PLUS-WAVE's picture
Deploy InfiniSplat ZeroGPU demo
41ff959 verified
Raw
History Blame Contribute Delete
7.45 kB
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn as nn
import torch.nn.functional as F
def bilinear_sampler(input, coords, align_corners=True, padding_mode="border"):
r"""Sample a tensor using bilinear interpolation
`bilinear_sampler(input, coords)` samples a tensor :attr:`input` at
coordinates :attr:`coords` using bilinear interpolation. It is the same
as `torch.nn.functional.grid_sample()` but with a different coordinate
convention.
The input tensor is assumed to be of shape :math:`(B, C, H, W)`, where
:math:`B` is the batch size, :math:`C` is the number of channels,
:math:`H` is the height of the image, and :math:`W` is the width of the
image. The tensor :attr:`coords` of shape :math:`(B, H_o, W_o, 2)` is
interpreted as an array of 2D point coordinates :math:`(x_i,y_i)`.
Alternatively, the input tensor can be of size :math:`(B, C, T, H, W)`,
in which case sample points are triplets :math:`(t_i,x_i,y_i)`. Note
that in this case the order of the components is slightly different
from `grid_sample()`, which would expect :math:`(x_i,y_i,t_i)`.
If `align_corners` is `True`, the coordinate :math:`x` is assumed to be
in the range :math:`[0,W-1]`, with 0 corresponding to the center of the
left-most image pixel :math:`W-1` to the center of the right-most
pixel.
If `align_corners` is `False`, the coordinate :math:`x` is assumed to
be in the range :math:`[0,W]`, with 0 corresponding to the left edge of
the left-most pixel :math:`W` to the right edge of the right-most
pixel.
Similar conventions apply to the :math:`y` for the range
:math:`[0,H-1]` and :math:`[0,H]` and to :math:`t` for the range
:math:`[0,T-1]` and :math:`[0,T]`.
Args:
input (Tensor): batch of input images.
coords (Tensor): batch of coordinates.
align_corners (bool, optional): Coordinate convention. Defaults to `True`.
padding_mode (str, optional): Padding mode. Defaults to `"border"`.
Returns:
Tensor: sampled points.
"""
sizes = input.shape[2:]
assert len(sizes) in [2, 3]
if len(sizes) == 3:
# t x y -> x y t to match dimensions T H W in grid_sample
coords = coords[..., [1, 2, 0]]
if align_corners:
coords = coords * torch.tensor(
[2 / max(size - 1, 1) for size in reversed(sizes)], device=coords.device
)
else:
coords = coords * torch.tensor(
[2 / size for size in reversed(sizes)], device=coords.device
)
coords -= 1
return F.grid_sample(
input, coords, align_corners=align_corners, padding_mode=padding_mode
)
class ResidualBlock(nn.Module):
def __init__(self, in_planes, planes, norm_fn="group", stride=1):
super(ResidualBlock, self).__init__()
self.conv1 = nn.Conv2d(
in_planes,
planes,
kernel_size=3,
padding=1,
stride=stride,
padding_mode="zeros",
)
self.conv2 = nn.Conv2d(
planes, planes, kernel_size=3, padding=1, padding_mode="zeros"
)
self.relu = nn.ReLU(inplace=True)
num_groups = planes // 8
if norm_fn == "group":
self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
if not stride == 1:
self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
elif norm_fn == "batch":
self.norm1 = nn.BatchNorm2d(planes)
self.norm2 = nn.BatchNorm2d(planes)
if not stride == 1:
self.norm3 = nn.BatchNorm2d(planes)
elif norm_fn == "instance":
self.norm1 = nn.InstanceNorm2d(planes)
self.norm2 = nn.InstanceNorm2d(planes)
if not stride == 1:
self.norm3 = nn.InstanceNorm2d(planes)
elif norm_fn == "none":
self.norm1 = nn.Sequential()
self.norm2 = nn.Sequential()
if not stride == 1:
self.norm3 = nn.Sequential()
if stride == 1:
self.downsample = None
else:
self.downsample = nn.Sequential(
nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3
)
def forward(self, x):
y = x
y = self.relu(self.norm1(self.conv1(y)))
y = self.relu(self.norm2(self.conv2(y)))
if self.downsample is not None:
x = self.downsample(x)
return self.relu(x + y)
class BasicEncoder(nn.Module):
def __init__(self, input_dim=3, output_dim=128, stride=4):
super(BasicEncoder, self).__init__()
self.stride = stride
self.norm_fn = "instance"
self.in_planes = output_dim // 2
self.norm1 = nn.InstanceNorm2d(self.in_planes)
self.norm2 = nn.InstanceNorm2d(output_dim * 2)
self.conv1 = nn.Conv2d(
input_dim,
self.in_planes,
kernel_size=7,
stride=2,
padding=3,
padding_mode="zeros",
)
self.relu1 = nn.ReLU(inplace=True)
self.layer1 = self._make_layer(output_dim // 2, stride=1)
self.layer2 = self._make_layer(output_dim // 4 * 3, stride=2)
self.layer3 = self._make_layer(output_dim, stride=2)
self.layer4 = self._make_layer(output_dim, stride=2)
self.conv2 = nn.Conv2d(
output_dim * 3 + output_dim // 4,
output_dim * 2,
kernel_size=3,
padding=1,
padding_mode="zeros",
)
self.relu2 = nn.ReLU(inplace=True)
self.conv3 = nn.Conv2d(output_dim * 2, output_dim, kernel_size=1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
elif isinstance(m, (nn.InstanceNorm2d)):
if m.weight is not None:
nn.init.constant_(m.weight, 1)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def _make_layer(self, dim, stride=1):
layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride)
layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1)
layers = (layer1, layer2)
self.in_planes = dim
return nn.Sequential(*layers)
def forward(self, x):
_, _, H, W = x.shape
x = self.conv1(x)
x = self.norm1(x)
x = self.relu1(x)
a = self.layer1(x)
b = self.layer2(a)
c = self.layer3(b)
d = self.layer4(c)
def _bilinear_intepolate(x):
return F.interpolate(
x,
(H // self.stride, W // self.stride),
mode="bilinear",
align_corners=True,
)
a = _bilinear_intepolate(a)
b = _bilinear_intepolate(b)
c = _bilinear_intepolate(c)
d = _bilinear_intepolate(d)
x = self.conv2(torch.cat([a, b, c, d], dim=1))
x = self.norm2(x)
x = self.relu2(x)
x = self.conv3(x)
return x