning

This commit is contained in:
Justin Kerr committed 2023-09-29 17:38:04 -07:00
1 parent 84b8241b4f
commit 53d8031884
1 file changed
+29 -12
+29 -12
View File
@@ -93,9 +93,9 @@ class GaussianSplattingModelConfig(ModelConfig):
"""training starts at 1/d resolution, every n steps this is doubled"""
num_downscales: int = 2
"""at the beginning, resolution is 1/2^d, where d is this number"""
cull_alpha_thresh:float = .01
cull_alpha_thresh:float = .005
"""threshold of opacity for culling gaussians"""
cull_scale_thresh:float = .5
cull_scale_thresh:float = .1
"""threshold of scale for culling gaussians"""
reset_alpha_every:int = 30
"""Every this many refinement steps, reset the alpha"""
@@ -106,12 +106,12 @@ class GaussianSplattingModelConfig(ModelConfig):
#set to 3 to effeectively turn off, it seems to make things worse? need to investigate
sh_degree_interval: int = 3
"""every n intervals turn on another sh degree"""
max_screen_size: int = 200
"""maximum screen size of a gaussian, in pixels"""
class GaussianSplattingModel(Model):
"""Gaussian Splatting model
TODO (jake-austin): Figure out how to print out on the training log in terminal the number of splats
Args:
config: Gaussian Splatting configuration to instantiate model
"""
@@ -128,6 +128,7 @@ class GaussianSplattingModel(Model):
self.means_grad_norm = None
init_scale = torch.log(torch.tensor(.01)).item()
self.scales = torch.nn.Parameter(torch.full((self.num_points,3),init_scale))
self.max_2Dsize = None
self.quats = torch.nn.Parameter(random_quat_tensor(self.num_points))
self.degree = 3
dim_sh = num_sh_bases(self.degree)
@@ -174,11 +175,20 @@ class GaussianSplattingModel(Model):
def after_train(self,step):
with torch.no_grad():
#keep track of a moving average of grad norms
weight = 1/self.config.refine_every
visible_mask = self.radii>0
if self.means_grad_norm is None:
self.means_grad_norm = weight*self.means.grad.detach().norm(dim=-1,keepdim=True)
self.means_grad_norm = self.means.grad.detach().norm(dim=-1,keepdim=True)
self.vis_counts = torch.zeros_like(self.means_grad_norm)
else:
self.means_grad_norm = weight*self.means.grad.detach().norm(dim=-1,keepdim=True) + self.means_grad_norm
self.vis_counts[visible_mask] = self.vis_counts[visible_mask] + 1
self.means_grad_norm[visible_mask] = self.means.grad.detach()[visible_mask].norm(dim=-1,keepdim=True) + self.means_grad_norm[visible_mask]
#update the max screen size
if self.max_2Dsize is None:
self.max_2Dsize = torch.zeros_like(self.radii)
newradii = self.radii[visible_mask]
self.max_2Dsize[visible_mask] = torch.maximum(self.max_2Dsize[visible_mask],newradii)
def refinement_before(self, optimizers:Optimizers, step):
print("Inside refinement before")
@@ -196,7 +206,8 @@ class GaussianSplattingModel(Model):
if self.step > self.config.warmup_length:
with torch.no_grad():
#then we densify
high_grads = (self.means_grad_norm > self.config.densify_grad_thresh).squeeze()
avg_grad_norm = self.means_grad_norm / self.vis_counts
high_grads = (avg_grad_norm > self.config.densify_grad_thresh).squeeze()
splits = (self.scales.exp().max(dim=-1).values > self.config.densify_size_thresh).squeeze()
splits &= high_grads
split_means,split_colors,split_opacities,split_scales,split_quats = self.split_gaussians(splits)
@@ -216,7 +227,6 @@ class GaussianSplattingModel(Model):
param_groups = self.get_param_groups()
for group,param in param_groups.items():
print(f"adding {len(add_idcs)} params to {group} optimizer")
self.dup_in_optim(optimizers.optimizers[group],add_idcs,param)
if self.step // self.config.refine_every % self.config.reset_alpha_every == 0:
@@ -230,6 +240,8 @@ class GaussianSplattingModel(Model):
param_state['exp_avg'] = torch.zeros_like(param_state['exp_avg'])
param_state['exp_avg_sq'] = torch.zeros_like(param_state['exp_avg_sq'])
self.means_grad_norm = None
self.vis_counts = None
self.max_2Dsize = None
def cull_gaussians(self):
"""
@@ -240,7 +252,12 @@ class GaussianSplattingModel(Model):
culls = (torch.sigmoid(self.opacities) < self.config.cull_alpha_thresh).squeeze()
#cull huge ones
culls = culls | (torch.exp(self.scales).max(dim=-1).values > self.config.cull_scale_thresh).squeeze()
#cull big screen space
if self.step>self.config.refine_every*self.config.reset_alpha_every:
culls = culls | (self.max_2Dsize > self.config.max_screen_size).squeeze()
self.means_grad_norm = self.means_grad_norm[~culls].detach()
self.max_2Dsize = self.max_2Dsize[~culls].detach()
self.vis_counts = self.vis_counts[~culls].detach()
self.means = Parameter(self.means[~culls].detach())
self.scales = Parameter(self.scales[~culls].detach())
self.quats = Parameter(self.quats[~culls].detach())
@@ -391,7 +408,7 @@ class GaussianSplattingModel(Model):
background = torch.rand(3, device=self.device)
else:
background = torch.zeros(3, device=self.device)
xys, depths, radii, conics, num_tiles_hit,self.cov3d = ProjectGaussians.apply(
xys, depths, self.radii, conics, num_tiles_hit,self.cov3d = ProjectGaussians.apply(
self.means,
torch.exp(self.scales),
1,
@@ -414,10 +431,10 @@ class GaussianSplattingModel(Model):
cx_delta = cx - W / 2
cy_delta = cy - H / 2
xys = xys.view(-1, 2) + torch.tensor([cx_delta, cy_delta], device=self.device)
rgb = RasterizeGaussians.apply(
rgb,gauss_ids_sorted = RasterizeGaussians.apply(
xys,
depths,
radii,
self.radii,
conics,
num_tiles_hit,
torch.sigmoid(rgbs),