Spaces:
Running
Running
File size: 5,709 Bytes
19fe404 f62c8b9 19fe404 f62c8b9 19fe404 f62c8b9 19fe404 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
from .common import CausalConv3d
class Upsampler(nn.Module):
def __init__(
self,
spatial_upsample_factor: int = 1,
temporal_upsample_factor: int = 1,
):
super().__init__()
self.spatial_upsample_factor = spatial_upsample_factor
self.temporal_upsample_factor = temporal_upsample_factor
class SpatialUpsampler3D(Upsampler):
def __init__(self, in_channels: int, out_channels: int):
super().__init__(spatial_upsample_factor=2)
if out_channels is None:
out_channels = in_channels
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = F.interpolate(x, scale_factor=(1, 2, 2), mode="nearest")
x = self.conv(x)
return x
class SpatialUpsamplerD2S3D(Upsampler):
def __init__(self, in_channels: int, out_channels: int):
super().__init__(spatial_upsample_factor=2)
if out_channels is None:
out_channels = in_channels
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels * 4,
kernel_size=3,
)
o, i, t, h, w = self.conv.weight.shape
conv_weight = torch.empty(o // 4, i, t, h, w)
nn.init.kaiming_normal_(conv_weight)
conv_weight = repeat(conv_weight, "o ... -> (o 4) ...")
self.conv.weight.data.copy_(conv_weight)
nn.init.zeros_(self.conv.bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.conv(x)
x = rearrange(x, "b (c p1 p2) t h w -> b c t (h p1) (w p2)", p1=2, p2=2)
return x
class TemporalUpsampler3D(Upsampler):
def __init__(self, in_channels: int, out_channels: int):
super().__init__(
spatial_upsample_factor=1,
temporal_upsample_factor=2,
)
if out_channels is None:
out_channels = in_channels
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if x.shape[2] > 1:
first_frame, x = x[:, :, :1], x[:, :, 1:]
x = F.interpolate(x, scale_factor=(2, 1, 1), mode="trilinear")
x = torch.cat([first_frame, x], dim=2)
x = self.conv(x)
return x
class TemporalUpsamplerD2S3D(Upsampler):
def __init__(self, in_channels: int, out_channels: int):
super().__init__(
spatial_upsample_factor=1,
temporal_upsample_factor=2,
)
if out_channels is None:
out_channels = in_channels
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels * 2,
kernel_size=3,
)
o, i, t, h, w = self.conv.weight.shape
conv_weight = torch.empty(o // 2, i, t, h, w)
nn.init.kaiming_normal_(conv_weight)
conv_weight = repeat(conv_weight, "o ... -> (o 2) ...")
self.conv.weight.data.copy_(conv_weight)
nn.init.zeros_(self.conv.bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.conv(x)
x = rearrange(x, "b (c p1) t h w -> b c (t p1) h w", p1=2)
x = x[:, :, 1:]
return x
class SpatialTemporalUpsampler3D(Upsampler):
def __init__(self, in_channels: int, out_channels: int):
super().__init__(
spatial_upsample_factor=2,
temporal_upsample_factor=2,
)
if out_channels is None:
out_channels = in_channels
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
)
self.padding_flag = 0
self.set_3dgroupnorm = False
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = F.interpolate(x, scale_factor=(1, 2, 2), mode="nearest")
x = self.conv(x)
if self.padding_flag == 0:
if x.shape[2] > 1:
first_frame, x = x[:, :, :1], x[:, :, 1:]
x = F.interpolate(x, scale_factor=(2, 1, 1), mode="trilinear" if not self.set_3dgroupnorm else "nearest")
x = torch.cat([first_frame, x], dim=2)
elif self.padding_flag == 2 or self.padding_flag == 4 or self.padding_flag == 5 or self.padding_flag == 6:
x = F.interpolate(x, scale_factor=(2, 1, 1), mode="trilinear" if not self.set_3dgroupnorm else "nearest")
return x
class SpatialTemporalUpsamplerD2S3D(Upsampler):
def __init__(self, in_channels: int, out_channels: int):
super().__init__(
spatial_upsample_factor=2,
temporal_upsample_factor=2,
)
if out_channels is None:
out_channels = in_channels
self.conv = CausalConv3d(
in_channels=in_channels,
out_channels=out_channels * 8,
kernel_size=3,
)
o, i, t, h, w = self.conv.weight.shape
conv_weight = torch.empty(o // 8, i, t, h, w)
nn.init.kaiming_normal_(conv_weight)
conv_weight = repeat(conv_weight, "o ... -> (o 8) ...")
self.conv.weight.data.copy_(conv_weight)
nn.init.zeros_(self.conv.bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.conv(x)
x = rearrange(x, "b (c p1 p2 p3) t h w -> b c (t p1) (h p2) (w p3)", p1=2, p2=2, p3=2)
x = x[:, :, 1:]
return x
|