(
self,
z: torch.Tensor,
is_first_frame: torch.Tensor,
*feat_cache: List[torch.Tensor]
)
| 166 | self.conv2 = CausalConv3d(self.z_dim, self.z_dim, 1) |
| 167 | |
| 168 | def forward( |
| 169 | self, |
| 170 | z: torch.Tensor, |
| 171 | is_first_frame: torch.Tensor, |
| 172 | *feat_cache: List[torch.Tensor] |
| 173 | ): |
| 174 | # from [batch_size, num_frames, num_channels, height, width] |
| 175 | # to [batch_size, num_channels, num_frames, height, width] |
| 176 | z = z.permute(0, 2, 1, 3, 4) |
| 177 | assert z.shape[2] == 1 |
| 178 | feat_cache = list(feat_cache) |
| 179 | is_first_frame = is_first_frame.bool() |
| 180 | |
| 181 | device, dtype = z.device, z.dtype |
| 182 | scale = [self.mean.to(device=device, dtype=dtype), |
| 183 | 1.0 / self.std.to(device=device, dtype=dtype)] |
| 184 | |
| 185 | if isinstance(scale[0], torch.Tensor): |
| 186 | z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view( |
| 187 | 1, self.z_dim, 1, 1, 1) |
| 188 | else: |
| 189 | z = z / scale[1] + scale[0] |
| 190 | x = self.conv2(z) |
| 191 | out, feat_cache = self.decoder(x, is_first_frame, feat_cache=feat_cache) |
| 192 | out = out.clamp_(-1, 1) |
| 193 | # from [batch_size, num_channels, num_frames, height, width] |
| 194 | # to [batch_size, num_frames, num_channels, height, width] |
| 195 | out = out.permute(0, 2, 1, 3, 4) |
| 196 | return out, feat_cache |
| 197 | |
| 198 | |
| 199 | class VAEDecoder3d(nn.Module): |
nothing calls this directly
no outgoing calls
no test coverage detected