Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,792 Bytes
1ea89dd 02b5a6d 1ea89dd 02b5a6d 1ea89dd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 |
import torch
import torch.nn as nn
from einops import rearrange
from unik3d.utils.constants import VERBOSE
from unik3d.utils.misc import profile_method
class ResidualConvUnit(nn.Module):
def __init__(
self,
dim,
kernel_size: int = 3,
padding_mode: str = "zeros",
dilation: int = 1,
layer_scale: float = 1.0,
use_norm: bool = False,
):
super().__init__()
self.conv1 = nn.Conv2d(
dim,
dim,
kernel_size=kernel_size,
padding=dilation * (kernel_size - 1) // 2,
dilation=dilation,
padding_mode=padding_mode,
)
self.conv2 = nn.Conv2d(
dim,
dim,
kernel_size=kernel_size,
padding=dilation * (kernel_size - 1) // 2,
dilation=dilation,
padding_mode=padding_mode,
)
self.activation = nn.LeakyReLU()
self.skip_add = nn.quantized.FloatFunctional()
self.gamma = (
nn.Parameter(layer_scale * torch.ones(1, dim, 1, 1))
if layer_scale > 0.0
else 1.0
)
self.norm1 = nn.GroupNorm(dim // 16, dim) if use_norm else nn.Identity()
self.norm2 = nn.GroupNorm(dim // 16, dim) if use_norm else nn.Identity()
def forward(self, x):
out = self.activation(x)
out = self.conv1(out)
out = self.norm1(out)
out = self.activation(out)
out = self.conv2(out)
out = self.norm2(out)
return self.skip_add.add(self.gamma * out, x)
class ResUpsampleBil(nn.Module):
def __init__(
self,
hidden_dim,
output_dim: int = None,
num_layers: int = 2,
kernel_size: int = 3,
layer_scale: float = 1.0,
padding_mode: str = "zeros",
use_norm: bool = False,
**kwargs,
):
super().__init__()
output_dim = output_dim if output_dim is not None else hidden_dim // 2
self.convs = nn.ModuleList([])
for _ in range(num_layers):
self.convs.append(
ResidualConvUnit(
hidden_dim,
kernel_size=kernel_size,
layer_scale=layer_scale,
padding_mode=padding_mode,
use_norm=use_norm,
)
)
self.up = nn.Sequential(
nn.Conv2d(
hidden_dim,
output_dim,
kernel_size=1,
padding=0,
padding_mode=padding_mode,
),
nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False),
)
@profile_method(verbose=True)
def forward(self, x: torch.Tensor):
for conv in self.convs:
x = conv(x)
x = self.up(x)
return x
class ResUpsample(nn.Module):
def __init__(
self,
hidden_dim,
num_layers: int = 2,
kernel_size: int = 3,
layer_scale: float = 1.0,
padding_mode: str = "zeros",
**kwargs,
):
super().__init__()
self.convs = nn.ModuleList([])
for _ in range(num_layers):
self.convs.append(
ResidualConvUnit(
hidden_dim,
kernel_size=kernel_size,
layer_scale=layer_scale,
padding_mode=padding_mode,
)
)
self.up = nn.ConvTranspose2d(
hidden_dim, hidden_dim // 2, kernel_size=2, stride=2, padding=0
)
@profile_method(verbose=True)
def forward(self, x: torch.Tensor):
for conv in self.convs:
x = conv(x)
x = self.up(x)
return x
class ResUpsampleSH(nn.Module):
def __init__(
self,
hidden_dim,
num_layers: int = 2,
kernel_size: int = 3,
layer_scale: float = 1.0,
padding_mode: str = "zeros",
**kwargs,
):
super().__init__()
self.convs = nn.ModuleList([])
for _ in range(num_layers):
self.convs.append(
ResidualConvUnit(
hidden_dim,
kernel_size=kernel_size,
layer_scale=layer_scale,
padding_mode=padding_mode,
)
)
self.up = nn.Sequential(
nn.PixelShuffle(2),
nn.Conv2d(
hidden_dim // 4,
hidden_dim // 2,
kernel_size=3,
padding=1,
padding_mode=padding_mode,
),
)
def forward(self, x: torch.Tensor):
for conv in self.convs:
x = conv(x)
x = self.up(x)
return x
|